This Is a Bad Bubble: Why AI Infrastructure Won't Age Like Railways

Part 4 of 5: The Great AI Infrastructure Build-Out of 2025

---

🧨 Is this a bubble?

The honest answer is: Yes. And not the good kind…

I know what you’re expecting. The standard tech-optimist take: “Sure, there’s froth, but the infrastructure will outlast the bubble—just like railways, utilities, and fiber optics. Investors lose, society wins, progress marches on.”
But here’s the uncomfortable truth: AI infrastructure is fundamentally different from past infrastructure booms. And that difference makes this a bad bubble, not a good one.

Let me show you why.

🧭 The Pattern We Want to Believe In

First, let’s acknowledge the historical pattern that everyone wants to apply:
Phase 1: The Build → New technology creates infrastructure opportunity
Phase 2: Financial Innovation → Novel financing accelerates deployment
Phase 3: The Overshoot → Optimism outpaces reality, capacity built ahead of demand
Phase 4: The Correction → Prices collapse, companies fail, investors lose money
Phase 5: The Prosperity → Infrastructure remains, enables new industries, society benefits
This pattern held for railways, utilities, and fibre optics.
But what if the infrastructure doesn’t remain? What if it becomes obsolete before it’s even fully utilised?

⏳ The Depreciation Problem: Why This Time Is Catastrophically Different

“A fibre optic cable laid in 2000 is still useful today. A GPU cluster built in 2023 may be obsolete by 2026.”
Let me be blunt: This completely destroys the historical parallel.

🧮 The Maths of Depreciation

Railways: 50–100+ years; upgrades don’t obsolete the tracks.
Fibre optics: 20–30+ years; improvements at endpoints.
GPUs: 2–4 years to obsolescence at the frontier; improvements at the hardware level.
Implication: Rail-like payback windows vanish. You’ve got ~3 years at the edge.

🪦 The Stranded Asset Scenario
Build $10B in 2024 → peak year, then rapid price compression as new silicon lands.
By years 4–10, you’re down to 30–40% of peak revenue; resale maybe 30–50%.
Result: You might recoup $6–7B, not $10B.

⚖️ The Training vs Inference Distinction Matters
Obsolete for training ≠ worthless for inference.
A100s (2020) still run profitably in 2025 (inference, fine-tuning).
AWS uses ~5-year depreciation—economic life outlasts frontier life.
Revised verdict: Bad bubble—but with 30–50% value retention over time.

🚫 Why the “Good Bubble” Thesis Doesn’t Hold

Good-bubble indicators:

1. Lasting value ❌

2. Enables new industries ⚠️

3. Society benefits ⚠️
   Bad-bubble indicators:

4. Massive debt ✅

5. Circular financing risk ✅

6. Unclear business models ✅

7. Stranded asset risk ✅

8. Leverage amplifies downside ✅
   Score: 5/5 bad, 0/3 good → Bad bubble.

🌍 The Geographic Stranding Risk: Building in the Wrong Places

Saudi/UAE/China are building capacity where demand is uncertain.
Rail/fibre followed demand; AI DCs follow geopolitics.
Risk: Underutilised assets, geographically stranded.

🧠 The AGI Counterfactual: What If I’m Wrong?

If AGI lands by 2027–28, current infra was under-built.
Economic value becomes trillions; today’s spend looks prescient.

📊 Why These Probabilities?

AGI by 2028 (15%) — diminishing returns + timeline slippage + revealed hedging.
Incremental progress (50%) — modal outcome.
Capabilities plateau (25%) — hard limits show up.
10× efficiency gains (10%) — compute needs fall sharply.

🔎 When Does the Correction Happen? Actual Leading Indicators

1. Debt maturities (2027–2029) — refinancing stress.

2. GPU utilisation — sub-50% → price collapse.

3. Capability vs capex — <10% capability growth with 50%+ capex = break.

4. Enterprise adoption — <30% F500 at scale by 2027 = weak demand.

5. Algorithmic efficiency — fast gains strand frontier GPUs.

   

🗓️ The Correction Timeline: Best Estimate

2025–26: Peak euphoria.
2027: Trigger (disappointing models, weak adoption, efficiency jump, debt stress).
2028–29: Correction (defaults, fire sales, consolidation).
2030+: Aftermath (power/cooling/space survive; GPUs commoditised).

🧱 Distinguishing Infrastructure Types: What Lasts?

Likely durable: Power & cooling, buildings/land, backbone networks.
Likely stranded: Specific GPU generations, bespoke chips with weak ecosystems, training-optimised clusters misaligned to new architectures.

🪞 The Reflexivity Problem: What Happens When Everyone Knows?

Less true belief → faster puncture. Volatility rises; exits crowd.

🗺️ The Scenario Framework

• AGI by 2028 (15%) — Winners: OpenAI, Oracle, NVIDIA; Losers: under-investors.

• Incremental progress (50%) — Winners: NVIDIA, Microsoft, revenue-backed apps; Losers: infra moonshots.

• Capabilities plateau (25%) — Winners: Apple, profitable incumbents; Losers: big spenders.

• Efficiency +10× (10%) — Winners: efficiency tools, open source; Losers: GPU hoarders.

📉 The Uncomfortable Conclusion: Follow the Evidence

Bad-bubble indicators dominate; depreciation is rapid; geography misallocates; reflexivity is fragile.
Strategy: Avoid infra unless you can absorb 30–50% losses. Back applications, workflows, data moats. Wait for 2027–29 resets.

💵 Key Variable to Watch

GPU resale values — H100 down >50% on secondary markets = overcapacity confirmed.

✅ Key Takeaways

1. Fast depreciation kills the railway parallel.

2. Bad bubble, not catastrophic — 30–50% residual value.

3. Training vs inference matters — older GPUs can earn.

4. Geography strands capital — politics ≠ demand.

5. AGI low-probability tail (10–15%) — not enough to justify spend.

6. Likely correction 2027–29 — watch debt, utilisation, capability, adoption, efficiency.

7. Not all infra equal — power/cooling/space > specific GPUs.

8. Reflexivity broken — faster deflation risk.

9. Invest after the flush — buy the survivors; fund durable moats.

🧩 Epilogue: Three Ways I Might Be Wrong

1. Inference tsunami — Economic life >> frontier life.

2. AGI odds understated — revealed preferences imply higher probability.

3. Geopolitical premium — “stranded” assets gain strategic value.

📐 Bottom Line: Confidence Intervals

Base case bad bubble (≈60% confidence).
But 40–50% combined chance I’m materially wrong via inference surge, AGI arrival, or geopolitics.
Next: scenario playbooks for each future.

📬 Coming Up

Part 5: The Application Layer Playbook (Why You Should Build)

---

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. 

---

For the ❤️ of startups

✌🏼 & 💙
Derek

---