The Great AI Infrastructure Build-Out Bubble
Part 0 of 5: What You’re About to Read
🎯 The Thesis in Three Sentences
AI infrastructure faces continuous margin collapse as GPUs obsolete every 2-4 years (not 20-50 years like railways or fiber), fundamentally changing who wins and who loses.
Infrastructure players—NVIDIA, cloud providers, CoreWeave—see this coming, which is why they’re desperately pivoting to applications and software even as they build more hardware.
This creates a golden opportunity for application builders: you benefit from continuously falling costs while infrastructure investors face margin compression. The question isn’t whether to build—it’s how to position correctly.
⚡ The Core Insight: Infrastructure Margins Collapse, Applications Capture Value
Everyone wants this to be like the railway boom of the 1840s. Infrastructure gets built ahead of demand, investors lose money in a spectacular crash, but society benefits from the tracks left behind. The railways enabled the Industrial Revolution. The fiber optic cables laid during the dot-com boom enabled Web 2.0.
The comparison is instructive, but not in the way people think.
A railway line built in 1850 was still carrying trains in 1950 (100-year useful life). Railway companies faced one buildout, then decades of stable operations. Margins were sustainable.
A fiber optic cable laid in 2000 still carries internet traffic today (25-year useful life). Telecom companies laid cable once, then operated it for decades. Infrastructure held value.
But a GPU cluster built in 2023 with H100 chips? Already being replaced by Blackwell in 2025 (2-year cycle for cutting-edge work). Retains maybe 30-50% of value for inference and secondary workloads, but the training premium vanishes. Infrastructure companies must constantly rebuild on depreciating assets. Margins compress continuously.
This changes who wins.
Infrastructure investors face permanent margin erosion. Every GPU generation obsoletes the previous one. Capital expenditure never stops, but pricing power continuously erodes as newer, better hardware floods the market.
Application builders capture permanent margin expansion. Your costs fall continuously. Model API prices that cost $10 per million tokens today will cost $1 in three years and $0.10 in six years. Your margins expand automatically, year after year, with zero effort.
And here’s the part that makes this series different: infrastructure players know this. That’s why they’re moving to your layer.
🔄 Why Infrastructure Players Are Pivoting to Applications
Pay attention to what NVIDIA is actually doing.
They’re not just selling GPUs. They’re investing in 50+ AI startups. Building software platforms. Moving up the stack toward applications and distribution. The world’s most successful infrastructure company is pivoting away from pure infrastructure plays.
OpenAI isn’t satisfied being an API provider. They built ChatGPT—a consumer application with 100 million users. They’re racing to own distribution, not just sell compute.
Microsoft isn’t content providing Azure infrastructure. They bundled Copilot into Office, embedding AI directly into applications.
Why? Because they see two trends:
First, continuous margin compression (happening now): Infrastructure margins compress as hardware obsoletes every 2-4 years. This is structural. It never stops.
Second, potential acute crisis (2027-2029): Debt maturities cluster. Overleveraged players face refinancing risk. GPU utilization might drop. This could trigger rapid correction.
For infrastructure investors, both trends are terrible. For application builders, both trends are excellent.
🗺️ What Each Part Actually Covers
Part 1: The Trillion-Dollar Loop—How It’s Financed
We examine three financing mechanisms:
Strategic ecosystem investment (Microsoft-OpenAI): Rational coordination
Vendor financing (NVIDIA investing in customers): Medium risk, historically questionable
Asset-backed lending (CoreWeave’s $14.5B GPU loans): High stranded asset risk
Key insight: These financing structures aren’t inherently problematic (they’re how infrastructure always gets built), but they create correlated exposure. When combined with fast depreciation, this matters.
Key variable: Debt-to-GPU-value ratios. If >2:1, overleveraged.
Part 2: The New Compute Cold War—Geopolitics
We analyze sovereign AI strategies:
US export controls: Porous and failing (Chinese firms getting 1M+ chips via gray markets)
China’s self-reliance: Rational long-term, but short-term inefficiency
Middle East: Buying GPUs ≠ AI capability (no talent, no ecosystem)
Europe’s caution: Actually smart (avoiding stranded assets)
Key insight: Hardware doesn’t equal capability. Talent and institutions matter more.
Key variable: GPU gray market prices in Singapore, Malaysia.
Part 3: Financial Forensics—Who Survives?
We rank tech giants by risk-adjusted positioning:
Tier 1 (Smart Money):
NVIDIA (9/10): Already pivoting to software/applications via 50+ startup investments
Apple (8.5/10): Not playing = maximum optionality
Tier 2 (Rational Bets):
Microsoft (7.5/10): Generating real AI revenue, but locked into OpenAI
Amazon (7.5/10): Platform play, latecomer risk
Tier 3 (High Risk):
Google (6.5/10): $85B defending search
Meta (6/10): High spending, zero direct revenue
Tier 4 (Moonshots):
OpenAI (4/10): $450B bet on AGI, needs 9-18× revenue growth
Oracle (3/10): $100B in debt, 85% probability of failure
Key insight: Even NVIDIA (the infrastructure winner) is diversifying into applications. That’s validation.
Key variable: OpenAI revenue. Needs $20B+/year by 2026.
Part 4: This Isn’t Railways—Why Depreciation Changes Everything
We analyze why fast depreciation destroys the historical parallel:
Good bubble indicators: 0/3 (infrastructure doesn’t last)
Bad bubble indicators: 5/5 for infrastructure investors
But here’s the twist: The same dynamics that make infrastructure a bad investment make applications an excellent opportunity.
Five scenarios:
AGI by 2028 (15%): You benefit if positioned correctly
Incremental progress (50%): Application margins expand continuously
Capabilities plateau (25%): Domain expertise > AI hype
Efficiency gains 10× (10%): Your COGS drop 90%, margins explode
Key insight: Application builders win in 85%+ of scenarios.
Key variable: GPU resale values, inference API pricing trends.
Part 5: The Application Layer Playbook—Why You Should Build
This is where the series pivots. Parts 1-4 document why infrastructure is problematic. Part 5 explains why that’s good news if you’re building applications.
For application builders:
Continuous margin expansion as costs fall 10× over 5 years
If acute crisis hits (2027-2029), even better: costs crater suddenly, acquire distressed competitors, hire talent at discounts, competition clears out
Position correctly: get to revenue in 18 months, stay flexible on infrastructure, build on commodity layer
For infrastructure investors:
Continuous margin compression (structural)
Potential acute crisis (debt + overcapacity)
Most should sit out or be extremely selective
The strategy: Build applications on rented, commodity infrastructure. Capture value at the software layer where even infrastructure giants are desperately trying to pivot.
📊 The Consistent Probability Framework
Note: Application builders with strong unit economics win in 85%+ of scenarios.
🎯 Key Variables: Your Dashboard
Track these to know which scenario is unfolding:
1. Inference API pricing (continuous trend)
GPT-4-level performance dropped from $20/M tokens (2022) to $0.07 (2024)
If this continues, your margins expand 10× over 5 years
Watch: Model pricing announcements, new provider launches
2. Debt-to-GPU-value ratios (acute crisis indicator)
If >2:1, overleveraged and vulnerable
Watch: CoreWeave, Oracle debt levels; GPU resale values
3. GPU utilization rates (demand signal)
If <60%, overcapacity confirmed
Watch: Spot pricing, cloud provider discounts
4. NVIDIA’s application investments (where smart money moves)
Already 50+ deals in 2025
Watch: Which application categories they’re betting on
5. OpenAI revenue growth (AGI scenario tracker)
Needs $20B+/year by 2026 to justify infrastructure
Watch: ChatGPT subscribers, enterprise deals
👥 Who This Series Is For!
Application Builders (you should read this):
Should you build AI applications? Yes, with discipline
How to position for infrastructure trends? (Build on commodity layer, capture margin expansion)
What happens if correction hits? (Even better for you—costs crater, acquire competitors)
Infrastructure Investors (this explains your risk):
Should you invest in infrastructure? Mostly no, very selective
How to position? (Sit out or <10% to NVIDIA/Apple, prefer application layer)
What’s your risk? (Continuous margin compression + potential acute crisis)
Founders Building Tools/SaaS:
Where should you build? (Application layer, developer tools, vertical SaaS, knowledge platforms)
How fast must you move? (18-month revenue rule)
How to capture falling costs? (Model-agnostic architecture, managed services)
Policymakers:
How to prevent 2008-style contagion? (Monitor debt ratios, prepare restructuring)
Are export controls working? (No—failing, pivot to talent retention)
⚖️ The Synthesis: Two Different Conclusions
For infrastructure investors: Sit this out or be extremely selective. Continuous margin compression plus potential acute crisis creates terrible risk-reward. Even NVIDIA sees this, which is why they’re pivoting to applications.
For application builders: Build now with discipline. You benefit from the exact same trends that hurt infrastructure investors. Continuous cost deflation expands your margins. If acute crisis hits, you benefit even more.
The key insight: Infrastructure stress (continuous or acute) transfers value from hardware to software, from infrastructure to applications. Position accordingly.
⭐ What Makes This Analysis Different
Most AI infrastructure analysis asks: “Is this a bubble?” or “Will AGI happen?”
This series asks: “Given that infrastructure margins compress continuously AND there’s potential acute crisis risk, who benefits and who loses?”
The answer: Application builders benefit. Infrastructure investors lose (except NVIDIA, who’s already pivoting).
Parts 1-4 document the infrastructure stress. Part 5 explains why that’s your opportunity if you’re building applications.
Now let’s dive in...
This is Part 0 of a 5-part series on the Great AI Infrastructure Build-Out of 2025.
The full series:
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 🙏🏼
🛑 Get a different job.
Subscribe below and follow me on LinkedIn or Twitter to never miss an update.
For the ❤️ of startups
✌🏼 & 💙
Derek