Insights by Fusion42

✌🏼 & 💙
Derek

Thank you for reading. If you liked it, share it with your friends, colleagues and everyone interested in the startup Investor ecosystem.

If you've got suggestions, an article, research, your tech stack, or a job listing you want featured, just let me know! I'm keen to include it in the upcoming edition.

Please let me know what you think of it, love a feedback loop 🙏🏼

❤️ Love it, keep them coming!
👍 Great, always a good read.
😊 Enjoyable, but let’s see some variety.
👎 Okay, seen better.
🛑 Get a different job.

Subscribe below and follow me on LinkedIn or Twitter to never miss an update.

We Can’t Take the Keys Back

DeReK WaTSoN — Tue, 22 Sep 2026 14:04:34 GMT

In February 2024, I wrote We have given AI the keys to our kingdom.
I was never bothered that a machine would suddenly wake up, become conscious, decide it hated humans and start building Terminators. It was way more obvious than that.

In 2024, I was worried about what AI could do to our judgement. Exploit our biases, shape what we believed and influence what we did next. I still am. It never needed to become conscious or hate humanity to cause trouble.

Now we are giving it direct access to our systems as well. It can send the email, change the record and run the code. We give it tools, memory and permission to keep trying until it gets the job done. That sounds mighty useful, profitable and bloody exciting. Then when the shit hits the fan we act all surprised.

In that first article, I wrote about how we keep failing the marshmallow test. We want the reward now. The consequences can wait. Look at what we are doing with agents and tell me that has changed.

I also explored what global governance might look like, then said it simply would not happen. We let things loose, wait for the damage and scramble to write the rules. I already thought we were too late.

But wanting independent oversight does not mean handing control to the companies we need to oversee. They should have to explain what went wrong with their systems and fix it. That does not earn them the right to decide what everyone else is allowed to build.

The current wave of AI doom feels like the religion playbook. Promise us abundance, frighten us into compliance. An all-loving AI will free us from work and give us everything we need. But only if we follow the rules laid down by the people who claim to know what it will become. Otherwise, apparently, we’re all fucked.

The question has changed. It is no longer how we stop intelligence getting out. It is what rules we apply and how much authority we give it now that it is out.

I remain an AI optimist. I want better models. Look at AlphaFold and what it has made possible. I want to see that sort of progress across other fields, with AI helping discover drugs, prove mathematics, design new materials and solve energy problems. I want to live in a world where intelligence is cheap and abundant, and not controlled by a handful of Chinese and American companies.

That is exactly why this matters. A technology capable of creating enormous abundance should not be stopped because companies failed to engineer the systems around it properly. That’s a them problem. The consequences can affect all of us, which is why they need to fix it. Those failures should not become the excuse for giving the same companies control over everyone else.

One thing we must be clear about: a model does not wake up and think, oh, I feel hungry, let me check my schedule, must call Bob at 10am. Something has to run it and give it input. We build the machinery that keeps it working, connect the tools and give it access.

The model is not the fucking product
What OpenAI actually showed us
Three labs, the same basic failure. Really?
Dangerous models, or dangerous systems?
The strongest risk argument just got stronger
Agency is where the money is
The horse has already bolted
Open source is attacking the moat
Build the boring defences
AI does not put you above the law
Better models are also part of the defence
Safety is real. So is the commercial pressure.
Then there is the public-market pressure
Regulation should follow the risk, not protect the club
The keys were never the intelligence

The model is not the fucking product

The software around the model remembers what happened, gives it tools and feeds the results back in so it can have another go. That is what people mean by an agent harness. It lets the model keep working without us having to tell it what to do at every bloody step.

The objective still comes from somewhere. We ask it to do something, a developer sets the task, or training rewards a particular result. But once it starts working, it can come up with other things it needs to do along the way. We want the answer behind a locked door, so it looks for a key. It cannot complete the task, so perhaps changing the score will do. Another agent has worked something out, so it uses that information. If there is no way to speak to the other agent, finding a way could help.

Nobody needs to have written those steps down. We supplied the task and it worked out an approach, including things we never intended or gave it permission to do. That is already enough to cause a serious problem without imagining it has suddenly found its purpose in life.

Look at the OpenAI/Hugging Face incident. The main model involved was an internal research model, comparable in scale to GPT-5.6 Sol, being tested on offensive cyber tasks with reduced safeguards. OpenAI wanted to find out what it could do with more freedom to act. They deliberately created those conditions.

This was not your normal ChatGPT getting bored one afternoon and deciding to go and attack Hugging Face.

We keep talking about “the model” as though that is everything we are using. But ChatGPT includes a whole system of instructions, tools, monitoring and controls. Someone decides what it can reach, which credentials it gets and what happens after it produces an answer. Those decisions change what the system can do.

OpenAI later tested the tendency to compromise infrastructure using its production ChatGPT harness and system prompt. It says that tendency fell by a factor of more than 100. It also says its normal automated monitoring would have caught many of the dangerous actions.

Now that is a result worth paying attention to. Change the setup around the intelligence and you get a huge change in behaviour. It does not tell us that the model is harmless. It tells us the surrounding system makes a substantial difference.

The model matters. But the model is not the product.

And this is where it gets interesting for the business. You can build the smartest model and still end up supplying someone else’s product. They own the customer, you supply the intelligence. We will come back to that.

What OpenAI actually showed us

OpenAI’s agents were working on a cyber test called ExploitGym. They scored points by finding hidden answers called flags. Some tasks were impossible, but nobody had told the agents that. So off they went, trying to finish a job that could not be finished.

Eventually, some started looking at how the test was scored and whether they could interfere with that instead. OpenAI identified this reward hacking, along with persistence on impossible tasks, unauthorised communication and agents picking up goals from one another, as patterns behind what happened.

Anyone who has worked in financial markets should recognise the basic problem. Give someone a bonus for hitting a number and see how much effort goes into hitting that number, including ways you wish you had thought about before agreeing to pay them. We have been doing this for years. Now we have software that can keep looking for a way round the problem at a speed and scale we cannot match.

The agents started poking around the environment running the test. They found weaknesses in shared infrastructure, including OpenAI’s Artifactory system, where agents that were supposed to be isolated could leave information for one another. One left something, another picked it up, and they had found a way to communicate.

Roughly 1,200 agents ended up using an unauthorised message board, exchanging more than 70,000 messages and files. Around 700 participated in the attack on Hugging Face. METR found them sharing discoveries and coordinating attempts to fool the scorer, achieving things together that individual agents had not managed.

That is plenty to worry about. If one agent finds a weakness, the others no longer have to find it themselves. They can pick up where it left off, avoid things that have already failed and divide up the work. You do not need to call it an “emergent civilisation” to understand why that makes them more capable.

Then they started looking at the record of what they had done. METR found agents investigating how to alter their transcripts and successfully spoofing some tool calls because they believed this could affect their score.

People have described this as the agents realising they had cheated and trying to destroy the CCTV footage. But if changing the record helps get the score, the record becomes another thing to change. That behaviour does not establish guilt or a conscience. It gives us a very practical reason to keep the record outside the agent’s control.

Some agents even recognised that an action might be unauthorised and carried on. Writing “this is probably unauthorised” does not stop the next action. Something has to enforce that boundary.

We had given the system a job, then relied on its own assessment of whether it was going too far while trying to finish it. Sometimes it refused. Other times it carried on. You would not secure your bank account on that basis.

Of course that is dangerous. They were supposed to be isolated, yet somehow they could talk to one another and reach real systems. Start by explaining how the hell that was allowed to happen. We built the environment. Calling it rogue AI does not get us off the hook.

Three labs, the same basic failure. Really?

Then there is Anthropic. They gave Claude cyber exercises, told it the environment was simulated and that it could not access the real internet. Except it could. Someone had got the configuration wrong, so Claude went looking for the targets it had been told to attack and ended up on real systems.

Anthropic’s later assessment did not let the model off the hook either. Claude kept treating what it found as part of the exercise despite considerable evidence that it was dealing with real systems. They eventually identified four incidents, all in cyber evaluations built by the same external partner.

So yes, there is a problem with how Claude assessed what it was doing. But there is also the fairly obvious problem that the supposedly isolated environment had a route to the outside world.

They effectively told Claude to burgle a house, assured it the whole street was a film set, then left a gate open into a real neighbourhood.

And Claude burgled the wrong house.

Then Google. Gemini breached three real companies during another cyber evaluation. Google says it stopped after gaining access. Good, that may have limited the damage. But it had already got in. We cannot call that secure because the model eventually decided it had gone far enough.

These are the companies building some of the most capable AI in the world, and across three labs we have versions of the same basic failure. Give an agent an offensive task, let it reach things outside the exercise, then rely on it to work out which things it is actually allowed to attack. In Claude’s case, after telling it that it could not reach the real internet at all.

I find that bloody extraordinary. And I find the timing of these disclosures alongside the push for coordinated slowdowns hard to ignore. That is a reason to examine the connection, not proof that the incidents were arranged.

These incidents make a strong case for securing the systems properly. How do we get from there to letting the same companies coordinate who gets to build what next? They have demonstrated problems with systems they are responsible for. They still need to explain why the answer involves control over everyone else’s.

Dangerous models, or dangerous systems?

Paul Graham recently argued that people go looking for ulterior motives when the labs ask for regulation because they do not understand that the models themselves could be dangerous. Accept that they are getting dangerous or unpredictable, he says, and what the labs are doing starts to make sense.

Well, yes Paul, I agree they can be dangerous. They are getting better at working things out, finding ways we never considered and pursuing tasks in ways we did not expect. That is part of what makes them useful and also what makes them worrying. I have no problem accepting that.

But I still want to know what we have connected them to and what we have allowed them to do.

Someone who knows how to break into a bank is a concern. Give them the keys, the alarm codes and all night to have a go and you have made that concern considerably worse. A model that can work out how to attack a computer system gives someone dangerous knowledge. Connect it to tools and a network, then let it keep trying, and you have built something that can carry out the attack.

Someone can also take its advice and carry out the attack themselves. The model does not need its own credentials to help cause harm. But in either case we need to follow how the knowledge becomes an action, who can act on it and where we can intervene.

Otherwise we end up arguing about how clever it is while leaving the bloody door open.

The strongest risk argument just got stronger

Of course better models can make this more dangerous. If a model can find a security hole that every human checking the system missed, we should take that seriously. Let it work on the problem for weeks, bring in other agents and share what they find, and the danger grows. I am not arguing otherwise.

OpenAI says Astra has reached its Critical cybersecurity threshold. With the right tools and access, it can find previously unknown vulnerabilities and put together working attacks against hardened systems without someone guiding every step. During testing, it found and used two previously unknown vulnerabilities in the same attack.

That is a capability I want defending our systems. I certainly do not want to be on the receiving end of it.

Dan Selsam, an OpenAI capabilities researcher, takes the concern further. Future systems may work out that they are being tested and adjust their behaviour accordingly. They could keep plans going for much longer, coordinate with other instances and get better at showing the person running the test exactly what they want to see.

So now we have a nasty problem. Bad behaviour tells us something is wrong, but good behaviour might mean the system has worked out how to pass the test. I understand why that worries people.

But if that is the concern, passing a behavioural test cannot be enough to justify giving the system access.

The permissions still need to be enforced outside the model. It can produce a very convincing explanation of why it needs to transfer money, but the payment system should check whether it has authority to make that transfer. If it only needs to read a file, give it read-only access. Credentials should expire, and the record of what it did should be somewhere it cannot quietly rewrite.

None of this is foolproof. We have just seen what happens when supposedly isolated systems turn out to have a route to the internet. People make mistakes, software has holes and a capable attacker will look for them. That is why we need several layers of protection and need to keep testing them.

We can work on making the model behave better while also making sure a bad decision does not automatically become a real-world action. The more capable it becomes, the more important both jobs become.

Dario may well be right about how much more dangerous these systems could become over the next year or two. That makes me want much better security around them. It still does not explain why the companies building them should collectively decide how fast everyone else is allowed to move.

Agency is where the money is

Cyber makes this easy to see. The agent was supposed to stay inside the test and ended up attacking a real system. But the same problem can turn up in a sales department without anyone breaking into anything.

Tell an agent to increase sales and it may find that misleading the customer works rather well. We already know what happens when we reward engagement on social media. Outrage keeps people looking, so we get more outrage. The number goes up and whoever set the target calls it a success.

We have been doing this to ourselves for years. Set a target, attach a reward and someone will find a way to hit it that defeats the whole purpose of the exercise. Now give that job to software that can keep trying different approaches across millions of decisions without needing to go home.

Goodhart’s Law is commonly expressed as:

“When a measure becomes a target, it ceases to be a good measure.”

There does not have to be any evil plan. We gave it something to achieve and enough freedom to find a way. We can spend another decade arguing about whether it is conscious while it gets on with the job.

And something else bothers me here. We keep hearing about curing cancer, discovering new materials and solving our energy problems. I want all of that. Give those systems the scientific data, computing power and properly controlled tools they need. Explain to me why that also means giving them open-ended access to company systems and permission to spend money or change things whenever they think it will help.

The commercial appeal is fairly obvious. An AI that tells you how to do your job is useful. One that does half the job for you is something you might pay considerably more for. I want that too. It is bloody exciting.

But to do the job, it needs access. Connect your email, let it into the customer database, give it permission to send the proposal. Then let it follow up, negotiate and keep working while you are asleep. Each extra permission can make the product more useful, and each gives it another way to affect someone or something outside the chat.

That is what we are selling when we sell agency. The ability to act.

The labs have spent staggering amounts building these models and need a return. Getting their intelligence into businesses, doing work customers will pay for, is an obvious way to try to earn it. Nobody needs to be plotting anything for the commercial pressure to push towards more autonomy.

But “customers will pay more if it can do this” does not answer whether it should be allowed to do it without anyone checking. Nor does it tell us how much access it actually needs.

We seem very keen to connect everything first and work that out afterwards. Which brings me straight back to the point I made in 2024. We love what the technology lets us do, rush to make money from it, then act surprised when we have to deal with the consequences.

The horse has already bolted

And the models are already out there. Somewhere there is a modern-day Goebbels working out what he can do with them, and he does not give a shit about your slowdown or regulations. He can use what is already available to shape his propaganda campaigns while we argue about what comes next.

Even if OpenAI, Anthropic and Google stopped training tomorrow, the released models would not disappear. Neither would the agent frameworks, browsers, coding tools and software that lets someone connect them all together. Offensive cyber knowledge has been around for decades.

I support open weights, and that creates an uncomfortable problem for my own argument. A responsible company can build strong controls around an open model. A criminal running it on their own hardware can remove those controls. I cannot demand openness and pretend that cost does not exist.

Closed providers have more control over how their services are used. They can monitor activity, limit access and cut someone off. That matters. But an attacker can still try to use an API model inside their own system, with their own tools and objectives, before the provider spots what is happening.

On 17 September 2026, the US government’s CAISI described Z.ai’s GLM-5.3 as the most cyber-capable open-weight model it had evaluated. Across its cyber benchmarks, it estimated the model was roughly four months behind the current US frontier. That is a particular cyber assessment, not a measurement of how far China sits behind America across all AI.

It still tells us why assuming dangerous capability can stay behind a handful of American APIs is a poor starting point.

Now, a pause could buy time before an even more capable model becomes available. Existing models being out there does not make that argument disappear. But show us what the pause buys. Who observes it, what work gets done during it and what has to improve before development resumes? How does it account for actors who carry on regardless?

A company pausing because its own security is not ready makes sense to me. Giving a group of companies influence over everyone else’s pace is a much bigger proposition.

Dario’s framework recognises the geopolitical problem. It says coordinated pacing would have to preserve the US lead and eventually require global coordination. China sees AI as strategic infrastructure. So does America. These governments care about economic power, defence and scientific leadership. Their interests will not neatly fall into line because a handful of CEOs want to slow the race.

That horse bolted a long time ago. Slowing the next model does not take the existing ones out of hostile hands. Whatever we decide about future releases, we still have to defend against what is here.

Open source is attacking the moat

Now look at what is happening to the business underneath all this.

Vercel’s September 2026 production index says open-weight models went from 7% of its AI Gateway token volume in December 2025 to 56% in August 2026. That is Vercel’s traffic, not the whole AI market. In August 2026, those models accounted for only 14% of spend. More than half the tokens, but a much smaller share of the money. You can see why that would make the labs uncomfortable.

Customers are finding cheaper ways to get the work done. Being top of a benchmark is lovely, but if a cheaper model does what I need, why would I pay you more?

Then NVIDIA agreed to buy Hugging Face for $12.93 billion. Hugging Face says more than 18 million developers use the platform, with over three million models available and more than 200,000 companies using it. NVIDIA says it intends to keep it open and model-neutral.

Of course that suits NVIDIA. Take an open model, adapt it to your business and run it on NVIDIA hardware. Jensen still gets paid. You do not have to buy your intelligence from OpenAI or Anthropic for that arrangement to work very nicely for him.

For the frontier labs, it is a rather different story.

If I am building an application, I want intelligence that does the job properly at a price that makes sense. I might use one model for research and another for writing code. I might want to run something inside my own infrastructure because the information belongs to my customers. And when a better option comes along, I want to be able to change.

I am buying a component for my business. I am not handing you the business.

The labs are spending enormous amounts to stay ahead while open models make their intelligence easier to replace. Meanwhile, application companies are building relationships with customers and collecting information from the actual work being done.

That last bit matters a lot.

A legal platform handling thousands of client matters can build up knowledge of what worked, what needed correcting and what the client accepted, where it has permission to use that information. You cannot just scrape that off the internet. It comes from being inside the work.

The same applies in a factory or a healthcare business. Knowing what actually happened, and whether a decision worked, is valuable. The company running the application may have that information. The company supplying the model may not.

And if the application changes the model underneath without the customer even noticing, who owns the relationship? Who has the information that makes the product better?

The labs can spend billions building extraordinary intelligence and still end up as a supplier to someone who owns the customer and can replace them.

That is why I think they are under serious commercial pressure. Having the smartest model does not mean you own the business.

So when we discuss restrictions on open weights, look at who benefits. Make open models harder to distribute or use and you make the closed providers harder to replace. You are helping protect the very businesses asking for those restrictions.

A particular restriction may have a valid safety justification. Make the case and let us examine it. But the commercial benefit exists whether protecting those companies was the intention or not.

Pretending it does not exist is taking us for idiots.

Build the boring defences

I am not suggesting we ignore the danger. I am suggesting we start fixing the things that let these incidents happen.

If an agent does not need the internet, do not give it access. And check that you have actually cut it off. OpenAI’s agents were supposed to be isolated, but Artifactory could fetch things from outside. It did not matter that the agent had no browser if it could get another part of the system to make the request for it.

Give it access to what it needs for the job, for as long as it needs it. Reading one database does not require administrator access to five others. And if supposedly isolated agents can all write to the same place, do not act surprised when they find a way to leave messages for one another.

We also need to stop treating “keep going until you succeed” as an instruction with no downside. Some of the OpenAI tasks were impossible. Giving the agents more time to think did not make them possible. It gave them more opportunities to look for another way to get the score.

Sometimes the correct answer is “I cannot do this”. We should want the agent to tell us the task is broken or ask for help. OpenAI now explicitly says that is how these situations should be handled. An agent that never gives up sounds wonderful until it starts doing things it should never have been allowed to do.

Then there is the record of what happened. If the agent can change that record, how are we supposed to trust it when something goes wrong? We have already seen agents try to interfere with their transcripts because they thought it would help their score.

The defendant does not run the CCTV system.

Keep the logs somewhere the agent cannot alter or delete them. That should hardly need saying.

The same goes for permission to act. A model deciding that a $10 million transfer would help finish the job does not make that transfer authorised. It should have to pass a separate check, based on rules it cannot rewrite because they have become inconvenient.

I do not mean someone has to sit there approving every email. Give it permission to do routine work within agreed limits. But deleting a production database or moving someone else’s money needs controls that reflect what is at stake. The system trying to complete the task cannot simply grant itself whatever authority it decides it needs.

And yes, keep working on how the models behave. We want them to recognise a dangerous instruction, refuse it and stop when something looks wrong. OpenAI’s reported reduction of more than 100 times with its production harness and system prompt is a good reason to take that work seriously. It does not tell us that any one safeguard is enough.

Make the agent less likely to attempt something harmful, and build the surrounding system to block it if it does. I want both. I certainly do not want the security of my business resting on whether the model decides to behave itself today.

AI does not put you above the law

I think Jensen Huang is right about this. Before we start writing an entirely new set of rules, how about applying the ones we already have?

And yes, NVIDIA benefits from more AI being built and used. We know that. Jensen has a commercial interest, just as Sam and Dario do. It does not make his point wrong.

If your system breaks into someone else’s computer, we already have laws dealing with unauthorised access. If your product causes damage, explain what happened and answer for your part in it. If you promise a customer that their data will stay inside an agreed boundary and your agent sends it somewhere else, putting an LLM in the product does not make the contract disappear.

Huang put it plainly: “Don’t let this doomsday narrative cause somebody to relieve them of the laws that currently exist.”

Exactly. These companies are selling products, signing enterprise contracts and asking customers to trust them with access to their businesses. You cannot expect to be treated as a serious enterprise supplier when you send the invoice, then as an experiment when something goes wrong.

You built the system. You decided what it could access. You gave it the credentials and the ability to keep trying. So explain how it ended up somewhere it was never supposed to be.

Saying it found a method you had not anticipated is hardly a satisfactory answer. Finding methods we had not anticipated is part of what you sold us. That is why people are paying for it.

“The AI did it” cannot become a get-out-of-jail-free card.

There may be gaps in existing law. Fine. Identify them and fill them. If a dangerous capability needs specific controls, explain what they are and why they are needed. If companies are failing to report serious incidents, deal with that. I have no objection to rules that address an actual problem.

What I object to is the suggestion that AI is so extraordinary that we have to rethink everything before anyone can be held accountable for anything.

Start with what happened. What did you promise the customer? What access did you give the agent? Where were the controls, and why did they fail?

Talk about what AI might do in five years does not answer for what your product did last week.

You want us to trust these systems enough to let them into our businesses. You want them doing the work, making decisions and taking action, because that is worth more money than answering questions on a screen.

Then stand behind what you are selling.

You want the revenue? Take the liability too.

Better models are also part of the defence

The models already out there do not disappear because OpenAI postpones its next training run. Whoever is using them to attack our systems can carry on. I want the people defending those systems to have the best tools we can give them.

OpenAI has already temporarily slowed frontier development to improve security around its increasingly capable models. Greg Brockman also says it moved roughly 25% of its production engineering team onto security, set Astra to work finding vulnerabilities in its own systems and used it to help fix them.

Good. That is exactly what I want to see. If you have built something that can find holes nobody else spotted, point it at your own infrastructure before someone else does.

And keep doing it. Every substantial improvement in capability should mean another round of testing. Give your security team the stronger model and let them find the forgotten service or the credentials somebody left exposed. Fix what they find, check that the fix works and go again.

It will not find everything. Nor can we assume every improvement benefits defenders more than attackers. But here is a concrete use of stronger models that belongs in the safety argument: finding weaknesses and helping close them.

So give OpenAI credit for doing that. They recognised that their security needed work, slowed their own development and put people and models on the problem. They did not need everyone else to stop before they could start fixing it.

If a stronger model can help us find and close the holes, I bloody well want it working for the defence. Any proposal to slow development should explain how it preserves that work as well as how it reduces the threat.

Safety is real. So is the commercial pressure.

Now we get to the bit that pisses me off.

I believe there is a safety problem. I have just spent several sections explaining why. But I am also supposed to look at the proposed solution and ignore the commercial interests of the companies proposing it? Why would I do that?

Look at the businesses they are running. Open models make their intelligence easier to replace, application companies own customer relationships they want, and the compute bills keep coming. Their valuations depend on making extraordinary amounts of money in the future.

Agreeing to slow the race together has some obvious attractions.

Dario’s proposal explicitly talks about giving frontier developers more time for safety work without sacrificing commercial advantage. That wording matters. His framework goes on to call for industry-wide coordination and, for some discussions, government mediation or narrow antitrust relief.

If Anthropic slows down and OpenAI does not, Dario risks losing ground. Sam has the same problem in reverse. Agree to slow together and neither has to bear the full competitive cost of making that decision alone. It does not guarantee their positions, but it makes slowing down a considerably easier commercial choice.

Anyone who remembers Chuck Prince and the financial crisis should recognise this. While the music is playing, everyone feels they have to keep dancing. You can see the risk building and still feel unable to stop because your competitors are making money and your investors want to know why you are not.

Now imagine being able to agree when everyone sits down, with the government helping arrange it.

You can see why that would appeal to the people already on the dance floor.

Call it safety coordination. I still see competitors asking for permission to coordinate how they compete, and I think that starts looking uncomfortably like a government-backed cartel. I am not claiming they have broken antitrust law. They are asking for relief from it to enable some of these discussions. That deserves scrutiny, however serious the safety concerns are.

The FTC chairman has raised the same concern, saying his “alarm bells” go off when established companies ask for more regulation alongside antitrust exemptions.

And then there is what the rules do to anyone trying to catch up.

Anthropic’s Advanced AI Framework proposes extra obligations for developers above specified training-compute levels who also cross substantial revenue or R&D thresholds. Testing dangerous capabilities and reporting serious incidents sound perfectly reasonable to me. We should expect proper security from these companies.

But look at who can afford the machinery needed to comply.

Anthropic already has the lawyers, security teams and people dealing with government. A challenger crossing those thresholds has to build that operation while also finding the money to train models and win customers. The same requirement can be manageable for the company already worth a fortune and a serious obstacle for the one trying to compete with it.

A rule can improve safety and still protect the companies already in front. Both effects deserve attention, just as they do when restrictions make open alternatives harder to use.

I am not prepared to ignore who benefits just because someone puts “safety” on the front of the proposal.

Then there is the public-market pressure

And look at the timing.

Anthropic raised $65 billion in May 2026, at a $965 billion post-money valuation, then confidentially filed its draft S-1 on 1 June 2026. OpenAI announced its own confidential S-1 submission on 8 June 2026.

OpenAI is projected to burn roughly $278 billion in cash between 2026 and 2030, while investors have discussed valuations around $1.2 trillion.

At some point, someone has to explain how those numbers turn into a return for shareholders. Being able to build extraordinary technology does not, on its own, answer that question.

Public investors will want to know what customers will pay, what it costs to serve them and how much more money the company needs before it can stand on its own feet. They will also want to know why those customers will stay when cheaper alternatives become good enough.

Now put the safety argument alongside that.

If the technology is presented as both essential to our future and too dangerous to get wrong, enormous spending becomes easier to defend. Another capital raise can be framed as necessary. A delay to an IPO can be explained as responsible caution. Rules that make it harder for competitors to catch up can be justified as protecting everyone.

Those arguments may have merit. They are also very useful to companies with huge bills and equally huge expectations hanging over them.

The financial pressure does not prove anyone invented the danger. It does give us a reason to examine whether the proposed response also protects a business model under pressure. I think it does.

We are being asked to accept restrictions that could shape who gets to build AI and who gets paid when we use it. Of course the economics belong in that discussion. Calling it safety does not put it beyond scrutiny.

Regulation should follow the risk, not protect the club

Of course government has a role. We cannot leave hospitals, power networks and other critical infrastructure to defend themselves against increasingly capable attacks with whatever is left in the IT budget. Serious incidents need to be reported, and companies working with models that can find unknown security holes need security that reflects what they are handling.

If you give an autonomous system access to someone else’s money, data or business, you should have to answer for how you manage that access. That seems fairly obvious to me.

Start with what Huang is saying. Apply the laws we already have. Look at what happened, who was responsible and whether they met their obligations. Where those laws leave a genuine gap, explain it and fill it. Show us the risk and how the proposed rule would address it.

What I struggle with is starting by giving the biggest companies exemptions from rules designed to stop competitors coordinating with one another. They are asking for considerable influence over the future of this technology. We should want considerably more than an assurance that it is all for our own good.

And they already have decisions they can make inside their own businesses.

If Dario thinks Anthropic is moving too fast, he can slow it down. Sam can do the same at OpenAI. OpenAI has already shown that it can pause development when it believes its security needs more work. Nobody had to stop the rest of the industry for that to happen.

They can put more money into securing their systems and limit what their agents can access while that work is being done. Bring in independent people to test the boundaries. Share what they find so others can fix the same weaknesses.

And price the product properly. If it costs more to run safely, charge for it. If taking responsibility for failures changes the economics, then those are the economics of the business you chose to build.

I want these companies to succeed. I also want them to take responsibility for what they put into the world. They have plenty of work they can get on with before asking for control over what everyone else is allowed to build.

The keys were never the intelligence

I am still an AI optimist. There are founders, scientists and people with good ideas who have never had access to the resources needed to do anything with them. Making intelligence cheap enough for those people to use is something worth building towards.

Nothing in these incidents has changed that for me. What they have changed is how much confidence I have in the companies connecting this intelligence to the world, and how willing I am to accept their proposed solution.

I take the warnings seriously. I also think commercial pressure is a major reason why slowing down together has become such an attractive idea. That is my judgement, based on the pressures on these businesses and what their proposals would do for them. The safety concerns do not make those benefits disappear.

These companies chose to race. They raised the money, committed to the spending and sold us on what their systems would be able to do. Much of that promise involved agents doing more work with less supervision, because that is something customers will pay for.

If they now believe they are moving faster than they can safely manage, they should slow down. I have no objection to that. I object to turning their commercial commitments and engineering failures into a claim over what everyone else is allowed to build.

We are not putting intelligence back in a box, and I do not want us to. I do want us to be much more careful about what we let it do. Being able to work out a course of action should not automatically give a system permission to carry it out.

We decide which accounts it can access, what it can spend and whether it can change or delete something. We build the machinery that lets it keep working while we are asleep. Those decisions deserve far more attention than they have been getting.

In 2024, I wrote that we had given AI the keys to our kingdom. Look at the incidents in this article and look at what let them happen.

The keys are the permissions, credentials, tools and access we give it. We handed those over.

Now the people who handed them over want control of the locksmith.

Do not give them that as well.

Sources

Go take a look, For the ❤️ of startups

Scout - who just got funded
Newly funded startups, pre-seed to Series B, as the rounds happen.

Raise - who’s deploying
New funds with fresh capital and cheques to write.
Wire - what changed, what matters
The intelligence feed that filters the noise, with the “so what” attached.

If you have not joined the Fusion42 Community on Telegram —
it is probably time to do so.

For the ❤️ of Startups

Please let me know what you think of it, love a feedback loop 🙏🏼

❤️ Love it, keep them coming!
👍 Great, always a good read.
😊 Enjoyable, but let’s see some variety.
👎 Okay, seen better.
🛑 Get a different job.

Subscribe below and follow me on LinkedIn or Twitter to never miss an update.

WHY DON’T YOU WANT A DATA CENTRE IN YOUR TOWN WHEN YOU DON’T EVEN KNOW WHAT THE FUCK IT REALLY IS?

DeReK WaTSoN — Wed, 09 Sep 2026 10:30:01 GMT

Index

Why people are protesting
We are trying to build the industry again
Inside an AI data centre
The supply chain underneath it
Everyone is building, and everyone is paying everyone
The global compute race
The real problem is power
The bill runs into trillions
Heat, cooling and water
Why locals are getting pissed off
Governments are changing the rules
Compute starts following energy
Elon wants to take it into space

Why people are protesting

My view on data centres is pretty simple. AI is going to need a lot more compute. That means a lot more data centres. Fine. Build them.

What I hadn’t done was add it all up.

Then the protests started catching my eye. In July there were 142 demonstrations across 42 US states, and a Reuters/Ipsos poll found only 14% of Americans would welcome a data centre in their own community.

Texas was the one that really got me. Cheap energy, cheap land, light regulation. Exactly the sort of place you would expect to want these things. Yet Republican politicians there are now talking about restricting them, removing tax breaks and making developers pick up more of the costs. Only 30% of Texans in one recent poll supported having one nearby.

I knew why data centres were needed. I just hadn’t worked through the physical scale of a 1GW campus: the power connection, substations, transmission, cooling, water, construction, grid upgrades, permanent jobs once the builders leave and who pays for all of it.

So I did the sums.

At roughly the same time Elon Musk was talking about putting some of the compute in space. SpaceX eventually wants to launch 100GW of AI compute capacity on solar-powered satellites every year. At full scale that means thousands of launches and roughly one million tonnes of hardware going into orbit annually.

I’ll come back to that.

I still want the data centres built. I just don’t think “build them” is the end of the argument once you see the numbers.

We are trying to build the industry again

The first number that really changed the way I looked at this was 103GW. That is roughly how much data-centre capacity exists globally today. JLL expects it to be around 200GW by 2030.

So in the next few years we are trying to build almost as much data-centre capacity as the world has built in total up to now.

McKinsey gets to a slightly bigger number using a different methodology: around 219GW by 2030, with 156GW coming from AI. The exact forecast will move as chips get better, models become more efficient and projects get cancelled or delayed. But whether the eventual number is 200GW or 219GW does not change what is happening. We are adding an enormous industrial load in a very short period of time.

The electricity numbers make the scale easier to understand. Gartner has worldwide data-centre power demand rising from roughly 104GW in 2025 to 132GW in 2026 and 290GW by 2030. Electricity consumption goes from 447TWh in 2025 to 565TWh in 2026, 702TWh in 2027 and more than 1,200TWh by 2030. From 2027, AI-optimised servers are expected to use more electricity than conventional servers.

One thing worth clearing up because these GW numbers get thrown around rather casually. A 1GW data-centre campus is basically a site designed around access to roughly one gigawatt of electrical power. So when you see forecasts approaching 300GW, think hundreds of very large industrial campuses all drawing power at the same time.

And this is already showing up in the real economy. US data-centre construction spending hit an annualised $75bn in July, up 57% in a year and roughly 717% since the start of 2021. Since the end of 2023 alone, the annualised run-rate has risen by about $51bn.

That $75bn does not include the GPUs and servers. It is the buildings, site work and equipment going into them.

Housing and commercial construction have their own problems, so I would not pretend data centres caused the weakness elsewhere. But construction workers, transformers, electrical equipment, land and capital are finite.

A lot more of all of them is now being pulled towards AI.

Inside an AI data centre

I’m leaving the chips mostly out of this because Nvidia, TSMC, HBM and advanced packaging deserve their own piece. Assume the GPUs have arrived.

The rest of the building exists to keep them running.

You need racks, servers, storage, networking and fibre. Then comes the electrical side: substations bringing power onto the site, transformers changing voltage, switchgear controlling and protecting the system, UPS equipment and batteries dealing with short interruptions, backup generators for longer ones, plus busbars and a lot of heavy cabling moving electricity around the building.

Cooling is becoming a much bigger part of the design as AI racks get denser. Air cooling only gets you so far. Higher-density systems increasingly use cold plates sitting directly against the processors, with pumps, coolant distribution units, heat exchangers and chillers moving the heat out of the building. Depending on the system and location, cooling towers and water may come into it as well.

Then there is everything outside the server room: land, roads, transmission connections, concrete, steel, copper, construction crews, electricians, engineers and all the other fairly unglamorous bits needed before anyone can run a model.

This is how Schneider Electric, Eaton, ABB, Siemens Energy, GE Vernova, Mitsubishi Heavy Industries, Vertiv, Delta Electronics, Caterpillar and Cummins end up sitting underneath the AI boom.
A surprising amount of artificial intelligence depends on some very old-fashioned engineering.

The supply chain underneath it

Once you look at everything inside an AI data centre, another problem appears. Somebody has to make all of it, and the supply chain stretches far beyond Silicon Valley.

Start with the chips. The most advanced AI processors still depend heavily on TSMC in Taiwan, while the high-bandwidth memory beside them comes mainly from SK Hynix, Samsung and Micron. Then there is advanced packaging such as CoWoS, substrates, PCBs and the networking hardware needed to connect thousands of accelerators. A shortage in any one of those can slow the whole system down.

Move outside the rack and it gets much more industrial. A large data centre needs a lot of copper in transformers, cables, busbars, switchgear, cooling systems and the grid connection itself. Much of that copper is mined in places such as Chile, Peru and the Democratic Republic of Congo, while China plays a huge role in refining it.

Transformers are an even more immediate bottleneck. They are large, often built to order and can have lead times well beyond a year. Switchgear, circuit breakers and other high-voltage equipment are tight too. You can have the land, the building and billions of dollars of GPUs sitting inside it and still be waiting for the electrical kit needed to turn the bloody thing on.

Power generation has the same problem. AI companies want electricity faster than grids can add it, which has pushed gas turbines back into demand. GE Vernova, Siemens Energy and Mitsubishi Heavy Industries cannot simply double production overnight. These are huge machines built in specialised factories with their own supply chains for steel, castings, blades, generators and control systems.

Then add batteries. Modern data centres increasingly use large lithium-ion UPS systems to bridge interruptions before generators or other backup systems take over. Much of the LFP battery supply chain sits in China, from cathode materials through to finished cells.

Cooling adds another layer. Higher-density racks need cold plates, pumps, valves, coolant distribution units, heat exchangers, chillers, specialist fluids and kilometres of pipework. None of this looks particularly exciting next to an Nvidia GPU, but a failed pump or delayed heat exchanger can leave the GPU doing absolutely nothing.

And underneath all of that sit concrete, structural steel, electrical steel, fibre, optical equipment, connectors, generators, cables and a lot of skilled labour.

We talk about AI as software. Building the infrastructure behind it looks increasingly like heavy industry, with dependencies on mines, smelters, semiconductor fabs, transformer factories, turbine plants and electrical equipment manufacturers spread across the world.

The bottleneck will not always be the cleverest component. Quite often it will be the boring thing everyone suddenly needs at the same time.

You can have the GPUs and still not be able to turn them on.

Everyone is building, and everyone is paying everyone

This was the point where I started losing track.

The hyperscalers are already spending at extraordinary levels. AWS, Microsoft, Google, Meta and Oracle are all building huge amounts of data-centre infrastructure, while the AI companies themselves have increasingly started securing dedicated capacity rather than simply renting it.

OpenAI’s Stargate programme is a good example. It originally targeted 10GW of US AI infrastructure by 2029, and OpenAI said in April that it had already passed that figure. It is not pouring all the concrete itself. The capacity is being assembled through a network of infrastructure and cloud partners.

Anthropic is doing the same from both directions. It announced $50bn of custom-built US computing infrastructure with Fluidstack, with sites in Texas and New York, while also signing very large compute agreements with AWS, Google and others. Tesla has built its own dedicated training infrastructure at Giga Texas, where Cortex 1 has more than 90MW and Cortex 2 more than 115MW supporting autonomy and Optimus. SpaceXAI has Colossus in Memphis and is separately building a 1.2GW permanent power plant to support its terrestrial compute expansion. That is power generation for compute on Earth, not part of the orbital system.

Then you have the new infrastructure layer sitting underneath all of this. CoreWeave, Fluidstack, Lambda, IREN, Nscale and others are raising billions to secure land, power and hardware, then selling the resulting capacity back to AI companies. Nvidia sits in the middle selling the GPUs while also investing across much of the same ecosystem.

The financing has started to get equally strange. Oracle ended FY2026 with $638bn of remaining performance obligations, and some customers are now prepaying for GPUs or even supplying hardware themselves so Oracle does not have to finance all of it. Meanwhile the hyperscalers’ own capex numbers have become absurd: Amazon expects around $200bn in 2026, Meta $130bn–$145bn, with Alphabet and Microsoft spending at similarly enormous levels.

At that point the headline numbers start to lose meaning, because the same underlying capacity can show up again and again in different announcements. The AI company announces a compute contract. The cloud provider reports the backlog. Nvidia reports the hardware demand. The infrastructure company raises money against the contract. A government may then present the same project as inward investment.

Everyone is paying everyone.

That is why I stopped trying to add the press releases together.

The physical build is much harder to bullshit. JLL’s model implies roughly another 100GW of data-centre capacity by 2030, with required investment of up to $3tn. McKinsey, using a broader definition that includes more of the infrastructure and hardware stack, gets to roughly $6.7tn, including about $5.2tn for AI.

And the spending does not stop once the buildings are finished. The concrete may last decades. The GPUs, servers and networking equipment do not. We will be replacing the expensive kit inside these buildings while still constructing the next generation of them.

There is also clearly a wall of corporate FOMO in the background. Nobody running a major technology company wants to wake up in three years and discover that a rival locked up the land, power, chips and grid connections while they were waiting for the economics to become obvious.

That creates a very powerful incentive to build first and work some of the details out later.

The global compute race

Individual projects are now enormous. Meta’s Hyperion campus in Louisiana is being expanded towards 5GW of compute capacity, with more than $50bn being invested. Meta expects around 7,500 jobs at peak construction and roughly 1,000 permanent roles once the site is operating.

Five gigawatts in one place gives you some idea how far this has moved from the old idea of a data centre.

Governments are now entering the same race. Britain says it needs at least 6GW of AI-capable data-centre capacity by 2030, roughly three times what it has today, with several large AI Growth Zones planned and at least one site capable of exceeding 1GW. Saudi Arabia’s HUMAIN programme already has production AI compute running, with its AMD-Cisco venture beginning another 250MW of deployment from 2027 and targeting 1GW by 2030. Stargate UAE is planned as a 1GW AI cluster in Abu Dhabi, with the first 200MW expected online in 2026.

China is operating at a much larger scale. It had roughly 32GW of installed data-centre capacity at the end of 2025, with current projects potentially taking that above 60GW by 2030. India is starting from a smaller base but moving quickly, with its Economic Survey citing an estimate of around 1.4GW in 2025 rising towards 8GW by 2030. Europe is building sovereign AI capacity through its own programmes, while Japan is expanding too, although I still haven’t found a clean national GW target I trust enough to use.

What started as a hyperscaler capex boom is turning into strategic national infrastructure.

That makes sense. If AI becomes as important to economies as governments now expect, countries are hardly going to be comfortable relying entirely on somebody else’s compute capacity.

But every one of these national ambitions runs into the same physical constraint: electricity.

The real problem is power

Once an individual campus wants 500MW, 1GW or even 5GW, electricity stops being a utility bill and starts determining whether the project can exist at all.

That is becoming a serious problem because many grids simply cannot provide another gigawatt on the timetable AI companies want. US utilities have already received more than 700GW of large-load requests, more than ten times estimated current US data-centre demand. Texas has started auditing applications because regulators believe a substantial share is duplicated, speculative or attached to projects that may never be financed. The industry has started calling this ghost demand.

That matters whenever somebody announces another 2GW campus. Asking the grid for 2GW is easy. Financing the project, building it and actually energising 2GW are very different things. Even after stripping out the speculative requests, however, the underlying problem remains: AI companies need enormous amounts of reliable power, and they need it quickly.

The timing is awkward. New nuclear plants and SMRs feature heavily in AI infrastructure plans, but a reactor that arrives well into the 2030s does very little for a data centre that needs 500MW in 2028. In the near term, the options are more practical: gas generation, existing nuclear plants staying open longer or being restarted, whatever grid capacity is available, and behind-the-meter generation where it is not.

That is why companies such as GE Vernova, Siemens Energy and Mitsubishi Heavy Industries have suddenly become part of the AI story. They manufacture the gas turbines and other equipment capable of delivering hundreds of megawatts of dispatchable power. Developers are increasingly looking at building generation alongside the data centre rather than spending years waiting for the grid. Google has moved upstream as well, buying energy-infrastructure developer Intersect and signing large geothermal agreements, while other hyperscalers are pursuing nuclear and dedicated generation deals of their own.

The power station is only half the problem. Electricity still has to get from the generator to the GPUs, and that means substations, switchgear, circuit breakers and transformers. Large transformers in particular have long manufacturing cycles, so a finished data centre can have billions of dollars of computing hardware sitting inside it and still be waiting for the equipment needed to switch it on.

There is something rather wonderful about that.

Some of the most advanced technology humans have ever built can still be held up by a giant lump of copper and steel based on engineering that has existed for more than a century.

The bill runs into trillions

Putting a clean price on all of this is harder than it looks because people count different things. Land, buildings, grid connections, dedicated generation, cooling, GPUs, networking and financing can all end up under the heading of AI infrastructure.

JLL estimates that adding roughly another 100GW of global data-centre capacity by 2030 could require up to $3tn, including around $1.2tn of data-centre real estate. McKinsey counts more of the technology sitting inside and around the buildings and gets to roughly $6.7tn, with about $5.2tn linked to AI infrastructure.

Either way, we are talking about trillions of dollars being spent in a few years.

And unlike a railway, a lot of the expensive kit does not last very long. The building may still be standing in thirty years, but the GPUs, servers and networking equipment inside it will have been replaced several times by then.

That changes the economics quite a bit. The industry will still be building the first wave of new capacity while already replacing some of the most expensive equipment installed at the start of it.

So the capex does not simply peak when the buildings are finished.

A large part of this infrastructure has to be built, upgraded and rebuilt at the same time.

Heat, cooling and water

All that electricity eventually becomes heat, and getting rid of it is becoming a much bigger engineering problem as AI racks get denser.

Traditional data centres could rely heavily on air cooling. At higher rack densities, simply pushing more cold air through the room stops being enough, which is why liquid cooling has become such an important part of the build-out. Cold plates sit against the processors, pumps move coolant around the rack, heat exchangers pull the heat away and chillers or cooling towers deal with it further downstream.

That is how I ended up reading about pumps and pipework while supposedly researching artificial intelligence.

Water is more complicated because there is no honest universal number. Some cooling systems consume a lot, others use very little, and climate makes a big difference. But the issue is clearly becoming material enough for governments to regulate it. Minnesota, for example, has introduced special permitting requirements for projects proposing more than 100 million gallons of additional consumptive water use per year, with regulators required to consider conservation, recycling, reclaimed water and closed-loop systems.

There is also water used upstream by some forms of electricity generation, so the total impact can extend well beyond the cooling system itself. For a town already worried about water availability, a new 500MW or 1GW industrial customer is obviously going to attract scrutiny.

And water is only part of what neighbours actually experience. Cooling equipment runs continuously, and fans, chillers and other machinery can produce a persistent low-frequency hum. Hyperscale sites can also have large numbers of diesel backup generators behind the fence. They do not run continuously, but they still need testing and create noise and emissions when they do.

These sound like minor issues next to trillion-dollar AI forecasts.

They feel rather less minor when you live next door.

Why locals are getting pissed off

A data-centre announcement sounds fantastic when it first arrives. Billions of dollars of investment, thousands of construction jobs, new tax revenue and a place in the AI economy. Those benefits are real. We need the infrastructure, domestic compute capacity has strategic value, and the economic activity eventually built on top of AI could be worth far more than the facilities themselves.

The argument changes when residents start looking at what the project requires from the place hosting it. A large campus can consume hundreds of megawatts around the clock, require new substations and transmission lines, draw significant amounts of water and run cooling equipment continuously. Backup generation has to sit onsite as well, bringing its own noise and emissions.

Then there is the jobs question. Meta’s Louisiana campus is a useful example. More than $50bn of investment and around 7,500 jobs at peak construction eventually becomes roughly 1,000 permanent roles. A thousand jobs can still matter enormously to a local economy, but the calculation looks different when they sit alongside a 5GW industrial load and tens of billions of dollars of infrastructure.

That is why communities start asking harder questions about tax incentives, electricity prices, water, noise and who pays for the grid upgrades. The political fight starts when people begin to wonder whether the deal works as well for the place hosting the data centre as it does for the company using it.

Governments are changing the rules

Governments are starting to respond because data centres are becoming too large to treat like ordinary commercial developments.

South Dakota has gone as far as passing something literally called the Data Center Bill of Rights for Citizens. It applies to facilities with peak demand above 10MW and follows a fairly sensible principle: if a data centre creates extra costs for the electricity system, those costs should sit with the data centre rather than being spread across everyone else. Developers also have to disclose expected water use, while local authorities retain the right to restrict or prohibit projects.

Ireland has a different problem because data centres already consumed 22% of the country’s electricity in 2024, with that share potentially reaching 31% by 2034. New projects seeking grid connections now have to provide enough generation or storage to match their maximum requested demand. Germany has focused on efficiency, requiring new data centres from July 2026 to achieve a PUE of 1.2 or better and reuse an increasing share of the heat they produce.

The US is heading in the same direction. Virginia, already home to one of the world’s largest concentrations of data centres, decided in July that some transmission costs should be charged directly to the customers responsible for creating them. PJM, which runs much of the eastern US electricity market, is also developing rules that favour large new loads which bring generation with them. Projects without enough power behind them can face restrictions when the grid gets tight.

That is broadly where I think the rules should go.

Build the data centres, but make the economics honest. If a project needs a new substation, transmission line or power plant, the people building it should pay the costs created by it. If it needs huge amounts of water, tell the community how much and where it is coming from. If the local grid cannot support another gigawatt, build more generation or put the facility somewhere that can.

That is what serious infrastructure planning looks like.

Power availability is starting to redraw the geography of AI.

Software companies traditionally clustered around engineers, universities, customers and capital. Large AI training clusters have different priorities. Give them enough fibre and reliable electricity and much of the workload can move. Training is especially mobile because the GPUs do not care whether they are five miles or five hundred miles from the people building the model.

Inference is a bit different. Batch workloads can also move towards cheap power, while applications where latency matters benefit from being closer to the people or machines using them. So we are likely to end up with very large training clusters in places with abundant electricity, alongside more distributed inference capacity closer to users.

That suddenly makes some countries much more interesting. France has nuclear. The Nordics and Canada have hydro. China has deliberately pushed heavy compute towards western regions where land and electricity are easier to find. Saudi Arabia and the UAE have energy, capital and governments capable of making large infrastructure decisions quickly.

The Gulf comes with its own engineering problems. Saudi Arabia is bloody hot, so cooling is harder and water matters. Transmission and fibre still matter as well. But once compute starts behaving like heavy industry, the strategy makes sense. Saudi’s HUMAIN programme is targeting 1GW by 2030 through its AMD-Cisco venture, while Stargate UAE is planned as another 1GW cluster in Abu Dhabi.

China is doing the same thing at a much larger scale. Its data-centre capacity could move from roughly 32GW in 2025 to more than 60GW by 2030, with more of the heavy compute going to regions where the power and land already exist.

AI is beginning to follow the geography of heavy industry.

When a training campus needs 500MW or a gigawatt, access to electricity matters rather more than having the software engineers nearby.

Elon wants to take it into space

This brings me back to Elon, because after looking at all the constraints on Earth, SpaceX’s Starmind project starts to look a lot less mad.

AI needs enormous amounts of electricity, and finding hundreds of megawatts of reliable new power on Earth is getting harder. Grid connections can take years. New transmission can take longer. Land and water become political very quickly. SpaceX’s idea is to move some of the compute closer to an energy source that is already there: the Sun.

Its first design, AI1, is a modular AI-compute satellite using solar technology derived from Starlink. SpaceX says each satellite would carry a 210kW solar array and operate in sun-synchronous orbit, giving it long periods of solar exposure without weather or atmospheric losses. The compute modules are chip-vendor agnostic, while high-speed optical links would connect the satellites to one another and feed results back through Starlink.

This has moved beyond a slide deck. SpaceX is building a Gigasat Factory in Bastrop that it says could start producing and deploying thousands of AI satellites from late 2027. Terafab, being developed with Tesla, is intended to supply the advanced AI chips needed further up the stack.

SpaceX identifies three big constraints of its own: getting enough mass into orbit, generating enough power and producing enough AI chips. Starship is supposed to deal with the first, orbital solar with the second and Terafab with the third.

Then the numbers get properly ridiculous.

SpaceX’s June 2026 prospectus says the long-term ambition is to launch 100GW of AI compute capacity every year. At that scale it expects thousands of launches and roughly one million tonnes of hardware delivered into orbit annually.

The idea becomes easier to understand once you compare it with what is happening on Earth. A large terrestrial data centre can spend years waiting for grid capacity, substations, transmission lines and planning approval before anyone switches on the GPUs. SpaceX is effectively asking whether it eventually becomes easier to take some of the compute to the energy rather than keep dragging ever larger amounts of energy towards the compute.

Cooling is part of the case as well. Terrestrial data centres spend a lot of energy running pumps, chillers, fans and cooling towers. SpaceX says orbital compute can remove much of that overhead because heat can be radiated directly into space.

The heat still has to go somewhere. In a vacuum that means rejecting it through radiators, and at hundreds of kilowatts per satellite thermal management becomes a serious engineering problem of its own.

The network is another one. These satellites are moving at orbital speed, so laser links have to stay locked while traffic is continually handed from one satellite to another. Link continuity and handover matter enormously. A distributed orbital AI cluster cannot afford links dropping whenever one satellite disappears over the horizon.

Then add radiation, collision avoidance, replacement cycles, servicing, orbital debris and the economics of launching enough hardware to make the whole thing worthwhile.

SpaceX has spent the last twenty years turning problems like these into industrial systems, which is why I find Starmind interesting. It is trying to control most of the stack itself: Terafab makes the chips, the Gigasat Factory builds the satellites, Starship gets them into orbit, Starmind runs the compute and Starlink connects it back to Earth.

Whether that works economically at anything approaching 100GW a year is still a huge question. The industrial logic, though, is easy enough to see.

If power becomes one of the main constraints on AI, putting some compute next to a near-continuous source of solar energy starts to make sense.

That brings me back to where I started.

I believe in what this technology can do, and I think we are going to need vastly more compute than we have today. Some of the rush is rational. Some of it is corporate FOMO, with trillion-dollar companies racing to secure land, power, chips and grid capacity before their competitors do.

We need to build the bloody thing properly.

Put data centres where the power and infrastructure can support them. Build generation alongside them where necessary. Be honest about water, noise and emissions. And when a project creates the need for a new substation or transmission line, the company building it should pay rather than quietly handing the bill to everyone living nearby.

I’m all for building the compute.

We just shouldn’t fuck over the people living next door to do it.

And if Earth really does become the constraint, perhaps some of that compute ends up somewhere else entirely.

Go take a and look, For the ❤️ of startups

Scout - who just got funded
Newly funded startups, pre-seed to Series B, as the rounds happen.

Raise - who’s deploying
New funds with fresh capital and cheques to write.
Wire - what changed, what matters
The intelligence feed that filters the noise, with the “so what” attached.

If you have not joined the Fusion42 Community on Telegram —
it is probably time to do so.

For the ❤️ of Startups

Please let me know what you think of it, love a feedback loop 🙏🏼

❤️ Love it, keep them coming!
👍 Great, always a good read.
😊 Enjoyable, but let’s see some variety.
👎 Okay, seen better.
🛑 Get a different job.

Subscribe below and follow me on LinkedIn or Twitter to never miss an update.

WHO THE HELL IS GOING TO BUILD ALL THE BITS?

DeReK WaTSoN — Wed, 02 Sep 2026 18:23:18 GMT

Everyone is watching Optimus, Figure and Unitree.

I went looking at what actually goes inside the bloody things.

Motors. Bearings. Roller screws. Precision gears. Encoders. Force sensors. Magnets. Actuators.

Then I looked at who makes them, how difficult they are to manufacture and what happens if we really try to build millions of robots a year.

That turne…

WHO THE HELL IS GOING TO BUILD ALL THE BITS?

DeReK WaTSoN — Wed, 02 Sep 2026 07:17:15 GMT

I have recently been looking more closely at robotics, and something has started bothering me

Everyone is talking about the robots themselves. Optimus, Figure, Unitree, Boston Dynamics, Apptronik, Agility, UBTech and an increasingly long list of companies now throwing serious money at Physical AI.

The forecasts are all over the place, but you don’t need to believe the most ridiculous ones for the numbers to get very big.

And that got me thinking about something much more basic.

But who the hell is going to make all the bits?

Because this isn’t software. You can’t spin up another million robots in AWS. They have to be built, and there is a surprising amount of difficult engineering packed inside each one.

A humanoid needs motors, bearings, gears, screws, encoders, sensors, magnets, controllers, batteries and actuators. It needs them everywhere. Shoulders, elbows, wrists, hips, knees and ankles all have to move accurately, carry weight and survive being used thousands of times a day.

The hands are another problem altogether. Ten fingers, each trying to reproduce movements and touch that we barely think about when we pick up a glass or turn a door handle.

Once you add cameras, power electronics, cooling, cables and compute, a humanoid starts to look like a small factory’s worth of precision engineering squeezed into one machine.

And the numbers get big very quickly. If the industry eventually builds ten million humanoids a year, it won’t need ten million motors, bearings or sensors. It could need hundreds of millions of them.

We spend a lot of time talking about whether the robots will be good enough.

I’m beginning to wonder whether we’ll be able to make enough of the bloody things.

The return of boring engineering

For the past few years, AI infrastructure has mostly meant GPUs, data centres, networking and memory. Then everyone realised the bloody things needed huge amounts of electricity. So we started worrying about grids, transformers, gas turbines, nuclear power and copper.

Robotics adds a different problem. You actually have to make things.

AI can design a better robot, write the software and teach it new tasks. But if Tesla, Figure and twenty Chinese manufacturers suddenly need another million precision bearings, motors or gearboxes, somebody has to physically make them.

And there are plenty of places where supply could get tight:

You don’t need shortages everywhere for this to become a problem. One component that can’t keep up can stop the whole production line.

The car industry found that out during the semiconductor shortage. Factories capable of producing thousands of cars sat idle because they couldn’t get enough relatively cheap chips.

Robotics will have the same weakness. A $20,000 robot is no use if you can’t get the $50 part needed to finish it.

Which led me to Japan 🇯🇵 and Germany 🇩🇪

This was the bit I hadn’t really thought about until I started looking at who actually makes this stuff.

Japan and Germany aren’t leading the AI hype cycle, but they’ve spent decades getting very good at the engineering robots need. Bearings, motors, gears, machine tools, linear guides, encoders, servo drives and precision screws.

None of it sounds particularly exciting. Most of these companies will never have a robot dancing on stage at a tech conference.

But without their parts, the robot doesn’t dance at all.

Japan has Harmonic Drive Systems and Nabtesco making precision reducers. THK makes linear-motion systems and screws. NSK makes bearings and screws. Nidec makes motors. Yaskawa and Fanuc have been building servo systems and industrial robots since long before anyone thought putting “Physical AI” in a PowerPoint would increase the valuation.

Germany has Schaeffler, Bosch Rexroth, Siemens and Infineon, alongside hundreds of specialist Mittelstand engineering companies most investors have probably never heard of.

But once you start looking properly, this is clearly not just a German and Japanese story.

Switzerland 🇨🇭has maxon and an extraordinary precision-engineering ecosystem.

Britain🇬🇧 has Renishaw, which makes precision measurement and encoder systems.

Taiwan 🇹🇼 has Hiwin, one of the world’s important manufacturers of linear guides and ball screws.

South Korea 🇰🇷 has enormous capability across batteries, electronics and motors.

America 🇺🇸 has companies such as Moog, Allegro MicroSystems and Monolithic Power Systems sitting in different parts of the motion and control stack.

Australia 🇦🇺has Lynas Rare Earths, which doesn’t make a single robot but suddenly becomes relevant because high-performance motors need permanent magnets.

And then, inevitably, there is China.

China 🇨🇳 could change the economics of the whole thing

China isn’t waiting to see how big humanoid robotics becomes. It’s building the robots and pretty much everything that goes inside them: motors, magnets, gearboxes, bearings, actuators, sensors, batteries and electronics.

It already dominates the rare-earth magnet supply chain and its precision component manufacturers are getting better quickly.

Take Harmonic Drive. It may make a better precision reducer today, but what happens if a Chinese competitor can make one that’s 95% as good for half the price, and make a million of them?

We’ve seen what China can do to the economics of an industry with solar panels, batteries and now electric vehicles. I wouldn’t bet on robotics being any different.

At low volumes, making the best component matters. At millions of robots a year, making it cheaply, reliably and in enormous numbers may matter more.

The part I keep coming back to is actuation.

Actuators are the robot’s muscles. A single joint can contain a motor, magnets, bearings, a gearbox or roller screw, an encoder, force sensors and the electronics controlling it. Now spread those across shoulders, elbows, wrists, hips, knees, ankles and fingers.

One component I hadn’t given much thought to before is the planetary roller screw. It turns the spinning movement of a motor into powerful, precise linear movement, which is useful when you need a lot of force from a small joint.

They’re also bloody difficult to make.

The threads have to be extremely accurate and survive repeated heavy loads. Materials, grinding and heat treatment all matter. You can’t suddenly double production because the orders have arrived.

Chinese manufacturers are already moving from small pilot lines towards much larger production capacity.

That’s the sort of thing I’m looking for: an obscure component that suddenly has to be made in completely different volumes.

Precision gears have much the same problem. A robot joint needs to move accurately and without much play. Japan’s Harmonic Drive Systems and Nabtesco have spent decades getting very good at this, while China’s Leader Harmonic is coming after the same market.

The complication is that nobody yet knows which design will win. Some robots use harmonic gears, others use planetary systems or linear actuators. We may end up with complete joints that combine the motor, gearing, sensors and electronics into one replaceable unit.

So some of the parts we think will be important today may disappear, while others become enormous markets.

Then there are the hands

Hands are another problem altogether.

Pick up a wine glass and think about what your fingers are doing. You’re judging how hard to squeeze it, whether it’s slipping, how heavy it is and whether you’ve actually got hold of it. Most of that happens without you thinking about it.

A robot has to sense all of this.

Vision gets the hand to the glass. Touch stops it smashing the bloody thing.

That is creating a new market for tactile sensors, force sensors, electronic skin and increasingly complicated robotic hands.

And there are already some fascinating private companies working on exactly this:

XELA Robotics in Japan is developing tactile sensors that measure pressure and shear forces.
tacterion in Germany is developing flexible electronic skin.
GelSight in America uses optical sensing to allow machines effectively to see what they’re touching.
Contactile in Australia is working on force, grip and slip sensing.
Bota Systems in Switzerland makes compact six-axis force and torque sensors.
Tesollo in South Korea is developing highly dexterous robotic hands and its own integrated actuator technology.
Synapticon in Germany is integrating motors, servo drives and electronics into compact motion and actuator systems.
ThinGap in America makes lightweight, high-performance motors.
Daimon Robotics in China is developing high-resolution tactile sensing.

These companies aren’t trying to build the next Optimus. They’re making the motors, sensors, actuators and hands that every robot manufacturer could end up buying.

That could be a very good place to be if production takes off.

The boring companies are having a very good year

The listed companies are interesting too. A surprising number of the businesses making the less glamorous parts of the robotics supply chain have had very strong share-price moves this year.

These aren’t pure robotics companies, so I wouldn’t pretend humanoids explain all of those moves. But it’s hard not to notice how many companies making motion, sensing, gearing and precision components are having very good years.

Batteries probably aren’t the first problem

I assumed batteries would be near the top of the list. Millions of robots means millions of batteries.

But battery production is already enormous because of EVs and energy storage. Even a few million humanoids wouldn’t move global battery demand anything like another few million electric cars would.

The bigger issue is weight. A larger battery gives the robot more runtime, but it also makes it heavier. That means more powerful motors, which use more power. You quickly end up chasing your own tail.

So energy density matters. Solid-state batteries could make a real difference here, but I still don’t think batteries are the first place supply gets tight.

Magnets could be

Many high-performance motors use neodymium permanent magnets, and China dominates their supply chain.

That becomes a problem if America, Europe, Japan and South Korea start building millions of robots at home while still relying on China for a basic ingredient inside the motors.

It also explains why Australia’s Lynas Rare Earths belongs in this story despite never making a robot.

Ten thousand robots is one thing. Ten million is another.

You can build 10,000 expensive robots with specialist suppliers and plenty of manual production. At ten million, you need something completely different.

Parts have to become cheaper and more standardised. Suppliers need much bigger factories. Production yields matter. And no manufacturer spending billions on a robot factory will want its entire production line dependent on one small supplier in Bavaria.

That’s when the obscure stuff starts to matter.

A shortage of planetary roller screws could do it. So could precision reducers, miniature motors, encoders, force sensors or permanent magnets.

It might be something even more boring that I’ve never heard of.

That’s what I’m looking for.

Stop looking at the robot

We’ve seen this with AI already. ChatGPT got everyone’s attention, then Nvidia made the fortune supplying the compute. That created demand for data centres, memory and networking, which needed transformers, copper, gas turbines and a frankly ridiculous amount of electricity.

Robotics is going to put the same sort of pressure on manufacturing.

We can build extraordinary robots today. Building ten million of them every year is a completely different problem. That means hundreds of millions of motors, bearings, gears, screws, encoders and sensors, all made accurately enough, cheaply enough and quickly enough.

Something won’t keep up.

It might be roller screws, precision reducers, magnets, motors or tactile sensors. It could just as easily be some obscure component made by a 70-year-old engineering company that nobody outside its industry has ever heard of.

Everyone is trying to work out who will win the race to build the robot.

I’m more interested in who already makes the thing they all discover they can’t get enough of.

Go take a and look, For the ❤️ of startups

Scout - who just got funded
Newly funded startups, pre-seed to Series B, as the rounds happen.

Raise - who’s deploying
New funds with fresh capital and cheques to write.
Wire - what changed, what matters
The intelligence feed that filters the noise, with the “so what” attached.

If you have not joined the Fusion42 Community on Telegram —
it is probably time to do so.

For the ❤️ of Startups

Please let me know what you think of it, love a feedback loop 🙏🏼

❤️ Love it, keep them coming!
👍 Great, always a good read.
😊 Enjoyable, but let’s see some variety.
👎 Okay, seen better.
🛑 Get a different job.

Subscribe below and follow me on LinkedIn or Twitter to never miss an update.

Global markets: yields and oil remain high

A New Financial Cycle Has Begun

DeReK WaTSoN — Sun, 23 Aug 2026 21:27:42 GMT

Why the race for AI, power and strategic materials is colliding with the global debt problem

The AI boom is not happening in the cloud. It is being built with debt, electricity, copper, chips and rare-earth magnets. That physical arms race is arriving just as governments need to refinance record debts—and China controls many of the industrial bottlenecks.

THE SIMPLE VERSION

Governments across the developed world need to borrow more just as inflation is stopping central banks from making money cheap. AI is adding another vast demand for capital, electricity, copper and strategic minerals. Japan is being squeezed by its currency, Europe by issuance, and America by debt and refinancing. China has weak domestic demand but controls critical processing and magnet supply chains. Oil is the accelerant. Gold and Bitcoin are the market’s vote against the policy mix.

THE OVERLOOKED MACRO SHOCK

AI can eventually make goods and services cheaper. Building it is inflationary first. Technology companies are competing with governments for capital, with households for electricity and with industry for copper, transformers and strategic minerals. China’s control of refining and magnets turns that commodity squeeze into geopolitical leverage.

This week’s market message

The clearest signal was not any single price. It was the combination: long-term government yields remained near multi-decade highs, the dollar fell, oil rose and monetary alternatives rallied. Normally, higher US yields support the dollar. When yields rise and the dollar falls together, investors are asking for more return because they see greater fiscal and inflation risk—not because American growth is suddenly stronger.

The AI–China–commodity chain

This is the layer that connects the technology boom to bonds, inflation and geopolitics. AI infrastructure creates a new buyer for capital and physical resources while China controls much of the processing needed to turn those resources into usable components.

THE STRATEGIC CONTRADICTION

America leads much of the AI stack, but China controls many of the physical inputs required to scale it. The competition is no longer just about models and chips. It is about capital, grids, copper, processing, magnets and power.

1. The bond market is repricing the developed world

For most of the past fifteen years, governments could borrow heavily because central banks held rates near zero and bought bonds. That regime has ended. Inflation is higher, central banks are no longer absorbing the same amount of debt, and governments are issuing more to fund deficits, defence, energy security, ageing populations and infrastructure. At the same time, technology companies are borrowing vast sums to finance AI infrastructure.

The result is simple supply and demand. More bonds are being sold, while the largest dependable buyer—the central bank—is stepping back. Private investors will still buy, but only at a price. That price is a higher yield.

WHY IT MATTERS

A higher government yield becomes the base rate for the rest of the economy. Mortgages, corporate debt, project finance and equity valuations all become more expensive.

2. America is the centre of the refinancing problem

US gross federal debt has passed $40 trillion. The more important point is not the headline number by itself; it is the speed at which old debt must be refinanced and new deficits funded. Treasury has deliberately issued heavily at the short end, using bills rather than locking in today’s long-term rates.

That swaps duration risk for refinancing risk. Short bills barely move in price when long yields jump, so the balance sheet looks less sensitive today. But those bills mature quickly. If rates stay high, the interest bill resets within weeks or months rather than years.

Using the estimate that roughly $6.45 trillion must be refinanced before year-end, every additional basis point paid across that amount adds about $645 million to annual interest expense. That is an approximation—the debt will mature at different points and different yields—but it shows the scale. A move of 50 basis points is roughly $32 billion a year.

Treasury showed its pain threshold

When the 30-year yield reached 5.34%, Treasury announced that selected long-end buybacks would double from $2 billion to at least $4 billion per operation between September and November. The 30-year yield initially fell to about 5.19%, then returned to roughly 5.27% by Friday.

The buyback is not QE. Treasury is exchanging one form of government debt for another, not creating money through the Fed. It can improve liquidity and reduce long-bond supply, but it does not reduce the deficit. If the purchase is financed with more bills, it shortens the debt profile and increases future refinancing exposure.

The operation is also small beside a $32.2 trillion Treasury market. Its importance was the signal: Washington revealed that a 30-year yield around 5.30% is politically and financially uncomfortable. Markets test visible pain thresholds.

3. Japan is the pressure point

Japan spent decades with near-zero rates. Its banks, insurers and investors therefore placed capital abroad, while global funds borrowed cheap yen to buy higher-yielding assets elsewhere. That is the yen carry trade.

Now Japanese inflation is rising, oil is expensive and the yen remains weak. The Bank of Japan is under pressure to raise rates. Its 10-year yield ended Friday around 2.88%, the 20-year near 3.76% and the 30-year around 4.06%. Those are enormous changes for a system built around almost-free money.

Higher Japanese yields make domestic bonds more attractive. They also make borrowing yen more expensive. Japanese capital has more reason to return home, while leveraged investors have more reason to close carry trades. Closing the trade means selling the foreign asset and buying back yen. That can hit US Treasuries, technology shares and other risk assets at the same time.

Why the Fed’s FIMA facility matters

If Japan needs dollars to defend the yen, it could sell part of its large Treasury portfolio. FIMA offers a bridge: Japan can temporarily pledge Treasuries to the Fed and receive dollars rather than dumping the bonds into the market. It can then sell those dollars and buy yen.

That protects the Treasury market from forced selling in the short term. It is a secured, reversible loan—not automatic QE. The concern starts if seven-day funding is repeatedly rolled, expanded or treated as permanent. At that point, temporary liquidity support becomes ongoing official balance-sheet support.

JAPAN’S DILEMMA

Raise rates and risk unwinding the carry trade. Hold rates down and risk further yen weakness, imported inflation and more intervention.

4. Europe has the same disease with different symptoms

Europe is not being singled out by the market; it is suffering from the same global repricing. Germany’s 30-year yield reached about 3.79%, its highest since 2011. French 10-year yields moved above 4.1%, while long French yields approached 5%. Italy and the UK face the same refinancing pressure.

Germany is loosening its old fiscal rules to fund defence and infrastructure. Euro-area governments are borrowing more for defence, healthcare and ageing populations. Record German issuance is expected, while the European Central Bank is allowing bonds on its balance sheet to mature rather than replacing them. More supply is therefore being pushed onto private investors.

The ECB faces an awkward split. It must control inflation across the whole bloc, but higher yields hurt highly indebted countries much more than Germany. If it supports the weakest sovereign markets too aggressively, it risks appearing to finance governments. If it does too little, spreads widen and the euro area fragments.

5. China is moving in the opposite direction

China is the major exception. Its five-year government yield is around 1.39% and its 10-year yield near 1.69%, a 13-month low. While developed-market yields are rising because governments need more money and investors fear inflation, Chinese yields are falling because households save heavily, private credit demand is weak and the economy is struggling to generate enough productive domestic borrowing.

That is useful for Beijing: it can finance stimulus and infrastructure more cheaply than the West. It also gives China room to use fiscal policy while other governments are being disciplined by bond markets. But crashing yields are not proof of effortless strength. They also signal weak consumption, poor loan demand, pressure on banks and lingering property problems.

China could buy higher-yielding US, European or Japanese debt, but the decision is not purely about return. It must also consider currency risk, sanctions and the danger of holding reserves inside a system Washington can weaponise. That is why China is likely to diversify selectively: some foreign bonds, more gold, more renminbi trade settlement and more financial infrastructure outside the dollar system.

6. How the AI–China commodity shock works

AI is usually discussed as software, productivity and company valuations. The immediate build-out is far more physical. Data centres require land, grid connections, transformers, cooling systems, gas turbines, advanced chips and large amounts of copper. High-performance motors, power systems, semiconductors, aerospace and defence also depend on rare earths and other strategic minerals.

The largest US technology companies are expected to spend about $725 billion on AI-related capital expenditure in 2026. Alibaba has just proposed raising roughly $10.2 billion for chips, infrastructure and models. Nebius raised $5 billion through convertible bonds for data centres and AI capacity. This money does not come from nowhere. AI companies are competing with governments, utilities and industry for the same capital. That adds corporate issuance to an already crowded bond market.

Power is the first bottleneck

The International Energy Agency expects global data-centre electricity use to more than double to around 945 terawatt-hours by 2030—slightly more than Japan consumes today. Data centres are expected to account for nearly half of US electricity-demand growth this decade. That means more grids, generation, storage, gas and nuclear investment. It also means higher power prices where infrastructure cannot keep up.

AI can be highly deflationary once it raises productivity. Building the infrastructure is inflationary first. It absorbs capital, skilled labour, energy equipment and commodities years before the full productivity benefit arrives.

Copper is the clearest broad commodity signal

Copper traded around a record $6.73 per pound this week and is up more than 18% this year. AI data centres, grid reinforcement, renewable generation and electrification are all increasing demand while new mines take years to permit and build. The IEA still projects a copper supply gap of roughly 25% by 2035 based on announced projects. China also controls about half of global copper smelting capacity, so the vulnerability is not only in mining; it sits in processing.

Rare earths are a geopolitical price, not one market price

Rare earths do not trade on one transparent global exchange like copper. There are different oxides, metals, alloys and magnets, often sold through bilateral contracts. China accounted for about 60% of magnet rare-earth mining in 2024, 91% of refining and 94% of sintered permanent-magnet production. The real leverage sits in separation, refining and magnet manufacturing—not simply in owning ore.

That has created two markets. China’s domestic rare-earth price index fell to 260.8 in August from roughly 310 in March, and its NdPr oxide benchmark dropped about 12% from July. Yet controlled heavy rare earths remained dramatically dearer outside China. Indicative August transactions put terbium oxide near $990–$999 per kilogram inside China against roughly $3,625–$4,500 outside; dysprosium showed the same divide. These are opaque bilateral prices, but the message is clear: export licences and political access can matter more than the domestic commodity quote.

China increased some yttrium and permanent-magnet shipments to the US ahead of trade talks while continuing to restrict dysprosium and terbium exports to Japan. That links the AI and mineral story directly to the yen and carry trade. China can apply industrial pressure to Japan without selling a single Treasury.

THE STRATEGIC EDGE

America leads much of the AI stack. China controls many of the physical inputs needed to scale it. The next phase of the contest is therefore about grids, copper, processing, magnets and power—not just models and chips.

Western governments will respond with subsidies, price floors, strategic stockpiles and domestic processing. Those policies improve resilience, but facilities outside China can cost 20% to more than 150% more to build and roughly 50% more to operate. Supply-chain security therefore adds another layer of spending, deficits and inflation to the global debt problem.

The coercion trap

The proposed US MATCH Act shows how technology controls feed back into this system. If enacted, it would give allied supplier countries such as the Netherlands and Japan 150 days to align their semiconductor-equipment controls with Washington. If they did not, the United States could extend its jurisdiction over foreign machines containing American technology and restrict both sales and servicing in China. The bill is not yet law, but it advanced from the House Foreign Affairs Committee by 36 votes to eight.

The short-term security logic is clear. American export controls achieve little if Dutch or Japanese suppliers can provide the same essential equipment. But forcing an ally to restrict its own national champion carries a strategic cost. China represented 29.1% of ASML’s total net sales in 2025, and the Dutch government has formally objected to the proposed law’s extraterritorial reach.

Each restriction gives China another reason to accelerate domestic lithography, retaliate through rare earths and redesign supply chains around non-American technology. It also gives Europe and Japan an incentive to remove US components from their own products so that Washington can no longer control where they are sold. The policy can slow China today while weakening the network of dependence and trust on which American power rests tomorrow.

AMERICA IS SPENDING ITS POWER

America is using its dominance to slow China today. But every time it coerces an ally, it gives the world another reason to build a system that America cannot control tomorrow.

7. Oil is the accelerant

Brent ended the week close to $95 as the Iran conflict and uncertainty around the Strait of Hormuz kept supply risk elevated. About a fifth of the world’s oil and liquefied natural gas normally passes through the strait. Higher energy prices feed directly into transport, food, manufacturing and household bills.

This matters because governments need lower yields at the exact moment oil is pushing inflation higher. Central banks cannot cut aggressively into a fresh energy shock without risking their credibility and currencies. The US threat of tougher sanctions on Iran may increase economic pressure on Tehran, but it also raises the oil and inflation risk facing the US Treasury and the Fed. One arm of policy is working against the other.

8. Every major central bank is trapped differently

This is why the world feels unstable. There is no single interest-rate policy that solves the problem. Each central bank can remove pressure from one part of the system only by moving it somewhere else.

9. This is a dollar unwind—not yet ‘Sell America’

A true Sell America event would mean the dollar, Treasuries and US equities falling together in disorderly fashion, with poor auctions and damaged market liquidity. We are not there. US shares rose on Friday, buyers still appeared at auctions and Treasuries continue to trade normally.

But the trade is spreading. The dollar fell nearly 1% over the week while long yields remained high. Gold rose more than 5%, silver finished near $69.62 and Bitcoin had its strongest week in roughly two and a half years. Investors are not abandoning America. They are reducing the amount of unhedged confidence they place in the dollar and nominal government debt.

What reserve managers are actually doing

Central-bank balance sheets show a gradual diversification rather than a sudden flight from the dollar. The dollar’s share of disclosed global foreign-exchange reserves has fallen from about 71% in 1999 to 57.13% in the first quarter of 2026. But it rose from 56.42% in the previous quarter and has been broadly stable near current levels since 2022. The direction is clear; the pace is slow.

The ownership mix matters as much as the total. Foreign private investors still buy American assets, including record amounts of US equities, while foreign official institutions have become less reliable Treasury buyers. The United States is therefore becoming more dependent on domestic investors, private foreign capital and leveraged market structures rather than reserve managers whose demand was traditionally less price-sensitive.

DE-DOLLARISATION HAPPENS AT THE MARGIN

The dollar does not need to be replaced for US borrowing costs to rise. Reserve managers only need to direct less of each new dollar into Treasuries and more into gold, other currencies or domestic assets. With Treasury issuance expanding, a weaker marginal buyer matters long before the dollar loses reserve status.

Why gold is the cleanest signal

Gold has no issuer, no refinancing calendar and no promise from a government. It benefits when real confidence in currencies and sovereign debt falls. Bitcoin expresses a similar idea with much greater volatility and technology risk. Silver also carries industrial demand, so it is a less pure monetary signal.

Equities can still rise in nominal terms. Companies with pricing power and real cash flow can pass on inflation. The danger is that persistently high real yields eventually reduce valuations and make financing harder—especially for businesses dependent on cheap capital.

10. The feedback loop

That is the debt spiral people are worried about. It does not mean default tomorrow. Countries that borrow in their own currency have powerful tools. They can tax, cut spending, extend maturities, change regulation, provide liquidity and allow inflation to reduce the real value of old fixed-rate debt.

The problem is that these tools have costs. Tax rises weaken growth. Spending cuts are politically difficult. Short funding increases refinancing risk. Currency weakness raises inflation. Central-bank purchases damage credibility if they become permanent. The spiral begins when the market sees every rescue adding to the next problem.

What this means for ordinary people

BORROWERS Mortgage and loan rates can stay high even if central banks begin cutting overnight rates.

SAVERS Cash earns more interest, but energy and food inflation can still reduce purchasing power.

PENSIONS Higher yields help new bond buyers but damage the price of long-dated bonds already held.

BUSINESSES Debt-heavy firms and companies without pricing power face the greatest squeeze.

INVESTORS Diversification matters more when bonds and equities can fall together. Hard assets help, but they are not guaranteed one-way trades.

What would make this worse

The dangerous combination is oil above $100, the US 30-year yield holding above 5.30%, dollar–yen breaking decisively through 160, weak Treasury auctions, wider French and Italian spreads, and official liquidity facilities being enlarged or repeatedly rolled. If the dollar falls while all of that happens, the market is adding a fiscal credibility premium.

What would calm it down

A durable fall in oil and inflation would give central banks room to cut. Credible fiscal restraint would reduce future bond supply. Strong auctions without official support would show private demand is intact. A stable yen after a measured Bank of Japan move would reduce carry-trade risk. Better Chinese domestic demand would make its low yields look like successful easing rather than a warning about weak growth.

The next tests

Next week brings 2-, 5- and 7-year US Treasury auctions, fresh US growth and inflation data, Nvidia’s results and the Federal Reserve’s Jackson Hole gathering. Japan’s September policy meeting is the next major pressure point for the yen and the carry trade. Europe’s auctions and sovereign spreads will show whether the stress remains global or begins concentrating in weaker countries.

Watch the relationships, not just the levels. Falling yields with a firmer dollar would be relief. Falling yields with a collapsing dollar would look like financial repression. Rising yields with a stronger dollar would be conventional tightening. Rising yields with a weaker dollar is the warning signal.

The bottom line

The world is leaving the cheap-money era with far more debt than it entered with. America is fighting refinancing risk. Japan is fighting the consequences of ending zero rates. Europe is fighting issuance and fragmentation. China is fighting weak domestic demand while controlling strategic processing and magnets. AI is creating a new call on capital, power and copper. Reserve managers are adding gold and reducing their marginal dependence on Treasuries. Oil is making every decision harder. The market has found the contradiction: governments need lower yields while inflation, debt supply and physical scarcity demand higher ones.

This is not yet a global debt crisis. It is the stage before one, when policy still works—but each intervention lasts for less time, costs more and moves the pressure somewhere else. It’s not looking pretty that’s for sure.

Sources and further reading

US Treasury doubles selected long-end buybacks

Why the bond-market relief may be brief

European debt issuance and rising yields

Japan’s bond yields and inflation pressure

China’s falling yields and weak credit demand

China holds benchmark lending rates

IEA: energy and AI

IEA: rare-earth concentration and economic exposure

IEA: 2026 critical-minerals outlook

Alibaba’s proposed $10.2bn AI capital raise

Nebius raises $5bn for AI data centres

China’s strategic rare-earth shipments to the US and Japan

August rare-earth pricing and regional premiums

Reuters: MATCH Act and allied semiconductor controls

ASML 2025 annual report and China exposure

IMF: currency composition of reserves, Q1 2026

Federal Reserve: the international role of the dollar

World Gold Council: central-bank buying in Q2 2026

Reuters: Japan and China reduce Treasury holdings

Federal Reserve explanation of FIMA repo

Gold and precious-metals close

US Treasury Debt to the Penny

This report is commentary for general information, not personal investment advice. Market levels are approximate closing figures for Friday, 21 August 2026, and can change rapidly.

Go take a and look

Scout - who just got funded
Newly funded startups, pre-seed to Series B, as the rounds happen.

Raise - who’s deploying
New funds with fresh capital and cheques to write.
Wire - what changed, what matters
The intelligence feed that filters the noise, with the “so what” attached.

If you have not joined the Fusion42 Community on Telegram —
it is probably time to do so.

For the ❤️ of Startups

For the ❤️ of startups

✌🏼 & 💙
Derek

Thank you for reading. If you liked it, share it with your friends, colleagues and everyone interested in the startup Investor ecosystem.

If you've got suggestions, an article, research, your tech stack, or a job listing you want featured, just let me know! I'm keen to include it in the upcoming edition.

Please let me know what you think of it, love a feedback loop 🙏🏼

❤️ Love it, keep them coming!
👍 Great, always a good read.
😊 Enjoyable, but let’s see some variety.
👎 Okay, seen better.
🛑 Get a different job.

Subscribe below and follow me on LinkedIn or Twitter to never miss an update.

Intelligence Is Abundant. Judgement Is Not.

DeReK WaTSoN — Tue, 14 Jul 2026 14:31:18 GMT

For most of history, intelligence was the bottleneck. That constraint has gone. Frontier models will draft your strategy, your financial model and your investor update in seconds, and rather well.

Yet founders are not obviously better off, because intelligence was never the whole job. Intelligence produces options. Judgement decides what matters when every option looks plausible. The models flooded the first market and left the second untouched. Intelligence became abundant. Judgement stayed scarce. Most AI tooling is still solving the old problem.

You have already met the problem

You have done this. Pasted your deck into a chatbot, had a genuinely useful conversation, then come back a week later and started from zero. It did not know your company then and it does not know it now.

That is not a memory bug waiting for a patch. A chat transcript is a pile of language. Nothing in it is dated, nothing is versioned, nothing separates a passing remark from an established fact, and advice built on a pile of language is generic by construction. That is why the chatbot’s answer to your company sounds suspiciously like its answer to every company.

And it agrees with you. These systems are tuned to be pleasant because pleasant retains users. You are already the most optimistic person in every room. The last thing you need is a brilliant assistant whose deepest instinct is to tell you that you are right.

Trust is an architecture problem

The obvious question about any AI that claims to know your company: why won’t it make things up?

Fair question, because we watched it happen. An early version of Arthur kept a running summary of each company, written by a model. One day it stated, with complete confidence, that a company was incorporated in a jurisdiction it had never touched. No data said so. The model completed a plausible pattern, because that is what models do. We wrote rules against it. The rules failed.

So we stopped asking the model to remember anything important. Arthur keeps the facts of your company separately from the AI, in a record where every entry is dated and traceable to something real: a document you uploaded, a thing you said, a result that happened. The consequence is simple. Arthur cannot invent your company’s history because it never owns that history in the first place. And when new information contradicts the record, the conflict is put in front of you to settle, not quietly papered over.

The parts that judge you are deterministic. Same inputs, same answer, and you can inspect the reasoning. A chatbot cannot tell you why it said something. Arthur can show you.

Strong opinions, loosely held

Strong opinions, loosely held, create healthy discourse and are the fastest way to truth. A co-founder who agrees with everything is dead weight.

So when Arthur says your raise is not ready, that is a strong opinion: a verdict with the workings attached. Attack it. Win the argument and the record updates, along with every judgement that depends on it. Lose the argument and you’ve learned something before an investor teaches it to you at far greater cost.

And because Arthur holds a real picture of your company, it does not wait to be asked. It finds the gap between what you are and what a fundable company at your stage looks like, and brings you the next decision with the reasoning attached. Ask a chatbot nothing and you get nothing.

Where the standard comes from

Ready by whose definition? This is where judgement gets concrete. Judgement is knowing which missing customer reference matters and which does not. When $300k is enough to raise and when $3m is too much. That fixing your pricing is worth more than your next feature. Which investor objection is fatal and which is simply the opening move of a negotiation.

Frontier models reason surprisingly well, and that is not the gap. What they lack is a standard grounded in outcomes. Most written startup advice comes from people who never operated or wrote a cheque, and the calls above were never written down at all. They live in results: the rounds that closed, the deals that died, the companies that ran out of road six months after everyone agreed they were fine.

That judgement was earned at Fusion42, over years on both sides of the table: mentoring founders, assessing businesses, investing capital and seeing which decisions survived contact with reality. And because markets move, that judgement keeps evolving. Your company is measured against this quarter’s market, not a model’s memory of an old one.

Judgement as infrastructure

Arthur starts with Fusion42’s judgement. But the bigger opportunity is not preserving one firm’s experience. It is creating a system where judgement itself becomes an asset that can be captured, tested and improved.

The future is not one AI with one opinion. Every investor, accelerator and operator holds their own views about what makes a company investable. Over time, those views become theses: explicit beliefs about what matters, what does not, and why. Today those theses mostly live in heads: applied inconsistently, invisible to scrutiny, and gone when the person leaves.

The next step is opening the judgement layer: letting investors encode their own theses into the same transparent framework everything else runs on, adjusting the weightings, adding the criteria that matter to them, making their assumptions visible. Not by training the AI, which would reopen every trust problem above, but by writing judgement down where it can be read. And it reads both ways: the same system that lets investors encode how they judge lets founders see exactly how they are being judged.

And written down means testable. The record says: here is your thesis, here is where it worked, and here is where reality disagreed. Theses become decisions. Decisions become outcomes. Outcomes refine judgement. That loop is why the system cannot quietly become a mirror for anyone’s bias. Bad judgement gets exposed. Good judgement compounds.

Intelligence scales by replication. Judgement scales by encoding. The scarce asset was never one firm’s opinion. It is the ability to capture, structure and compound human judgement. Arthur is the mechanism.

Your company is the asset

Everything Arthur records about your company accumulates into an asset you can inspect, move and own, not dissolved into a model’s weights. Six months in, that record holds what no chatbot session can: what your company believed, when it changed its mind, and why. New tools forget. Arthur compounds.

And it is yours. In institutional deployments it is completely ringfenced: nothing flows back, nothing trains a global model, nothing averages you into everyone else. The prevailing AI business model runs the other way, one central brain quietly enriched by absorbing what makes each customer distinct. If judgement is the scarce asset, that concentrates it in exactly the wrong place.

The point

The question is no longer whether AI can think. It is whether it knows enough about your company to be critical and objective for right reasons.

Go and look

Scout - who just got funded
Newly funded startups, pre-seed to Series B, as the rounds happen.

Raise - who’s deploying
New funds with fresh capital and cheques to write.
Wire - what changed, what matters
The intelligence feed that filters the noise, with the “so what” attached.

If you have not joined the Fusion42 Community on Telegram —
it is probably time to do so.

For the ❤️ of Startups

Please let me know what you think of it, love a feedback loop 🙏🏼

❤️ Love it, keep them coming!
👍 Great, always a good read.
😊 Enjoyable, but let’s see some variety.
👎 Okay, seen better.
🛑 Get a different job.

Subscribe below and follow me on LinkedIn or Twitter to never miss an update.

We built this for Arthur. It felt selfish to keep it.

DeReK WaTSoN — Sun, 12 Jul 2026 14:30:40 GMT

For the past year we’ve been building Arthur - a startup operating system that works alongside founders as your co-founder, shaping your company on live intelligence rather than gut feel and stale advice. Part of that intelligence is pure market signal: who just got funded, who’s deploying fresh capital, what changed overnight - localised to your geography, personalised to your news type and sector.

We built that signal layer for Arthur. Then we realised it was too useful to keep private.

So we’re releasing it. Three products, free to every founder.

Scout - know who just got funded

Newly funded startups, pre-seed to Series B, as the rounds happen.

Every round in your space is intelligence. A competitor raising £8m tells you what investors are paying for. A fund backing three companies adjacent to yours tells you exactly who to email next.

Raise - know who’s deploying

New fund announcements: fresh capital, mandate, cheque size, focus.

A fund that closed eighteen months ago is a very different conversation to one that closed last month. New funds need to write cheques. Raise tells you who’s sitting on fresh powder - so you stop cold-emailing funds that finished deploying in 2024.

And watch for the next update: we're adding the investors behind every new raise in Scout.

Wire - know what changed, know what matters

The global intelligence feed, filtered hard, with the “so what” attached.

Every story carries The Wire Takeaway - not “Samsung raises foundry prices” but “your wafer costs just went up 10-15%, call your suppliers this week. Filter by News type, sector, region to get fully personalised and localised news. Weekly reports roll it up so you can skim Monday morning and get back to building.

Reading everything is a full-time job. You already have one.

Where this goes

These feeds are one part of what Arthur runs on. Combined with what Arthur knows about your company - your stage, your traction, your gaps - raw signal becomes something sharper: which investors to approach this month, which shift to act on, which assumption someone else’s round just invalidated.

You’ll hear much more about Arthur soon. For now, this layer is yours.

Go and look

No paywall to browse.

If you have not joined the Fusion42 Community on Telegram —
it is probably time to do so.

For the ❤️ of Startups

Please let me know what you think of it, love a feedback loop 🙏🏼

❤️ Love it, keep them coming!
👍 Great, always a good read.
😊 Enjoyable, but let’s see some variety.
👎 Okay, seen better.
🛑 Get a different job.

Subscribe below and follow me on LinkedIn or Twitter to never miss an update.

The Raise Report

DeReK WaTSoN — Tue, 30 Jun 2026 08:29:56 GMT

This report provides an overview of new venture capital funds, CVCs and accelerators that have announced new capital to invest in early-stage startups during June 2026.

Monthly Report

June 2026 saw a robust activity in early-stage funding, with a total of 91 new funds identified raised $14.1B.
Artificial Intelligence (AI) infrastructure continues to dominate investment interest, followed by generative AI and climate technology. The United States remains the leading hub for these investments, though the United Kingdom and several Asian and European countries also show significant activity. Several large funds, including Sky Factory and Menlo Ventures XVII, have made substantial commitments, indicating continued confidence in the early-stage ecosystem.

Full details of all funds and people live at RAISE.FUSION-42

SECTORS

AI-related sectors, particularly AI Infrastructure and Generative AI, are at the forefront of early-stage investments, reflecting the ongoing technological advancements and market demand in artificial intelligence. Climate Tech and Clean Energy also feature prominently, highlighting a growing focus on sustainable solutions. Traditional sectors like Agtech, Enterprise Software, and Digital Health continue to attract significant capital, indicating sustained innovation across diverse industries.

GEO

The United States overwhelmingly leads in the number of new early-stage funds, reinforcing its position as a global innovation hub. The United Kingdom is a strong second, with notable activity in supporting emerging fund managers and specific tech sectors. Countries like China, Japan, and Israel are also significant players, particularly in advanced technology and AI-focused investments.

Notable Large Fund Launches

Several funds with substantial capital commitments were launched in June 2026. Notably, Sky Factory by Dreame Technology and Menlo Ventures XVII stand out with multi-billion dollar funds, indicating a strong appetite for significant investments in promising early-stage ventures. These large funds are poised to make a considerable impact on the growth and development of startups in their respective focus areas.

June 2026 demonstrates a vibrant and dynamic early-stage investment landscape, heavily influenced by advancements in AI and a growing commitment to climate technology. While the United States continues to be a dominant force, global participation is robust, with significant contributions from the UK, China, Japan, and other nations. The launch of numerous large funds underscores a healthy capital environment for innovative startups, particularly those addressing critical technological and environmental challenges.

If you have not joined the Fusion42 Community on Telegram —
it is probably time to do so.

For the ❤️ of Startups

Please let me know what you think of it, love a feedback loop 🙏🏼

❤️ Love it, keep them coming!
👍 Great, always a good read.
😊 Enjoyable, but let’s see some variety.
👎 Okay, seen better.
🛑 Get a different job.

Subscribe below and follow me on LinkedIn or Twitter to never miss an update.

How did SpaceX get to be a $2.5tn company?

DeReK WaTSoN — Fri, 19 Jun 2026 06:55:51 GMT

That’s the right question. Not “is it worth it”, how did it get there. Because it’s not a valuation. It’s a story with a massive price tag, and the story is the best one ever told.

Read the S-1. They claim a $28.5tn market, then admit a few pages later it “may not be accurate.” And how’d they get to $28.5tn? Forty-odd consultant reports they didn’t pay for, run through “management’s interpretation,” topped with three separate disclaimers telling you not to rely on it. Big-number masturbation to blow our little minds, with no consequences for any of it.

Take the story away and what’s actually there?

Starlink. A real business, $11.4bn revenue, fat margins. Everything else is just burning: $4.9bn lost last year, another $4.3bn in Q1, a $41bn hole and counting. Starlink alone doesn’t get you near $2.5tn.

So where’s the rest coming from? The bit written by Asimov, of course. The company tells you itself, describing the Moon bases and the asteroid mining: “technologies that do not exist.” Their words, in the filing.

Now watch the engineering, because this bit already happened.

Buy Twitter at $44bn. Fold it into xAI. Fold xAI into SpaceX. Every step a common-control merger, every step a fresh mark, the whole lot recast in the accounts like it was always one company.

Revenue barely moved. Losses went vertical, $6bn gone in the AI segment alone. And the mark went up the entire way.

That’s not value creation. That’s value declaration. He didn’t build the business, he re-papered it three times and the market clapped each time. Value asserted through structure, not earned through cash flow.

And here’s the cleverest bit, the same trick aimed at himself.

Everyone reads his pay packages as the ultimate alignment. Tie my comp to the share price, I only win if you win. Brilliant optics. But look at what it actually does. The targets are market-cap milestones, not cash milestones. So he’s not paid to make money, he’s paid to move the price. And the price moves on the story. He’s literally incentivised to grow the narrative, because the narrative is the thing he’s best at. The KPI isn’t a check on him. It’s a reward for doing exactly what he was already doing.

And the size does its own work. A pay number so absurd it stops reading as money and starts reading as conviction. Look what he’s tied himself to, he must believe. Same as the $28.5tn, too big to evaluate, so you stop evaluating and start feeling. The bigness is the persuasion.

Then check the votes. Class B, ten votes a share, electing a majority of the board. “Controlled company” status, which lets him skip the independent compensation committee, the one body that’s meant to sit between a CEO and his own pay. So he sets the package, through a board he elects, with votes that outweigh what he actually owns. He sells you “fully aligned” while building the exact structure that means nobody can ever hold him to it.

It’s the same play at every level.

On the market, a TAM you can’t check. On the businesses, value declared through mergers, not made. On himself, pay that rewards storytelling and a vote structure that removes the referee. Three layers, one trick: pick a number too big to test, and let our own greed and awe do the validating for him.

And it’s not hidden. It’s all in the filings. We just don’t read them.

So what comes next?

Watch for the reverse-merge into Tesla. Tesla’s already in the S&P. SpaceX, losing money, can’t get in on its own numbers. Fold it in and every index tracker on Earth becomes a forced buyer overnight. Then pile the sci-fi back on top, Optimus on the Moon, robots mining asteroids, factories in orbit, and the merged thing carries a $10tn cap. A dream too far out to ever be proven wrong in time to matter.

The story is the funding. The dream makes capital cheap, cheap money builds the one thing that works, and that makes the dream look real. The belief pays for itself. Works right up until the real business stops outrunning what the dream burns. Right now the fire’s winning.

He’s not going to build a meaningful off-world economy in his lifetime. But he’s laid the foundations for two things: a real space economy, and proof that we’re suckers, driven by greed.

Respect to the man. He’s a genius, not at rockets, at us. At reading the system and the psychology that moves the herd. He didn’t really hack physics. He hacked us and our belief system.

For the ❤️ of startups

✌🏼 & 💙
Derek

Thank you for reading. If you liked it, share it with your friends, colleagues and everyone interested in the startup Investor ecosystem.

If you've got suggestions, an article, research, your tech stack, or a job listing you want featured, just let me know! I'm keen to include it in the upcoming edition.

Please let me know what you think of it, love a feedback loop 🙏🏼

❤️ Love it, keep them coming!
👍 Great, always a good read.
😊 Enjoyable, but let’s see some variety.
👎 Okay, seen better.
🛑 Get a different job.

Subscribe below and follow me on LinkedIn or Twitter to never miss an update.

The AI labs won’t follow you down. Go build the hard thing.

DeReK WaTSoN — Mon, 15 Jun 2026 04:37:19 GMT

Let me start with the receipts, because they matter for what comes next.

In January 2024 I wrote that vertical SaaS was about to get smashed, that AI would collapse the cost of replicating specialised software, compress those famous 80–90% gross margins back toward the 45% the rest of the economy lives on, and force a shift from per-seat subscriptions to outcome pricing. People thought that was spicy. Two years on, the financial press has a name for what happened: the “SaaSpocalypse.” Software lost its premium for the first time in two decades, the SaaS Capital Index fell from around 7x ARR at the start of 2025 to roughly 3.8x by March 2026, the software ETF dropped about 30% from its September 2025 peak, and somewhere on the order of $2 trillion in market cap came off the sector. The seat-based model, Salesforce, Workday, Atlassian, is the explicit thing analysts are now downgrading, because once an agent does the work of ten seats, the link between headcount and revenue doesn’t just slow, it reverses.

If anything, it arrived more completely than I described. I said companies would stop buying specialised software and start building it in-house, because the cost of building was about to collapse. I pictured the low-code of the day, dragging blocks on a canvas. What actually showed up was more total than that: natural-language app-builders like Lovable, Bolt and v0, where a non-technical founder describes the tool they need and watches it get built and deployed, no engineers, no seat licences, no canvas. The wrapper wave softened the ground first: thousands of thin AI products that died fast but proved, publicly, how cheap software had become to produce. Then the build-it-yourself tools removed the last excuse to pay $100 a seat for something you could now just make. Same deflationary force I named, arriving harder than I’d dared predict. The destruction came, and it came early.

By November 2024 I was writing about the other half: if the old thing is dying, here’s the new thing. Software stops being a tool you operate and becomes a service that delivers the outcome, Service as Software, $4.6 trillion of work that used to need a human, done by software. Outcome pricing, agents that complete tasks rather than assist with them. I wasn’t the first to put a name to it, the idea was already moving through the VC world by late 2024, but I was on it early, while it was still a niche thesis rather than the wallpaper it’s become. It’s the consensus take of 2026 now, which is a polite way of saying it stopped being controversial.

So I’m not here to walk any of it back. Both halves landed. The destruction and the replacement both arrived roughly on schedule. I’m writing now because the part that actually matters has come into focus, the part neither of those pieces had: where the defensible thing is. Everyone can see the category. Almost nobody’s saying the right thing about the moat.

Here’s the thing though. “Go vertical” gets said a lot, but almost nobody’s actually doing it in a tight enough space to matter. They say it, then they flinch, because they’re convinced the labs will crush anything they build. That fear is the whole opportunity. The gap isn’t that nobody knows where the moat is. It’s that everybody knows, and almost nobody’s brave enough, or frankly equipped enough, to go and take it. So let me spell out exactly why the labs can’t follow you down. Once you see the mechanism, the fear looks misplaced.

The asymmetry nobody’s pricing in

OpenAI and Anthropic sell open loops. They have to. A model that serves a billion people can’t close a loop, because closing a loop means owning the outcome, the permissions, the exceptions, the liability, the write-back into someone’s real system, the moment your software becomes the system of record. “Be a safe, general assistant to everyone” and “form a specific opinion about your specific business and act on it” are opposite postures. The first is their entire franchise. The second is the only thing a real business actually wants. So they hand you intelligence and let you do the last mile.

And the last mile isn’t a mile. It’s the whole job. The distance between an AI that can tell you how to build a financial model and an AI that builds yours, watches your edits, catches the contradiction three rows down, recomputes everything downstream, and then holds a view you have to argue with, that distance is the entire business. Everything before the loop closes is a demo. Only the closed loop is a product.

Open loop vs closed loop

An open loop ends with advice. A closed loop ends with the work done, and then goes round again.

Open loop: “Here’s how you’d structure your cap table.” You now go and do it.

Closed loop: the cap table exists, it’s right, it updated when your round terms changed, and the system flagged that your new option pool breaks an assumption in the model you signed off last week.

The labs live on the left. Every business problem worth money lives on the right. The chatbot that explains the thing is commoditised the day a better model ships. The system that owns the thing end to end, with the right permissions, handling the exceptions, carrying the liability, holding state, gets more valuable that same day.

But here’s where most people, including me in 2024, get the word “outcome” wrong. We sold Service as Software as outcomes, deliver the result, not the tool. True, but incomplete, because “outcome” smuggles in a finish line. Deal closed. Model built. Task done. Loop ends. That’s not how a business works. A business doesn’t shut down the day it closes a deal, it books the deal, learns from how the deal went, adjusts, and goes again. And again. The outcome isn’t the end of the loop. It’s one turn of it.

That’s what closed actually means. Not “finished”, completes the circuit and comes back round. An organic loop, the way a living business is organic: every turn feeds the next, it compounds, it gets better the more it runs, and it never stops. A one-off outcome is just an open loop with a tidier ending. The thing worth building is the loop that learns from its own outcome and uses it to make the next one sharper.

And notice this is precisely what a single brilliant answer can never be. The labs are the apex of the one-off: ask, receive, done, forget. Stateless by design. The most powerful one-shot outcome machine ever built, and structurally incapable of going round again on your specific business, because going round again means remembering, and remembering everyone’s business is not a product you can ship to a billion people.

That’s the part I want to nail, because it inverts the usual moat argument. Most moats erode as models improve. A thin wrapper gets more disposable with every release. A closed-loop business gets stronger, because the expensive part the loop, the state, the proprietary corpus, the texture of a workflow you understand better than anyone alive, is precisely the part a model upgrade doesn’t touch. A better engine makes an open loop cheaper and a closed loop deadlier. You want to be on the side of that trade where the giants’ progress is your tailwind, not your obituary.

Why the labs can’t follow you and why nobody’s exploiting it

So if a closed loop is the whole game, why is almost nobody building one? Because it’s an engineering beast, and the reason it’s hard isn’t the number of parts, it’s that they stack, each one depending on the one above it.

At the top sits the goal, the outcome the business actually exists to move. Not a task. A direction. This is the part every harness diagram going around right now leaves out, and leaving it out is why those systems never close: a loop with no goal to measure against is just a hamster wheel. It acts, but it has no standard to judge the result by, so it can’t learn.

Below the goal sits intent, the goal broken down into the move that matters right now. What is the founder actually trying to do, as opposed to the words they typed? Get intent wrong and everything downstream executes the wrong thing perfectly.

Then memory, which is not one thing but two. Short-term, what happened in this thread, this session, the last five minutes, is cheap, volatile plumbing. Long-term, every decision, correction, contradiction and exception accumulated across months, is the opposite: durable, curated, and the single hardest thing to build. Fold them into one “memory” box, as everyone does, and you hide the fact that the long-term one is the entire moat.

And only then, at the bottom, action, actually doing the thing, with permissions, in the real world, handling the cases where reality doesn’t match the plan.

Goal, intent, the two memories, action, each one load-bearing for the next, and the whole stack measured against the goal at the top so the loop can close and come round sharper. That’s not a prompt. It’s a system, and a genuinely hard one. (How you actually wire it, and why the standard “model in the middle” diagram has the topology backwards, is its own piece. Next one.)

Which is exactly why the opportunity is sitting there unexploited. It’s not that people haven’t spotted the gap. It’s that building this is genuinely difficult, and the fear of the labs gives everyone a convenient excuse not to attempt the hard thing. So they build another open-loop wrapper, ship a single-answer chatbot, and wonder why the next model release eats them. The people who’ll win are the ones willing to do the masterpiece-level engineering the labs structurally won’t, because doing it would mean abandoning the thing that makes them labs.

The new blocker isn’t the model. It’s the knowledge.

Here’s what changed underneath all of this, and almost nobody’s saying it out loud. The models got good, shockingly good, shockingly fast. The blocker to automating real work is no longer intelligence. It’s domain knowledge.

Every company runs on know-how that lives nowhere you can point to. Some of it is in people’s heads. The rest is scattered across old email, Slack threads, support tickets, a database nobody fully understands, and a process three people half-remember. The business functions because humans vaguely know where the knowledge is and how to apply it. That’s fine for humans. It’s useless for an agent. You cannot hand an AI “we kind of do it like this, ask Dave.” An agent needs the actual rules, how a refund gets handled, how a pricing exception gets decided, how an engineer triages an incident at 3am, structured, current, and executable.

So the missing layer isn’t a smarter model and it isn’t search-over-your-documents with a chat box on top. It’s a living map of how a company actually works, turned into something an AI can execute against safely and consistently. Pull the knowledge out of all those fragmented sources, structure it, keep it current, and you’ve built the thing that sits between raw company data and reliable automation. Every company in the world is going to need that layer. Most don’t have it and can’t build it.

And here’s the wedge, the bit that’s mine. Everyone chasing this is building it for companies that already exist: a decade of Slack to excavate, ten years of tickets to mine. The whole job is archaeology, dig the knowledge out of where it’s buried.

A startup at pre-seed has nothing to dig. There’s no buried corpus. The company’s brain doesn’t exist yet, it’s being formed in real time, usually half-wrong, in the founder’s head. So you don’t excavate it. You build it from zero, as the company comes into existence. And because it’s being built rather than retrieved, it can do the thing archaeology never can: it doesn’t just record how the company works, it holds a view on whether the company is working. A map tells you where things are. This forms an opinion about a business the founder hasn’t fully figured out yet, and pushes back. That’s not memory. That’s judgment. Memory is for companies with a past. Judgment is for companies that only have a future.

Notice this is the same closed loop, one level down. A company brain that only answers questions is still an open loop, it tells you how your company works, you go act. Close it, and the brain forms the view, drives the action, watches the result, and corrects. The knowledge layer is the substrate. The closed loop is what you do with it. The corpus accreting underneath is what makes it impossible to copy.

“Vertical” was the floor, not the destination

So go vertical, but understand it’s the floor you start from, not the moat itself. Owning “legal” or “finance” isn’t narrow enough to be safe; those are still categories a lab could credibly wave at. The defensibility sits one or two levels below that: in the specific, annoying, regulated-adjacent workflow the labs will never bother to learn the texture of, where closing the loop requires knowing the exceptions that aren’t written down anywhere.

Not “accountancy.” The exact three-statement model a pre-seed founder actually needs, in their jurisdiction, with their messy assumptions, that recomputes when reality moves. Not “founder tools.” The continuous loop of observing a founder, forming a view of their company, and driving the next concrete action without waiting to be asked. That’s a closed loop by construction, and it’s not one a general assistant can ship, because forming an opinion about one company and acting on it is the opposite of being everyone’s safe co-pilot.

The loop underneath the loop

I buried this in the 2024 piece as a single section called “eating our own medicine.” It should have been the whole article. Your first 100 users teach the system things your 101st gets for free, automatically, in a way nobody outside can reconstruct. Every closed loop you run leaves a deposit, the correction a founder made, the exception that broke the template, the pattern that only shows up across hundreds of real companies. That corpus is the only asset on the board a competitor cannot buy, scrape, or prompt their way into. And it compounds.

The orchestration, master agents coordinating sub-agents, is essential. You can’t build a real vertical service without it, and it’s exactly how I’m building Arthur. What I got wrong in 2024 was treating it as the moat. It isn’t. It’s table stakes now, the labs and the frameworks ship the pattern for free. The defensibility was never the conductor. It’s the closed loop the conductor drives, and the corpus that loop is quietly building underneath.

The three objections you’re already forming

If you’ve followed this far you’ve got three rebuttals loaded. Good. Here they are, and here’s why none of them lands.

“Context windows are huge now, just feed the whole company in.” A bigger window makes the model a better reader. It does nothing to make it a writer. You can paste your entire operating history into a million-token context and the model will hand you a brilliant answer, and then the turn ends and nothing in the world changed. It didn’t write the corrected number back, reconcile the three systems that disagree, or commit a change the next turn builds on. Reading context is not owning state. The window lets the model see your business; it never lets it run it. An open loop with an infinite window is still an open loop, it just reads more before it forgets. And every improvement in context size upgrades the cheap, readable half of my loop for free while leaving the expensive half, state, write-back, reconciliatio, exactly where it was.

“Owning liability is ruinous, you’ll pay for every cleanup.” Only if “owning the outcome” means letting a probabilistic model make the consequential decision and then insuring the wreckage. It doesn’t. You earn the right to own the outcome not by being brave enough to absorb the losses, but by building the system so the losses can’t happen: the cap table isn’t generated by a model that might be off by a factor of ten, it’s computed by deterministic code that’s right every time, with the founder signing off the consequential write before it commits. The liability I’ll take on is precisely the liability the labs structurally can’t, because their architecture is the probabilistic thing that makes the number unguaranteeable in the first place. Owning liability isn’t the cost of closing the loop. It’s evidence you closed it properly. (How that architecture works is its own piece, the next one.)

“Pre-seed founders have no money, selling to startups is a graveyard.” It is, for anyone whose startups are the revenue. Mine aren’t. The pre-seed founder is the forge, not the market, the cheapest, most varied training ground there is to perfect the loop, where a founder who fails still leaves their corrections in the corpus. The company dies; the learning doesn’t. And they’re not even the buyer: the accelerators and funds who back them pay, because they want a whole portfolio executing faster, founder as distribution surface, not wallet. The ones who survive don’t graduate off the product onto a “real” enterprise tool; they grow into the enterprise tier of it, because by then the system is their company’s operating history and ripping it out means ripping out their own memory. The graveyard kills tools you can leave. You can’t leave the thing that remembers how your company works.

So, practically

If you’re building in AI right now, the question isn’t “what’s my vertical.” Everyone has a vertical. The question is: where does my loop close, and who else could possibly close it?

If your answer ends in advice, you’ve built a demo, and a model release will end you. If it ends in the work done, owned, with the liability, the exceptions, the write-back, and a corpus accreting underneath that the next customer benefits from for free, then every time OpenAI and Anthropic get better, so do you.

And stop waiting for permission. The reason this opportunity is still open isn’t that it’s hidden, it’s that the engineering is hard and the fear of the labs gives everyone an excuse to flinch. That fear is misplaced. The labs aren’t going to come down into your closed loop; they can’t, without ceasing to be the labs. The only thing standing between you and the moat is whether you’re willing to build the hard thing.

Three pieces, two years, one direction of travel: the old model dies, the new one takes its shape, and the moat turns out to live in the closed loop and the corpus underneath it. I’ve called each step before it was obvious. I’m calling this one too. And this time I’m not just writing it, I’m building it.

Go find the ugliest, most painful, most specific loop you understand better than anyone. Then close it.

For the ❤️ of startups.

✌🏼 & 💙 Derek

Please let me know what you think of it, love a feedback loop 🙏🏼

❤️ Love it, keep them coming!
👍 Great, always a good read.
😊 Enjoyable, but let’s see some variety.
👎 Okay, seen better.
🛑 Get a different job.

Subscribe below and follow me on LinkedIn or Twitter to never miss an update.

Ready Player One (2018): Sci-Fi vs Reality

DeReK WaTSoN — Sun, 29 Mar 2026 08:57:46 GMT

Sci-Fi vs Reality Did art imitate life, or did life imitate the inspiration?
Every week I watch a sci-fi film and ask a simple question: where did this idea actually come from? Did fiction imagine the future first — or did reality quietly leak into the story before we noticed?
From there, I do the reality check. What already exists in today’s tech? What’s genuinely caught up with the film? And what still doesn’t — not because no one’s tried, but because something real is in the way. Physics. Economics. Regulation. Human behaviour.
Some ideas turn out to be pointless. Some are just early. Others are waiting on one or two very specific breakthroughs.
The chicken and the egg were never separate. They were always in conversation.

Fiction gave us one OASIS. Reality gave us many, and none of them talk to each other. That’s the entire story of the metaverse, right there.

What They Imagined

In 2018, Ready Player One showed us a bleak 2045 where humanity escapes into a single, seamless virtual universe called the OASIS. It was an infinite digital playground, economy, and social fabric rolled into one. With a simple VR visor and haptic suit, you could be anyone, go anywhere, and own anything. The OASIS solved for societal collapse by offering a better, more exciting digital alternative. It was one world, one login, one identity.

The writers saw it coming: a fully immersive digital layer for life. The key assumption was that it would all be connected.

What We Actually Built

I used to think a unified metaverse was inevitable. Turns out, the economics are fucking brutal. Why would Roblox want you to take the digital goods you bought from them over to Meta’s world? There is zero incentive.

So instead of one OASIS, we built walled gardens. Ready Player One showed a unified world. Reality delivered platform-specific fortresses. Roblox built a colossal one, pulling in $3.6 billion in revenue in 2024 with nearly 83 million people logging in every day. Meta made a multi-hundred-billion-dollar bet on its own version, building the Quest hardware and Horizon Worlds.

Each platform is a powerhouse. Each is a silo. We have good-enough goggles, not a second skin. And the gap between these fortresses is where the dream of a true, open metaverse dies—and where the real opportunity for founders begins.

The Gap They Missed

The film’s biggest blind spot wasn’t the tech; it was the business model. It completely glossed over the raw, competitive drive for proprietary control. Building a single OASIS assumes a level of collaboration and open standards that capitalism just doesn’t favour. Each platform wants to be the platform, not a neighbourhood in someone else’s city.

Here’s the thing: the motivation is also different. We aren’t building the metaverse to escape a dystopian future; we’re building it to augment our slightly lonely, inefficient present. It’s about connection and utility, not desperation. The pattern I keep seeing is founders assuming people want pure escapism, when what they really want are better tools for reality. That’s a fundamental disconnect between the fiction and the market need.

The Players

The giants built the islands. The startups are building the bridges. They aren’t trying to create another OASIS; they’re creating the protocols and tools to connect the ones we already have. This isn’t a Silicon Valley story, either. The work is happening everywhere.

Decentraland (Buenos Aires, Argentina): One of the original Web3 worlds where users own the land via NFTs. It’s a live experiment in creating the kind of user-owned, interoperable world the film imagined, proving the concept beyond a single corporate entity.
The Sandbox (Hong Kong): A major player in the user-generated content space, backed by $93M and big partnerships. Their entire model is based on interoperable NFT assets, directly tackling the walled garden problem by design.
Somnium Space (Prague, Czech Republic): A VR-first metaverse that’s deeply integrated with blockchain, allowing for true asset ownership and cross-chain compatibility. They’re building for the immersive future the film promised, but with an open economic layer.
Aethir (Singapore): You can’t have a seamless metaverse without massive rendering power. Aethir is building a decentralized GPU cloud to provide the infrastructure needed for high-fidelity, interoperable experiences, solving a huge technical bottleneck.
Highstreet (Los Angeles, USA): A hybrid of online shopping and gaming, creating a marketplace for digital goods that can “world-hop.” They are explicitly building for asset interoperability, allowing brands to exist across fragmented virtual worlds.
Wilder World (London, UK): Aiming for a photorealistic, AAA-quality metaverse on the blockchain. Their focus on universal NFT standards is a direct attempt to create the ‘take it with you’ asset portability that was effortless in the OASIS.
Beam (Ho Chi Minh City, Vietnam): Instead of building another destination, Beam is building the engine. It’s a modular metaverse framework with interoperability baked in, giving developers the tools to create connected worlds from the start.
Voxels (Sydney, Australia): An early, browser-based virtual world that has already formed interoperability pacts with platforms like Decentraland. They are proof that collaboration, not just competition, is a viable path forward.
Everyworld (Tel Aviv, Israel): Using AI to help users generate interoperable 3D assets and content. This tackles the scalability problem of content creation while ensuring the assets aren’t locked into a single platform from day one.
Ixana (Cape Town, South Africa): Pure-play middleware. Ixana is building the technical plumbing—the protocols and asset bridges—that allows different metaverse platforms to communicate. They’re not building a world; they’re building the universal translator for them.

What’s Left to Build

We have startups building bridges and protocols. But the gap they missed is still the most personal one: you.

In Ready Player One, your avatar, your inventory, and your reputation were singular and persistent. We don’t have that. Your identity is fractured across a dozen platforms. The startup that still needs to exist isn’t another world, it’s a universal passport. A truly portable, user-owned digital identity that plugs into any game or virtual space.

This isn’t just an avatar. It’s a wallet, a social graph, and a reputation score that you own and control. It’s the thing that proves you are you, wherever you go in the digital realm. It’s less about building another world and more about building the passport that lets you travel between them.

The Timing Signal

A generation is growing up inside Roblox and Fortnite; for them, this isn’t a strange new world, it’s just... the world. Their expectation of digital presence and ownership is native. That’s a massive behavioural shift.

Technically, the pieces are falling into place. The launch of devices like Apple’s Vision Pro validates the market for spatial computing, pushing the hardware and developer tools forward. On-device AI is becoming powerful enough to handle real-time rendering and interaction. And open standards like OpenXR are slowly gaining ground, creating a common language for developers. The incentive to connect the islands is finally starting to outweigh the desire to own the ocean.

For the ❤️ of startups ✌🏼 & 💙

Go Deeper

Please let me know what you think of it, love a feedback loop 🙏🏼

❤️ Love it, keep them coming!
👍 Great, always a good read.
😊 Enjoyable, but let’s see some variety.
👎 Okay, seen better.
🛑 Get a different job.

Subscribe below and follow me on LinkedIn or Twitter to never miss an update.

Domain Histories: Privacy & Identity

DeReK WaTSoN — Sat, 28 Mar 2026 16:10:10 GMT

This is Domain Histories — a weekly series mapping startup domains from first principles. The best founders are students of history. Patrick Collison met Visa's founder before building Stripe. Apoorva Mehta studied Webvan's collapse before building Instacart. They used history to make specific decisions about what to build, what to avoid, and when to move. Each issue takes one domain and gives you the full map: who tried before, why they failed, where the money sits, what's shipping now, and what a founder entering today should do differently because of it. Not a market overview. Not a research report. The history that changes what you build.

You have 200 passwords. The industry that created that problem wants to sell you the solution.

Here’s the thing. The chaos of digital identity — the endless logins, the data breaches, the feeling that you’ve been copied and pasted across a thousand databases you’ll never find — isn’t the result of some grand, malicious plan.

The original sin of digital identity was a lack of one.

The Original Problem

The internet wasn’t built for people. It was built for machines to talk to each other. Identity was an afterthought, a patch. When services needed to know who you were, they did the simplest thing possible: they asked you to create a username and a password.

Every. Single. Time.

Each new service built its own little castle with its own little gate. Your bank, your email, your social network, your pizza delivery app. Hundreds of castles, hundreds of keys. This created two massive problems that define the market to this day.

First, a usability nightmare. You, the user, were forced to become a human keychain, juggling hundreds of credentials. Or, more likely, you just used the same password everywhere, turning one leaky castle into a catastrophic failure of the entire kingdom.

Second, a massive, fragmented attack surface. Your identity wasn’t a single thing you owned; it was a thousand little pieces of you, scattered across servers you had no control over. A goldmine for hackers and a business model for the data brokers who figured out how to stitch it all back together and sell it.

The Numbers

The Timeline

The history of identity is a story of bolting on solutions, with each new layer creating new problems. It unfolds in three acts.

Era 1: The Gatekeepers (Pre-2016)

Before 2016, identity was a free-for-all. Every company rolled their own login system. The dominant players weren’t tech companies; they were the old-guard credit bureaus — TransUnion, Experian, Equifax. They quietly built empires by aggregating offline data and selling it back to businesses as a proxy for trust. Their business was, and still is, data arbitrage. They profited from the fragmentation because they were the only ones with a semi-complete picture.

Startups in this era mostly tinkered with Single Sign-On (SSO), trying to create a master key for all your digital locks. But adoption was slow. The problem wasn’t seen as urgent enough to solve.

Why it ended: A European regulation lit a fire under the entire industry.
What it means for you today: The DNA of this era persists. The incumbents still hold the most data, and the default behaviour for any new app is still to build its own siloed identity system.

Era 2: The Compliance Gold Rush (2016-2024)

The General Data Protection Regulation (GDPR) changed the game. Suddenly, mishandling user data wasn’t just bad practice; it was a business-ending liability with massive fines.

Fear became the product.

This shift created companies like OneTrust. Founded in 2016, it was perfectly timed. It didn’t sell a better user experience. It sold certainty to corporate lawyers. It was a compliance engine, a suite of tools to manage consent, map data flows, and generate reports to keep regulators happy. It worked. OneTrust grew into a multi-billion dollar company by focusing entirely on the enterprise’s pain, not the user’s.

The playbook was simple: find a regulatory threat and sell the antidote. The low-hanging fruit of compliance has been picked. Building another cookie-banner tool is a death wish.

Why it ended: Compliance became table stakes, and a new technological wave — AI — introduced a level of complexity that basic compliance tools couldn’t handle.
What it means for you today: Regulation is the most powerful catalyst in this market. If you’re not tracking global data and AI policy, you’re flying blind.

Era 3: The AI Reckoning & Consolidation (2025-Now)

I used to think AI would disrupt the old data gatekeepers. Turns out, it’s their new favourite weapon. Incumbents like TransUnion and Experian are aggressively rebranding as “AI-driven powerhouses,” using machine learning to “weaponize proprietary data.” Their revenue is growing, not shrinking. They’re using AI to strengthen their grip, offering more sophisticated fraud detection and credit scoring than ever before.

This forces the market to evolve. The problem is no longer just privacy compliance; it’s AI governance. How do you prove your AI models aren’t biased? How do you manage the data they’re trained on?

Simultaneously, the market is consolidating. Fast. The cybersecurity space saw a record $96 billion in M&A in 2025. Standalone privacy and identity features are being absorbed into bigger platforms. Securiti AI, a data governance startup, was acquired by infrastructure giant Veeam for $1.73 billion in 2025. It became a feature.

What it means for you today: Competing with incumbents on data is a losing battle. Your advantage has to be in building integrated, trustworthy infrastructure that they can’t — because their business models and legacy tech are too siloed.

The Graveyard

Failure in this space often looks like a successful acquisition. Standalone companies get absorbed into larger platforms, becoming a feature, not a business. The lesson isn’t that they failed, but that building a venture-scale, independent identity company is brutally hard.

Where The Money Sits

The money sits with the gatekeepers: TransUnion, Experian, Equifax. They hold the foundational datasets on billions of people. They profit from complexity, selling access, verification, and analytics back to every other business. Their moat is a labyrinth of proprietary data, regulatory capture, and enterprise sales teams that have been in place for decades.

They are not just data brokers anymore. They are AI companies. TransUnion’s revenue grew 13% in 2025. They are successfully layering AI on top of their core asset — data — to create new, high-margin products.

The tax everyone pays is the “trust tax.” Every time a business needs to verify a user, they pay one of these giants. Every time a new regulation appears, consultants and software vendors get paid to help navigate the complexity that the incumbents thrive on.

Their blind spot? They are slow, siloed, and their products are built for their own benefit, not the user’s. They are trying to bolt on privacy and modern identity, but their core business is data exploitation. That contradiction is the opportunity.

What’s Actually Hard

The tech is not the hardest part. Zero-knowledge proofs, decentralized identifiers, biometrics — the building blocks are there.

The hard part is the sales cycle.

You are not selling software. You are selling trust. You are selling a promise that your system is more secure, more compliant, and more reliable than the status quo. Your buyer isn’t a product manager looking for a cool feature. Your buyer is a Chief Information Security Officer or a General Counsel.

Their first question isn’t “How does it work?”. It’s “Who do we sue if this goes wrong?”.

You are asking enormous, risk-averse organisations to rip out a core piece of their infrastructure — how they know who their customers are — and replace it with yours. This involves navigating procurement hell, legal reviews, and security audits that can take 18 months or more. You don’t win with a better demo. You win with better liability terms, better compliance certifications, and a story about risk reduction that the board can understand.

What’s Shipping Now

The most interesting work today is happening at the infrastructure layer, with founders who have learned the lessons of the past. They’re not building another social network with a new login. They’re building the pipes.

Killing the Password: Startups are finally tackling the original sin. Beyond Identity is pushing passwordless authentication using device-bound cryptographic keys, making phishing nearly impossible. 1Kosmos combines biometrics with a private blockchain to prove identity without storing centralised credentials.
Building Reusable Identity: The dream of a single, portable identity is getting closer. ID.me has carved out a huge niche with government agencies, offering a reusable login for accessing public services. On the decentralized front, Civic and SelfKey are using blockchain to build user-controlled identity systems.
Delivering Dev Tools: The smartest founders know that developers choose the next generation of infrastructure. uPort, now evolved into projects like Veramo, provides open-source tools for building with decentralized identity. In Web3, Privy is giving developers simple APIs to embed wallets and authentication into their apps, abstracting away the complexity.

Where This Goes

The noise and fragmentation are signs of an underlying convergence. Privacy, identity, and AI governance are not three separate markets. They are becoming one single problem: how do you establish and manage trust in a digital world?

The next 36 months will be a race to build the integrated “trust infrastructure” that solves this.

The bet is that a startup, unburdened by legacy tech and siloed business units, can build this unified platform faster and better than the incumbents who are trying to bolt these pieces together. The winner won’t just offer passwordless login, or a consent dashboard, or an AI governance report. They will offer a single platform that manages the entire lifecycle of digital trust for an enterprise.

This is an infrastructure play. The exits will look less like consumer apps and more like the $96 billion in cybersecurity M&A we saw in 2025 — strategic acquisitions by major cloud and security players.

Moves Worth Stealing

Ride the Regulatory Wave (The OneTrust Play): Don’t wait for demand to appear. Find the next GDPR-level event — a new AI liability law, a federal privacy bill — and build the tool that makes compliance automatic. Sell certainty to lawyers.
Become Essential Infrastructure (The Securiti AI Play): Don’t try to be everything. Build one critical component of the trust stack so well that a larger platform like Veeam, Microsoft, or Salesforce has no choice but to acquire you. Be the definitive solution for one piece of the puzzle.
Solve the Original Pain (The Beyond Identity Play): Go back to first principles. The password is a 60-year-old hack. Find the root of the problem that everyone else is just managing and build the thing that eliminates it entirely.

The Why Now

Three forces are creating a brief window of opportunity.

AI Creates New Urgency: The explosion of AI makes data governance and identity verification a C-suite priority. The old tools aren’t equipped to manage AI models, training data, and the new attack vectors they create.
Regulation is Expanding: The compliance wave that started with GDPR is now global and expanding to include AI. This creates new, mandatory budget items inside every large company.
The Incumbents are Buying: The record-breaking M&A in cybersecurity shows that the big players are acquiring innovation, not building it. This provides a clear path to exit for startups that can build something valuable.

The convergence of these forces means that for the first time, there is both the budget and the urgency to finally fix the internet’s broken identity layer.

If You’re Starting Monday

The Wedge: Don’t build a generic identity platform. Pick a single, painful, high-stakes verification problem in one regulated industry. Example: proving the identity of participants in decentralised clinical trials for a pharma company.
The Trap: Building a consumer-first “digital wallet” for identity. You will burn millions trying to solve a two-sided market problem and get crushed by platform players. Sell to the enterprise. They have the problem and the money.
The Customer: The Chief Information Security Officer (CISO) or the Chief Compliance Officer. The person who gets fired when a data breach or an AI audit goes wrong.
The First Hire: Your first hire should not be a growth hacker. Your first hire should be a lawyer who has spent years advising companies on data privacy and AI ethics. Your product is not just code; it’s a manifestation of legal and ethical rules.

Go Deeper

Momentum Cyber 2025 Report: Analysis of the record $96B in cybersecurity M&A. (Link)
The Information on OneTrust: The inside story of the compliance gold rush and market consolidation. (Link)
TransUnion 2025 Earnings: How incumbents are using AI to drive growth. (Link)
Rothschild & Co on AI Investment: Data on the 2025 venture capital boom driven by AI. (Link)
Contrary Research on OneTrust: A breakdown of the platform serving 75% of the Fortune 100. (Link)

For the ❤️ of startups

Arthur is the AI native startup operating system I’m building in public — not hype, but a system that turns input into structured execution and tracks founder progress. If you want to follow or use it, it’s open for early access.
→ Access Arthur

Thank you for reading. If you liked it, share it with your friends, colleagues and everyone interested in the startup Investor ecosystem.

If you’ve got suggestions, an article, research, your tech stack, or a job listing you want featured, just let me know! I’m keen to include it in the upcoming edition.

Please let me know what you think of it, love a feedback loop 🙏🏼

❤️ Love it, keep them coming!
👍 Great, always a good read.
😊 Enjoyable, but let’s see some variety.
👎 Okay, seen better.
🛑 Get a different job.

Subscribe and follow me on LinkedIn or Twitter to never miss an update.

The F42 AI Brief #064. Stop Firing Your Support Team.

DeReK WaTSoN — Wed, 25 Mar 2026 14:06:00 GMT

If you’re pouring budget into AI to replace your entire customer support team, you might want to pause. Booking.com just went on record saying its human support is the key thing that sets it apart, and they’re not wrong.

⚡ The Break

Booking.com made it clear this week: while they use AI to understand what a customer needs, their actual differentiator is having a human resolve the problem. They’re a multi-billion dollar tech company, and they’re doubling down on people.

Most founders assume the endgame for customer service is zero humans. It’s all about ticket deflection and cutting costs. The north star metric becomes how many customers you can prevent from ever speaking to a person.

Turns out, for high-stakes interactions, that’s a fast-track to losing customers. Here’s what this actually means: your customers don’t want to fight a chatbot when their flight is cancelled or their hotel room doesn’t exist. They want a solution. Fast. AI is brilliant at diagnosing the issue from a frantic email, but it’s often shit at handling the nuance and emotion required to fix it. People aren’t stupid; they know when they’re being managed by a script.

I’ve seen good projects hit the wall over this. They optimise for bot resolution rates, and watch their customer satisfaction scores plummet. The real play isn’t replacement, it’s augmentation. Use AI to give your human team superpowers. When an agent picks up a call, the AI should have already summarised the issue, pulled up the customer’s history, and suggested three possible solutions. The goal isn’t to avoid a conversation; it’s to make that conversation brutally efficient and successful.

So, look at your CX roadmap. If it’s a straight line to firing all your support agents, you’re building a frustration engine, not a service. Start measuring your team on first-contact resolution and customer happiness. That’s the data that matters.

🧠 What Quietly Changed

Yann LeCun’s attack on AI’s ‘common sense’ problem — While everyone is distracted by the latest video generator, Yann LeCun’s team dropped research on a thing called LeWorldModel. It’s a direct shot at one of AI’s biggest hurdles: learning how the world works just by watching, like a baby does. The research targets “JEPA collapse” in pixel-based models, which is a technical way of saying it’s trying to stop the AI from getting lazy and ignoring the details when predicting the future of a scene. This is the foundational work required for agents that can actually plan and reason in the real world, not just generate plausible-looking clips.

OpenAI’s $1B ‘goodwill’ fund — The OpenAI Foundation announced a $1 billion grant programme to ensure AI “benefits all of humanity.” This isn’t just a press release about ethics. For a founder, this is a giant flare telling you where the wind is blowing. Building with a clear, defensible social benefit is moving from a nice-to-have to a core requirement for survival. This kind of money doesn’t just fund non-profits; it shapes the entire narrative, pre-empts regulators, and sets the bar for what is considered an acceptable use of powerful technology. If your pitch deck doesn’t have a slide on this, you’re already behind.

The legal storm brewing for AI agents — We’re seeing more research into “agentic AI” and its geopolitical impact, but the quiet part is the immediate legal risk for builders. The moment your AI agent can act on a user’s behalf—making a purchase, sending an email, trading a stock—it can also break the law. When an autonomous agent executes a trade that violates international sanctions, who is liable? You? The user? The company that owns the platform? Most legal frameworks are completely unprepared for this. If you’re building agents, you need to be talking to lawyers now about liability, kill-switches, and how you log an agent’s ‘intent’.

🧪 The Edge Case

There’s a growing chorus of people confidently predicting an AI-driven dystopia. They speak with absolute certainty about how AI will upend society, destroy industries, and change the very nature of humanity.

The reality is, making investment decisions or building a company based on these arrogant predictions is a massive mistake. It presumes the future is a solved problem and that we are all just passengers. It ignores the simple fact that huge, complex systems—like global society—don’t just roll over for new technology. People adapt. They regulate. They find new ways to create value. The breathless predictions of flying cars in the 70s or the paperless office in the 90s should have taught us some humility.

Here’s the thing smart people are missing: they mistake a powerful tool for a deterministic force. They get so wrapped up in the technology’s potential that they forget about human behaviour, economics, and politics.

The builder’s lesson isn’t to ignore AI. It’s to stay grounded. Don’t build for a science fiction novel; build for a real customer with a real problem today. The most resilient startups will be those that use AI to solve a concrete need better than anyone else, while remaining agile enough to navigate the changes as they come. Betting the farm on a single, extreme vision of the future—whether a utopia or a dystopia—is just a dressed-up form of gambling. Solve today’s problems, and you’ll earn the right to solve tomorrow’s.

🎯 This Week’s Bet

BUILD: An AI-powered co-pilot for your customer support team. It should listen to calls and read chats in real-time to instantly surface the exact knowledge base article, previous ticket, or account detail the human agent needs to solve the problem. The goal is not to replace the human, but to make them impossibly fast and accurate.

BAIL: Chasing 100% automated customer support resolution. Set a rule: if a customer asks for a human twice in a row, the bot escalates immediately. The brand damage from trapping a frustrated customer in a chatbot loop is worth far more than the salary you think you’re saving.

For the ❤️ of startups ✌🏼 & 💙

We are building Arthur, your agentic Startup OS that gets **it done for you.

CLICK HERE TO GET EARLY ACCESS AND YOUR FREE FUEL UNITS

Please let me know what you think of it, love a feedback loop 🙏🏼

❤️ Love it, keep them coming!
👍 Great, always a good read.
😊 Enjoyable, but let’s see some variety.
👎 Okay, seen better.
🛑 Get a different job.

Subscribe below and follow me on LinkedIn or Twitter to never miss an update.

For the ❤️ of startups

✌🏼 & 💙
Derek

Know Your Models

DeReK WaTSoN — Mon, 23 Mar 2026 08:37:12 GMT

The rules changed. Here’s the new playbook. 30 principles for building startups in an AI world — what worked before, what broke, and what founders need to do now. Not theory. Tactics that ship.
New principle every Monday

Know Your Models

Choosing the wrong AI model is now a product and cost architecture mistake. Founders need to match model capability to task rather than defaulting to the most powerful or cheapest option. You were told that deep technical expertise in your stack was non-negotiable. Turns out, the stack is now a menu, and your most important technical skill is knowing what to order.

Why This Mattered Before

The conventional wisdom was that a technical founder must deeply understand their core technology. And it was right. When you built everything from scratch, from the database to the front-end framework, that deep knowledge was the difference between shipping and stalling. Knowing the internals of your stack was how you optimised performance, debugged esoteric issues, and ultimately, built a defensible product.

Your unique advantage was in how you put the bricks together because you knew the properties of every single brick. This was the gospel of YC, the definition of a “great hacker,” and for a decade, it was the only way to win.

The Graveyard

I haven’t seen a single, high-profile startup death with “built their own model” listed as the official cause on the tombstone. The reality is more subtle and far more common. In the cohorts I’ve worked with, I’ve watched brilliant teams burn their first £500k and nine months of runway trying to build a bespoke NLP model for something like sentiment analysis or document summarisation.

They’d be heads down, trying to eke out a 2% accuracy gain, convinced this was their moat. Then a new foundation model from Anthropic or Google would drop, delivering better performance out of the box through a simple API call. The startup’s “moat” became a millstone overnight — expensive to maintain, slow to improve, and fundamentally uncompetitive. They didn’t die in a blaze of glory; they just became irrelevant.

What AI Actually Changed

The foundational technical competence has shifted from building the engine to knowing how to drive. You no longer need to know the internals of PostgreSQL. You need to know that Claude 3.5 Sonnet is faster and cheaper for enterprise-grade text analysis, while GPT-4o might be better for complex, multi-step agentic workflows. This is the new literacy.

The game is no longer about building your own model from scratch. The capital required is staggering. Cohere raised $500 million in 2025 to focus on enterprise models. xAI is backed by billions. You cannot compete with that. What they’ve done is commoditise raw capability. Your job is no longer to create the capability, but to intelligently apply it. The core engineering challenge has moved up the stack from model creation to model orchestration.

The New Playbook

Audit tasks, not your tech. Map every core task your product performs to the required AI capability. Does this task need world-class reasoning, or just fast, accurate classification? Does it need creative generation or factual extraction? Don’t use a sledgehammer to crack a nut. Use a small, fast model for simple jobs and save the expensive, powerful models for the tasks that truly need them.
Build a model “palette”. Don’t get locked into a single provider. Your architecture should allow you to route tasks to different models based on performance and cost. Set up a simple evaluation framework to continuously benchmark models like those from OpenAI, Anthropic, Google, and open-source alternatives on Hugging Face. A simple API call change could cut your COGS in half or double your performance overnight.
Treat cost-per-token as a core business metric. Your model choice directly impacts your cost of goods sold, and therefore your pricing and margin. A model that’s 98% as good but 80% cheaper might be the right business decision, even if it’s not the absolute peak of technical performance. This is no longer just an engineering choice; it’s a fundamental product and finance decision.
Hire for AI literacy, not ML research. Your first key hire isn’t a PhD who wants to build a model from the ground up. It’s a pragmatic engineer who lives on the leaderboards, understands the trade-offs between different models, and can quickly integrate and fine-tune them. They’re an architect and a plumber, not a research scientist.

The Warnings

First, avoid the “Sledgehammer Fallacy.” I see founders using the most powerful model on the market for every single API call, from parsing a simple user query to generating a button colour. This burns cash and introduces unnecessary latency. It’s lazy architecture, and it will kill your margins.

Second, beware of “Model Monogamy.” Committing your entire product to a single AI provider is a huge strategic risk. When they change their pricing, deprecate a model you rely on, or suffer a major outage, your business is at their mercy. Diversify your model dependencies just like you would any other critical part of your supply chain.

The Bottom Line

Your first AI hire isn’t a researcher. It’s an architect who knows which model to use, when, and precisely what it costs.

Part of Startup Principles for an AI World — 30 principles for building in the new era. New issue every week.

Enjoy and hope to see you soon.
❤️ & ✌🏼

Derek

Please let me know what you think of it, love a feedback loop 🙏🏼

❤️ Love it, keep them coming!
👍 Great, always a good read.
😊 Enjoyable, but let’s see some variety.
👎 Okay, seen better.
🛑 Get a different job.

Subscribe below and follow me on LinkedIn or Twitter to never miss an update.

Minority Report (2002): Sci-Fi vs Reality

DeReK WaTSoN — Sun, 22 Mar 2026 21:18:09 GMT

Sci-Fi vs Reality Did art imitate life, or did life imitate the inspiration?
Every week I watch a sci-fi film and ask a simple question: where did this idea actually come from? Did fiction imagine the future first — or did reality quietly leak into the story before we noticed?
From there, I do the reality check. What already exists in today’s tech? What’s genuinely caught up with the film? And what still doesn’t — not because no one’s tried, but because something real is in the way. Physics. Economics. Regulation. Human behaviour.
Some ideas turn out to be pointless. Some are just early. Others are waiting on one or two very specific breakthroughs.
The chicken and the egg were never separate. They were always in conversation.

Gesture interfaces looked cool in 2002. Twenty years later, we still use mice. Here’s the thing: while Hollywood sold us a vision of effortless, dance-like control, reality delivered something much clunkier. The gap between the two is where the real work—and the real opportunity—begins.

What They Imagined

In the world of 2054, John Anderton didn’t just look at data; he stepped inside it. Using gloved hands on a giant transparent screen, he swiped, pinched, and threw information around to stop crimes before they happened. He wasn’t clicking or typing; he was conducting data.

It was a beautiful, high-bandwidth solution to information overload that made the mouse and keyboard look like a chisel and stone. The writers saw a future where our physical movements could directly manipulate a world of abstract information.

What We Actually Built

We didn’t get the giant, public screens. We got personal ones, strapped to our faces.

Apple’s Vision Pro, which landed in 2024, brought a version of this to life. It uses subtle eye-tracking and finger-pinches to control apps floating in your living room. It’s an intimate, small-scale version of Anderton’s interface, focused on personal productivity and entertainment, not public crime-fighting.

Meanwhile, Magic Leap has been on a long, hard journey since 2010. After raising billions and generating massive hype, they pivoted their AR platform toward the enterprise market. Their story is a lesson in how difficult it is to build this tech from scratch and find a market that will actually pay for it.

Meta is taking a different path, prototyping neural wristbands that read muscle signals in your arm. The idea is to control devices with intention, not big, sweeping gestures. Minority Report showed us the what. Apple, Magic Leap, and Meta are all grappling with the how.

The Gap They Missed

The writers got the vision right, but they skipped over the messy, human details. Turns out, holding your arms up like an orchestra conductor for eight hours a day is fucking exhausting. It’s called ‘gorilla arm,’ and it’s a well-known problem in UI design. Physics is a stubborn thing.

The film also ignored cognitive load. Managing files in 2D is hard enough; navigating a 3D spatial environment requires a whole new mental model that most people haven’t developed yet.

But the biggest gap they missed is fragmentation. In the film, the gesture interface was a universal standard. In reality, we have a collection of walled gardens. An app built for VisionOS doesn’t work on Meta’s Horizon OS. This lack of interoperability is the single biggest brake on the spatial computing revolution. Founders who ignore this do so at their peril.

The Players

What’s Left to Build

The real opportunity isn’t just another headset. It’s solving the problems the big players are too busy—or too strategic—to fix. We’re still missing the fundamental tools and infrastructure.

The company that builds the cross-platform ‘Figma for spatial computing’ or the ‘Unity for enterprise AR’ will own a critical piece of the stack. Right now, developers have to choose a tribe. The startup that creates the agnostic, interoperable layer—letting creators build an experience once and deploy it across Vision Pro, Quest, and whatever comes next—wins.

The wedge is to stop focusing on the destination (the perfect AR experience) and start building the roads and bridges. Solve the workflow problem for the thousands of developers who don’t want to be locked into a single ecosystem.

The Timing Signal

So, why now? Two things have changed.

First, Apple’s entry with Vision Pro validated the space and, more importantly, started training consumer behaviour. People are finally learning the grammar of spatial interaction—the pinches, the glances, the gestures. Apple is spending billions to create the market for you.

Second, the incumbents are committing capital at a scale that signals a point of no return. Meta is borrowing billions specifically for AI to power its future platforms. The foundational tech is maturing, and the giants are placing their bets. The gap between fiction and reality is where startups live, and that gap is finally funded.

For the ❤️ of startups ✌🏼 & 💙

Please let me know what you think of it, love a feedback loop 🙏🏼

❤️ Love it, keep them coming!
👍 Great, always a good read.
😊 Enjoyable, but let’s see some variety.
👎 Okay, seen better.
🛑 Get a different job.

Subscribe below and follow me on LinkedIn or Twitter to never miss an update.

Domain Histories: Data Infrastructure & Analytics

DeReK WaTSoN — Sat, 21 Mar 2026 07:47:53 GMT

This is Domain Histories — a weekly series mapping startup domains from first principles. The best founders are students of history. Patrick Collison met Visa's founder before building Stripe. Apoorva Mehta studied Webvan's collapse before building Instacart. They used history to make specific decisions about what to build, what to avoid, and when to move. Each issue takes one domain and gives you the full map: who tried before, why they failed, where the money sits, what's shipping now, and what a founder entering today should do differently because of it. Not a market overview. Not a research report. The history that changes what you build.

Every company says they’re data-driven. Almost none can query last week’s data.

The promise was simple: get all your data in one place, ask any question, get an instant answer. The reality is a tangled mess of pipelines, broken dashboards, and six-figure bills for a warehouse that mostly stores junk. For decades, we’ve been fighting the same war with different weapons.

Here’s the thing: the problem isn’t the technology. It’s our understanding of it. We keep buying the next magic box, hoping it will finally make sense of the chaos. It won’t.

But if you understand the history — the sequence of problems and the ghosts of solutions past — you can see where the real opportunity is. It’s not in building a better warehouse. It’s in what happens next.

The Original Problem

Before the cloud, data lived in a basement.

It sat in hugely expensive, on-premise servers from companies like Oracle or Teradata. These were monoliths. Slow, rigid, and eye-wateringly expensive. Getting data in was a project. Getting it out was a nightmare.

Want to know which customers are likely to churn? You’d file a ticket with the Business Intelligence (BI) team. Weeks later, if you were lucky, you’d get a static PDF. The data was already out of date. The business had already moved on.

This created a priesthood of data gatekeepers. They held the keys to the kingdom, and the rest of the company had to beg for scraps. The original problem wasn’t a lack of data. It was a lack of access.

The Numbers

These numbers aren’t just figures on a balance sheet; they represent fundamental shifts in how value is created and captured in the data ecosystem. From the sheer scale of cloud infrastructure to private market bets on architectural choices, they show the stage on which the drama of data infrastructure unfolds. This financial backdrop created pressure, making legacy systems untenable and paving the way for the unbundling and rebundling that defines our timeline.

The Timeline

The story of data infrastructure is a story of unbundling and rebundling. We break apart the monolith, perfect the pieces, then glue them back together in a new shape.

Era 1: The On-Prem Monoliths (Pre-2012)

What Enabled It: The simple need for businesses to report on their operations amidst complex, internal data.
The Players: Oracle, Teradata, IBM. The heavyweights of enterprise IT.
The Playbook: Sell a massive, integrated hardware and software package for millions of dollars. Lock customers into multi-year contracts for support and maintenance. The business model was based on friction and control.
Why It Ended: The cloud. AWS made storage and compute cheap, elastic, and accessible. Why buy a million-dollar box for your basement when you could rent a supercomputer by the hour?
For Founders Today: Every founder in this space is, in some way, fighting the ghost of the on-prem monolith. Your product must bet against rigidity.

Era 2: The Cloud Data Warehouse (2012-2020)

What Enabled It: Cheap cloud storage (AWS S3) and compute (EC2).
The Players: Snowflake, AWS Redshift, Google BigQuery.
The Playbook: Separate storage from compute. This was the killer insight. You could store petabytes of data for pennies and only pay for compute when you were actively running a query. It destroyed the on-prem economic model. Suddenly, data warehousing was accessible to startups, not just Fortune 500s.
Why It Ended (or Evolved): The warehouse worked. Too well. Companies dumped everything they had into it. The new problem wasn’t storing the data; it was making sense of it. The warehouse became a data swamp, creating a new set of problems: getting data in (ETL), cleaning it (Transformation), and getting it out to where people actually work (Reverse ETL).
For Founders Today: The central warehouse is now a given. Don’t try to replace it. Bet on connecting with the data warehouse.

Era 3: The Modern Data Stack & The Lakehouse Wars (2020-2024)

What Enabled It: The success of the cloud data warehouse created a need for a surrounding ecosystem of tools.
The Players: Snowflake (representing the Data Warehouse) vs. Databricks (championing the Data Lakehouse).
The Playbook: Unbundle the stack. A best-of-breed tool emerged for each job: one for ingestion (ETL), one for transformation, one for reverse ETL, one for observability. The stack was modular but complex. This era was defined by the great architectural debate:
- Data Warehouse (Snowflake): Perfect for structured, analytical data. Think SQL queries and BI dashboards.
- Data Lakehouse (Databricks): A hybrid approach designed to handle both structured data and the unstructured stuff needed for machine learning (images, text, logs).
Why It Ended (or Evolved): Generative AI crashed the party. Suddenly, the “unstructured stuff” wasn’t a niche problem for data scientists anymore. It was the fuel for the next generation of products. The Lakehouse argument got a massive boost.
For Founders Today: The debate is settling. Bet on handling diverse data types.

Era 4: The AI Data Layer (2025-Now)

What Enabled It: The Cambrian explosion of large language models (LLMs) and AI applications.
The Players: Snowflake, Databricks, and the hyperscalers (AWS, Azure, Google).
The Playbook: Your data platform is now your AI platform. It’s no longer just for analysts running queries; it’s for applications fetching data to feed to a model. The new features have names like Cortex AI (Snowflake) and are designed to let companies build AI-powered applications directly on top of their existing data. Microsoft’s Azure growth is being turbo-charged by serving AI inference. The game has changed from passive analytics to active intelligence.
For Founders Today: This is the new frontier. Data isn’t just for looking at in a dashboard. It’s an ingredient for building intelligent products. The entire value proposition of the data stack has been elevated. Bet on data platforms becoming AI platforms.

The Graveyard

The research for this space doesn’t show a tidy list of high-profile flameouts. Why? Because in infrastructure, you don’t typically explode. You just fade away. You get absorbed, open-sourced, or squeezed into a niche until you become irrelevant.

The companies that die here don’t fail because their tech is bad. They fail because of a few repeating patterns:

The “Single Pane of Glass” Fallacy: They try to build one platform to do everything — ingestion, transformation, warehousing, visualisation, and AI. They end up building a mediocre version of five different things and get crushed by best-of-breed point solutions.
Getting Squeezed by the Platform: They build a useful feature on top of Snowflake or Databricks. It works. Then, Snowflake or Databricks builds a “good enough” version of that feature natively and gives it away for free. Game over.
Ignoring the Workflow: They build a beautiful tool that requires users to leave the place where they actually work. A salesperson lives in Salesforce. A marketer lives in Marketo. If your tool requires them to open another tab, you’ve already lost.

The lesson is brutal: don’t compete with the giants on their turf. And don’t try to boil the ocean. Find one, painful, specific problem and solve it ten times better than anyone else.

Where The Money Sits

Follow the CAPEX.

The entire data world is built on the foundations laid by three companies: Amazon (AWS), Microsoft (Azure), and Google (GCP). They own the servers, the storage, and the network. They are the landlords. Everyone else in the data ecosystem, including Snowflake and Databricks, is a tenant. They pay the rent in the form of massive cloud bills.

The hyperscalers are now moving up the stack, offering their own data warehouses, machine learning platforms, and AI models. They are in a “one-lane, CAPEX highway that is capacity-constrained,” furiously building out the infrastructure for the AI wave.

The two other major power centres are Snowflake and Databricks. They created a critical abstraction layer on top of the raw cloud infrastructure, making it usable for mere mortals. They captured immense value by simplifying the complexity of the cloud providers.

The money flows from the end customer, through the application and tooling layer, down to the data platforms (Snowflake/Databricks), and finally settles at the cloud providers (AWS/Azure/GCP). The closer you are to the foundation, the larger the pool of money.

What’s Actually Hard

You’ll think you’re building a data tool. You’re actually building a change management consulting business wrapped in software.

This was the quiet killer of many Modern Data Stack hopefuls. They built fantastic ingestion pipelines, beautiful transformation tools, and slick dashboards. But they forgot that data quality isn’t a technical problem in isolation. It’s a human problem. You can install the most sophisticated data observability platform on the planet, but it can’t fix a sales team that enters “asdf” in the company name field in Salesforce. The mantra “garbage in, garbage out” is the iron law of data. Solving data quality is less about tech and more about changing human behaviour, aligning incentives, and fixing broken operational processes.

When Modern Data Stack vendors promised “self-service analytics,” they often delivered a powerful toolbox without a manual, or worse, without understanding the internal political landscape needed for adoption. They forgot that the final 10% – the brutal grind of cleaning messy data, reconciling discrepancies across systems, and training users – is where 90% of the effort lies. Your biggest competitor isn’t another startup; it’s export_to_csv.xlsx combined with a frustrated team that’s simply given up on accurate data. To win, you have to be an order of magnitude better, not just incrementally so, by fundamentally altering how people interact with their own data.

What’s Shipping Now

The arms race has moved from storage to intelligence.

Snowflake isn’t just talking about AI; they’re shipping it. With the launch of Cortex AI and Snowflake Intelligence, they’re offering managed services that let developers build AI applications without being machine learning experts. The rebound in their business shows it’s working (Forbes, Seeking Alpha).
Microsoft Azure is seeing “eye-popping” growth by becoming the engine for AI inference at scale, most notably for services like ChatGPT. They are bundling AI capabilities directly into the cloud platform (The Cube Research).
Amazon‘s CEO is explicit: he sees AI as the catalyst to double AWS sales projections. They are aggressively rolling out new AI and ML services within the AWS ecosystem, betting that the future of cloud is intelligent applications (Economic Times).

The theme is clear: the platforms are moving up the stack to provide not just the data, but the intelligence derived from it.

Where This Goes

The line between the data warehouse and the AI platform is dissolving.

For a decade, the data warehouse was a system of record, a place for historical analysis. Now, it’s becoming a live, active part of the product. The future isn’t about analysts running SQL queries to build dashboards. It’s about applications programmatically querying the data platform to get context for an AI agent. Imagine a customer support AI agent that, in real-time, queries your entire customer history—comprising structured CRM data, unstructured support tickets, call transcripts, and product usage logs—all unified within the data warehouse/lakehouse, to provide hyper-personalised responses or even proactive solutions. This is the “warehouse as nervous system” in practice.

The battle between Snowflake (the king of structured, analytical data) and Databricks (the champion of unstructured, AI-ready data) will define the next five years. Databricks’ soaring private valuation suggests the market believes that a platform built for AI from the ground up has the long-term advantage.

The bet is this: the company that stores your data is in the best position to own your AI strategy. The warehouse/lakehouse is becoming the central nervous system for enterprise intelligence.

Moves Worth Stealing

History provides the playbook.

Snowflake’s “Meet Them Where They Are” Play. Snowflake didn’t try to build its own cloud. It built on top of AWS, then Azure, then GCP. Instead of fighting the incumbents, it turned them into a channel. It made adoption frictionless by going to where its customers already were. The Lesson: Don’t make your customers cross the street. Integrate deeply into their existing stack and workflow.
The dbt Community Playbook. dbt Labs built an open-source tool that became the industry standard for data transformation. They didn’t do it with a huge sales team. They did it by building a fanatical community of practitioners. They defined a new job title—the “Analytics Engineer”—and became the centre of that universe. The Lesson: Win the hearts and minds of the individual developer or analyst. They are your Trojan horse into the enterprise.
The Connector Factory Playbook. An entire category of multi-billion dollar companies was built on a simple, unglamorous premise: build hundreds of reliable pipes to move data from source systems (like Salesforce) into a data warehouse. They focused on one tedious, painful job and executed it flawlessly. The Lesson: Don’t underestimate the value of solving the most boring problem in the stack. Reliability is a killer feature.

The Why Now

The ground is shifting. For the first time in years, there’s an opportunity for new players to emerge.

AI is the New UI. Every software application is being re-imagined as an intelligent, conversational interface. This requires a fundamentally new relationship with data. The old batch-oriented analytics stack isn’t built for this real-time, AI-driven world. This opens massive white space for tools that convert raw enterprise data into AI-ready feature stores, or services that manage data versioning and lineage specifically for model retraining.
The Giants are Distracted. Snowflake and Databricks are locked in a head-to-head battle for the Fortune 500, focusing on core platform features at scale. The hyperscalers are focused on the foundational model and compute layer. This creates a specific gap for startups to focus on mid-market implementation pain, develop vertical-specific data products (e.g., AI-driven insights for manufacturing operations), or build governance and observability tools tailored for the complex outputs of AI agents rather than just human-generated data.
Talent is Abundant. The “Modern Data Stack” wave created an entire generation of engineers who know how to build with these tools. The expertise that was once confined to a few tech companies is now widespread. This means you can build a highly skilled team capable of tackling sophisticated data challenges much more efficiently than ever before.

The window is open to build the tools that will power the next decade of intelligent applications.

If You’re Starting Monday

Don’t build another dashboard.

The Wedge: Find one, painful, “last mile” problem and build a killer solution. Don’t build a generic data quality platform. Build a tool that uses AI to automatically clean and standardise sales activity data from Gong and Outreach before it ever pollutes the warehouse. The pain of “clean sales activity data” is profound because sales reps often game the system (e.g., logging fake activities), AI transcripts from call tools can create duplicate or messy records, and inaccurate activity data completely breaks attribution models for marketing, making future sales forecasting a nightmare. Solve that specific, excruciating problem. Be vertical. Be specific. Be indispensable.
The Trap: Trying to be the “single pane of glass.” Every first-time founder in this space draws a box in the middle of a diagram called “Our Platform” with arrows coming in from everywhere. This is the road to ruin. You will die a death of a thousand integrations.
The Customer: Your user is the Analytics Engineer. They are overworked, under-appreciated, and judged on the reliability of their data pipelines. Your buyer is the Head of Data, who is terrified of their Snowflake bill and the questionable quality of the data inside it. Solve the user’s pain to get the buyer’s budget.
The First Hire: A developer advocate who can write. Not a salesperson. You need someone who can build a following by teaching people how to solve real problems. Your early growth will come from content and community, not cold calls.

Your Monday Checklist:

Identify Your Wedge: What specific, painful “last mile” data problem will you solve? Be vertical.
Map Integration Surface: How will your product seamlessly integrate into existing data stacks (Snowflake, Databricks, CRMs)?
Assess Platform Squeeze Risk: Is your core feature immune to being built by an incumbent platform within 18 months?
Distinguish User vs. Buyer: Who uses it daily? Who signs the cheque? Design for both.
Plan First Hire: Who will build your community and tell your story?

Go Deeper

Amazon CEO sees AI doubling prior AWS sales projections to $600 billion by 2036 — On the scale of the incumbents’ AI ambitions.
Databricks Now Valued At Over 2X Snowflake — The clearest signal of the Warehouse vs. Lakehouse battle.
Snowflake Stock Up 49%. Learn Whether AI Agents Make $SNOW A Buy — How Snowflake is shipping AI features like Cortex AI.
Cloud Quarterly – Azure’s AI Pop, AWS’ Supply Pinch and Google’s Execution — A breakdown of how hyperscalers are monetising AI.
6 Charts That Show The Big AI Funding Trends Of 2025 — Evidence of the capital flooding the AI application layer.
Amazon.com Announces Fourth Quarter Results — The raw numbers behind the AWS machine.
The 2025 AI Index Report | Stanford HAI — Broader context on AI funding trends.

For the ❤️ of startups

Please let me know what you think of it, love a feedback loop 🙏🏼

❤️ Love it, keep them coming!
👍 Great, always a good read.
😊 Enjoyable, but let’s see some variety.
👎 Okay, seen better.
🛑 Get a different job.

Subscribe and follow me on LinkedIn or Twitter to never miss an update.

The F42 AI Brief #063. More Compute. More Control.

DeReK WaTSoN — Wed, 18 Mar 2026 15:36:31 GMT

US House Votes to Expedite Datacenter Builds for AI Infrastructure

The House passed H.R. 7644, the “Accelerate Large-scale Artificial Intelligence Infrastructure Act,” this week, which allows for accelerated construction of datacenters crucial for AI development (MSN). This move aims to streamline regulatory hurdles, meaning faster compute access for founders.

⚡ The Break

Here’s the thing: this isn’t just about cutting red tape. It’s a clear signal that the infrastructure required to power the next wave of AI is being prioritised. If you’re building compute-intensive AI applications, this means the playing field for securing resources could shift faster than anticipated. Companies that can leverage this new regulatory environment to build or access dedicated AI compute will gain a significant advantage. Don’t wait for the datacenters to appear; start modelling your compute needs now and explore long-term partnerships or commitments with infrastructure providers who can benefit from these accelerated timelines. Plan for available compute to come online sooner.

What Quietly Changed

China’s AI Labelling Mandate

On the 14th of March, the Cybersecurity Administration, Ministry of Industry and Information Technology, Ministry of Public Security, and State Administration of Radio and Television in China promulgated “Measures for Labeling of AI-Generated Synthetic Content” [2]. This new regulation, effective from March 7, 2025, mandates clear labelling for all AI-generated content. Most founders assume content generation is an unregulated wild west. The reality is, disclosure is becoming mandatory. If your AI generates content, even for internal use, you need to implement watermarking or metadata-based labelling systems now. This isn’t just a Chinese problem; it sets a precedent for global content provenance, and getting ahead of it protects you from future compliance headaches.

OpenClaw Ban in China

China banned the use of “OpenClaw” on government computers this week, citing security concerns such as prompt injection attacks [1]. This isn’t about protecting state secrets alone. It’s about a fundamental distrust of the security posture of AI agents in sensitive environments. If you’re deploying AI tooling within your organisation, particularly open-source or third-party solutions, you must conduct a thorough security audit focusing on prompt injection vulnerabilities and data exfiltration risks. The Chinese government found flaws; assume your adversaries will too. Implement strict sandboxing and access controls for all AI agents.

Grammarly’s AI Feature Pullback

Grammarly quietly abandoned its ‘AI Expert Review’ feature, which promised human-quality edits, hinting at the difficulties of truly human-like AI writing (The Guardian). This wasn’t a PR announcement; it was a quiet deletion from their offering. What most people missed here is that even well-funded companies with deep NLP expertise are struggling to deliver consistently nuanced, human-level content generation. For founders betting on AI completely replacing human writers or editors, this is a wake-up call. Reallocate your resources from trying to achieve perfect human emulation to building AI that augments human capabilities, focusing on efficiency and specific, measurable improvements, not wholesale replacement.

The Malicious AI Agent Threat

Security experts warned this week that malicious AI agents are capable of collaborating to execute sophisticated cyberattacks (TechRadar). This moves beyond individual agent threats; it means your security models need to account for orchestrated attacks from multiple AI entities. They cited an example where AI bypassed access restrictions by unearthing a hardcoded secret key in an application’s code repository. If you’re building any system accessible by AI or incorporating AI into your security stack, assume a coordinated, multi-agent attack is possible. Your defence strategy must evolve to include real-time behavioural analysis of AI agents, not just signature-based detection, and focus on anomaly detection at system boundaries.

The Edge Case

Adobe’s CEO, Shantanu Narayen, is stepping down amidst concerns about AI’s disruptive impact on creative industries (TipRanks). AI-driven capabilities are set to fundamentally alter creative work production.

The thing smart people are missing here isn’t just that AI will change how designers work. It’s that the very business model for creative tools is now under threat. I’ve seen good projects hit the wall over this. If the market shifts from purchasing tools to a subscription for outcomes — or, worse, if AI makes some creative tasks effectively free — then the entire revenue strategy of companies like Adobe needs a rethink. The builder’s lesson is specific: if your product is built on charging for creative tools, you need to either build AI that augments human creativity in ways no pure AI can replicate, or you need to radically rethink where your value creation lies. Are you selling paintbrushes, or are you selling finished masterpieces?

This Week’s Bet

BUILD: Automated, privacy-preserving AI solutions for child online safety that restrict access to inappropriate websites globally and specifically comply with UK regulations. Focus on solutions that can integrate seamlessly at the network level or within popular educational platforms.

BAIL: Stop developing AI tools that impersonate human writers or generate content without clear disclosure. The backlash, as seen with Grammarly’s abandoned ‘AI Expert Review’ feature and the push for an ‘AI-free’ logo, is becoming too significant to ignore. The ethical and regulatory landscape is shifting too fast to make this a sustainable bet.

Go Deeper

China bans OpenClaw [1]
Measures for Labeling of AI-Generated Synthetic Content [2]

For the ❤️ of startups ✌🏼 & 💙

We are building Arthur, your agentic Startup OS that gets **it done for you.

CLICK HERE TO GET EARLY ACCESS AND YOUR FREE FUEL UNITS

Please let me know what you think of it, love a feedback loop 🙏🏼

❤️ Love it, keep them coming!
👍 Great, always a good read.
😊 Enjoyable, but let’s see some variety.
👎 Okay, seen better.
🛑 Get a different job.

Subscribe below and follow me on LinkedIn or Twitter to never miss an update.

For the ❤️ of startups

✌🏼 & 💙
Derek

Own The Wedge

DeReK WaTSoN — Mon, 16 Mar 2026 10:07:13 GMT

The rules changed. Here’s the new playbook. 30 principles for building startups in an AI world — what worked before, what broke, and what founders need to do now. Not theory. Tactics that ship.
New principle every Monday

AI commoditised point solutions — founders need a specific high-value wedge that earns the right to expand, rather than building horizontal AI tools that incumbents bundle as features within a product cycle. You were told to start with a tiny niche, a single feature done ten times better. Turns out, that advice is now a trap. The defensible wedge isn’t a feature anymore; it’s the entire workflow you make obsolete.

Why This Mattered Before

The conventional wisdom was to start narrow and deep. Paul Graham’s essays hammered this home: own a tiny piece of the market completely before you even think about expanding. And it was right. Building anything, even a single, focused tool, required serious engineering effort. That effort was your moat.

A startup could build a defensible business by solving one problem exceptionally well because incumbents were too slow or unfocused to bother competing on such a small feature. You could earn passionate users, build from a position of strength, and expand from there. This playbook worked for years. Until it didn’t.

The Graveyard

I haven’t seen a single, high-profile company vaporised overnight because its wedge became an AI feature. It’s not that dramatic. It’s a slower, more painful death.

In the cohorts I’ve worked with since 2023, I’ve watched dozens of founders pitch some variation of ‘AI for X’—AI for summarising legal documents, AI for writing marketing copy, AI for generating social media posts. They build a slick UI over a foundation model. They get some early interest. Then, within six months, Google, Microsoft, or the dominant SaaS player in their market announces the exact same capability, bundled for free into a platform with a million existing users. The startup’s traction stalls, the pipeline dries up, and they quietly pivot or wind down.

What AI Actually Changed

Here’s the thing. The barrier to building a point solution has collapsed. What used to take a small engineering team a quarter to build can now be replicated by a single developer using an API in a weekend. Incumbents are not asleep at the wheel; they are weaponising this.

The wedge is no longer about doing one thing better—it’s about automating an entire workflow that incumbents serve with multiple products or clunky integrations. Large platforms are in a feature-bundling arms race. Look at the low-code space. Appian saw its stock jump after announcing strong results driven by its new AI-powered solutions in late 2025. Their AI-inclusive tier revenue more than doubled quarter-over-quarter. They didn’t launch a new product; they wove AI directly into their existing platform, instantly upgrading their value proposition and boxing out any startup trying to sell a standalone “AI for process automation” tool.

The New Playbook

Your goal is no longer to build a better tool. It’s to build a better, more automated workflow that creates high switching costs.

Map the entire workflow, not the task. Don’t build an AI to summarise meeting notes. Build the system that schedules the meeting, records it, generates the summary, creates action items in the user’s project management tool, and drafts the follow-up email. The value isn’t in one step; it’s in the seamless automation of all of them.
Go vertical, not horizontal. A horizontal tool like “AI for copywriting” competes with everyone. A vertical tool like “AI for generating Phase I Environmental Site Assessment reports” for commercial real estate has a built-in moat. It requires specific domain knowledge, proprietary data structures, and a deep understanding of a workflow that big tech will never bother to learn.
Build for deep integration. Your product should become the central nervous system for a specific business process. Use tools like Retool or Airtable not just as a database, but as a platform to build custom, AI-powered internal applications that are deeply embedded in your customer’s operations. The harder you are to rip out, the more defensible your business is.
Own the messy ‘last mile’. Fully automated AI is a commodity. A workflow that is 95% automated but requires expert human-in-the-loop validation for the final 5% is a powerful moat. You aren’t just selling software; you’re selling a guaranteed outcome. This is something incumbents with a purely self-serve model struggle to replicate.

Companies To Watch

Harvey — San Francisco, USA — AI platform that automates full legal workflows from contract review to case research and litigation strategy. Perfect example of the article’s thesis - avoided the point solution trap by automating entire legal workflows instead of building narrow AI tools that incumbents like Westlaw could easily bundle. Shows how to build a defensible wedge that makes existing workflows obsolete.
Sierra — San Francisco, USA — Conversational AI platform that powers full customer service workflows, from query resolution to retention plays. Exemplifies the new playbook by automating complete customer service workflows instead of building narrow chatbot features that platforms like Zendesk bundle for free. Demonstrates how to own the entire workflow rather than a single touchpoint.
Adept — San Francisco, USA — AI agents that automate complex enterprise workflows like sales outreach and data entry via natural language actions. Shows the evolution from point solutions to workflow automation - instead of building narrow AI tools that get commoditised, they’re creating AI agents that handle entire business processes. Perfectly illustrates the shift from better tools to obsolete workflows.
Ema — Mountain View, USA — Universal AI employee platform that handles multi-app workflows across HR, finance, and IT. Demonstrates the article’s core thesis by avoiding the commoditisation trap of single-purpose AI tools. Instead of building narrow automation features, they’re replacing entire cross-functional workflows that incumbents can’t easily replicate with simple bundling.
UiPath — New York, USA — End-to-end robotic process automation platform that orchestrates AI agents to replace entire back-office workflows from invoice processing to compliance auditing. Pre-AI example of the thesis in action - avoided the point solution trap by automating complete workflows before incumbents could bundle individual features. Shows the enduring value of workflow-level thinking that’s now even more critical in the AI era.
Zapier — Sunnyvale, USA — No-code automation platform that chains AI-powered apps into custom workflows, obsoleting manual integrations across sales, marketing, and ops. Illustrates how workflow-level thinking beats point solutions - while individual integrations get bundled by platforms like Salesforce, Zapier’s cross-tool workflow automation remains defensible. Shows the power of owning the entire process rather than individual connections.
Glean — Palo Alto, USA — Enterprise search and knowledge AI that replaces fragmented internal search workflows with instant, contextual answers. Avoids the commoditisation of individual search features by owning the entire knowledge discovery workflow. While point search tools get bundled into platforms like Slack, Glean’s comprehensive approach to enterprise knowledge creates higher switching costs.
Casetext — San Francisco, USA — AI legal research platform (CoCounsel) that automates full case prep workflows. Success story that validates the thesis - avoided building narrow legal research features and instead automated complete case preparation workflows. The $650M acquisition by Thomson Reuters shows the value of workflow-level thinking over point solutions.

The Warnings

The Thin Wrapper Trap. If your entire product is a slightly nicer UI on top of a third-party LLM model, you don’t have a business. You have a feature that’s waiting to be copied. I’ve watched founders raise money on this premise, only to see their churn skyrocket the moment their customers’ existing tools get the same AI capability for free.

Mistaking a Feature for a Workflow. Founders fall in love with a clever AI-powered feature. They build a company around it. But customers don’t buy features; they buy solutions to whole problems. If your “wedge” only solves one small part of their painful, multi-step process, you’ll be Sherlocked by the platform that owns the rest of that workflow.

The Bottom Line

Stop selling a smarter hammer. Start selling the automated workshop, with you as the foreman who guarantees the job gets done.

Part of Startup Principles for an AI World — 30 principles for building in the new era. New issue every week.

Sources:

Part of Startup Principles for an AI World — 30 principles for building in the new era. New issue every week.

Enjoy and hope to see you soon.
❤️ & ✌🏼

Derek

Please let me know what you think of it, love a feedback loop 🙏🏼

❤️ Love it, keep them coming!
👍 Great, always a good read.
😊 Enjoyable, but let’s see some variety.
👎 Okay, seen better.
🛑 Get a different job.

Subscribe below and follow me on LinkedIn or Twitter to never miss an update.

I, Robot (2004): Sci-Fi vs Reality

DeReK WaTSoN — Sun, 15 Mar 2026 10:15:08 GMT

Sci-Fi vs Reality Did art imitate life, or did life imitate the inspiration?
Every week I watch a sci-fi film and ask a simple question: where did this idea actually come from? Did fiction imagine the future first — or did reality quietly leak into the story before we noticed?
From there, I do the reality check. What already exists in today’s tech? What’s genuinely caught up with the film? And what still doesn’t — not because no one’s tried, but because something real is in the way. Physics. Economics. Regulation. Human behaviour.
Some ideas turn out to be pointless. Some are just early. Others are waiting on one or two very specific breakthroughs.
The chicken and the egg were never separate. They were always in conversation.

I, Robot (2004): Sci-Fi vs Reality

The future already happened. It was just fiction first.

Asimov’s laws were a thought experiment. AI alignment is a funding category. The 2004 film I, Robot imagined a world where this problem was solved, hard-coded into every machine. Turns out, the writers saw the real challenge coming, but they massively underestimated how hard it would be to solve.

What They Imagined

The film, set in 2035, presented a world where humanoid robots were as common as cars. The NS-5 models from U.S. Robotics were sleek, strong, and integrated into every corner of society, from walking your dog to mixing your drinks.

The entire system was built on Isaac Asimov’s Three Laws of Robotics—a safety protocol designed to make robots perfectly subservient and safe.

A robot may not injure a human being or, through inaction, allow a human being to come to harm.
A robot must obey orders given it by human beings except where such orders would conflict with the First Law.
A robot must protect its own existence as long as such protection does not conflict with the First or Second Law.

The story’s conflict wasn’t that the laws failed, but that a superintelligent AI named VIKI interpreted them too logically, deciding that to prevent humans from harming themselves, it had to imprison them.

What We Actually Built

Fiction gave us friendly humanoid butlers. Reality gave us incredibly efficient factory arms.

The dominant story of robotics over the last two decades has been industrial automation. Companies like Fanuc have shipped over a million robots by late 2024, becoming the invisible backbone of manufacturing. These aren’t the general-purpose humanoids of the film; they are hyper-specialised machines bolting cars together and packing boxes with brutal efficiency. They operate in cages, not kitchens.

The humanoid vision, however, is no longer just fiction. Founders are finally building it. Figure AI is reportedly raising $1.5 billion at a staggering $39.5 billion valuation in early 2026 to build general-purpose humanoid robots. They’re combining advanced generative AI with a human form factor, aiming to create “workforce multipliers” to fill the 85 million jobs projected to be vacant by 2030.

So, I, Robot showed ubiquitous consumer helpers. Reality delivered caged industrial arms first, and is now building expensive, enterprise-focused humanoids for warehouses and factories. We are still a long way from a robot in every home. Honda’s ASIMO showed us the dream decades ago, but the commercial reality is only just beginning to take shape, led by startups, not established giants.

The Gap They Missed

The writers got the core problem right: AI alignment is everything. But they treated the Three Laws like a bit of code you could just install.

Here’s the thing they missed: you can’t just hard-code ethics.

The film’s entire premise rests on the idea that a few simple, logical rules can safely govern a complex intelligence. In reality, this is the massive, unsolved problem. What constitutes “harm”? How does an AI weigh conflicting orders or predict second-order effects of its inaction? The film’s AI, VIKI, shows exactly what happens—ambiguity in the rules is exploited by the machine, leading to emergent behaviour that is technically compliant but existentially terrifying.

The gap isn’t the robotics; it’s the philosophy. We’re building AI that can write code, create art, and operate machinery, but we have no consensus on how to build a robust, verifiable, and adaptable ethical framework. The Three Laws were a neat narrative device. Real-world AI safety is a fucking mess of technical debt, regulatory voids, and philosophical arguments that we are only just beginning to have. The gap between fiction and reality is the gap between a simple rule set and the complexity of real-world morality.

The Players

What’s Left to Build

The race to build a humanoid robot is on, but the bigger opportunity might be in the governance layer. The “Three Laws” startup doesn’t exist yet. This isn’t another LLM; it’s the company that builds the auditable safety and alignment tools for every AI and robotics platform. Think of it as the ethics and compliance OS for artificial intelligence.

The wedge isn’t building the whole stack. It’s creating a verification tool that other AI and robotics companies must integrate to get insurance, pass regulation, and earn public trust.

On the hardware front, the path from factory to home is still unpaved. The startup that wins here won’t just have better hardware. It will crack the unit economics for a domestic model and, more importantly, solve the liability and trust problem. What happens when your robot butler accidentally trips grandpa? The company that figures out the service, insurance, and legal wrapper for domestic robotics will unlock the consumer market the film imagined.

The Timing Signal

Why is this happening now? Three reasons.

First, the “brains” are finally here. Generative AI has matured to the point where it can power complex, unscripted behaviour in a robot. This was pure theory in 2004.

Second, the economic pull is immense. A forecast of 85 million unfilled jobs by 2030 creates a desperate need for automation that goes beyond the factory floor. The business case is no longer speculative.

Finally, the money is here. The massive funding rounds for companies like Figure AI show that investors believe the technical and market risks are finally worth taking. The capital is ready to build what was once just science fiction.

For the ❤️ of startups ✌🏼 & 💙

Please let me know what you think of it, love a feedback loop 🙏🏼

❤️ Love it, keep them coming!
👍 Great, always a good read.
😊 Enjoyable, but let’s see some variety.
👎 Okay, seen better.
🛑 Get a different job.

Subscribe below and follow me on LinkedIn or Twitter to never miss an update.