Date: January 20, 2026
I. The Thesis: Infrastructure vs. Vaporware
We are not in a Dot-Com Bubble (2000); we are in a Telecom/Fiber Bubble (1996). The distinction is critical to understanding the current crash.
The 2000 Dot-Com Bubble (Valuation Fraud)
Driven by companies with zero revenue and broken unit economics (e.g., Pets.com). The assets were vapor. When the bubble popped, the companies evaporated, and nothing was left but unused office chairs.
The 1996 Telecom Bubble (CapEx Overbuild)
Driven by massive capital expenditure to lay fiber optic cables for an internet traffic boom that eventually arrived—but not fast enough to save the builders.
- The Reality: Companies like Global Crossing and WorldCom spent billions digging trenches. They went bankrupt because they built capacity for 2010 demand in 1999.
- The Legacy: Investors lost everything, but the fiber remained. That cheap, distressed infrastructure eventually allowed the “Internet 2.0” (YouTube, Netflix) to emerge a decade later.
The 2026 Reality: The “Field of Dreams” Risk
We are overbuilding GPU clusters exactly like we overbuilt fiber.
- CapEx Overload: Hyperscalers (Microsoft, Google, Meta) are projected to spend over $500 billion in CapEx this year.
- The Overbuild Risk: The danger isn’t that the tech is fake; it is that the infrastructure is being built faster than the monetization layer can catch up.
- The Physical Reality Check: Paradoxically, the bubble is being saved from inflating even faster by physical constraints. We are running out of power (electricity) and grid capacity. If power were infinite, the oversupply of compute would be catastrophic by now.
II. The “Juicy” Risks: Circular Economics & Hardware Rot
Why this bubble is more dangerous than 1996.
1. The Financial Ouroboros (Circular Revenue)
The market is being propped up by “Cloud Round-Tripping.”
- The Mechanism: Hyperscalers (Google, Microsoft, Amazon) and Nvidia invest billions of cash into AI startups (e.g., Anthropic, CoreWeave, Mistral).
- The Catch: The investment often comes with a “cloud credit” requirement. The startup must spend that money back on the investor’s cloud/GPUs.
- The Illusion: Microsoft invests 1B for Compute → Microsoft reports this as “Cloud Revenue Growth” to Wall Street.
- The Risk: This isn’t organic market demand; it is accounting gymnastics. When the VC money stops, this circular revenue vanishes instantly.
2. The Depreciation Bomb (Hardware Rot)
The “Fiber” analogy from 1996 has one fatal flaw: Durability.
- 1996: Fiber optic cable lasts 25–30 years. Even if the company went bankrupt, the asset remained valuable for decades.
- 2026: Nvidia H100/Blackwell GPUs have a useful economic lifespan of 3–5 years before they become obsolete e-waste.
- The Threat: If this infrastructure isn’t monetized immediately, it doesn’t just sit there like dark fiber; it rots. Billions in CapEx are racing against a depreciation clock that is ticking much faster than revenue is growing.
III. The Wrapper Extinction Event: Getting “Sherlocked”
“You know those handcars, the little machines that people stand on and pump to move along on the train tracks? That’s [your startup]. Apple is the steam train that owns the tracks.”
— Steve Jobs (paraphrased) to developers regarding Sherlock
To understand why 90% of AI startups are dying in 2026, you have to look at Apple in 2002.
The History: Watson vs. Sherlock 3 (2002)
In the early 2000s, Karelia Software released Watson, a tool that “wrapped” the web (stocks, movies, flights) into a native Mac interface. It was brilliant.
- The Kill: In 2002, Apple released Sherlock 3 directly into the OS. It did exactly what Watson did, but for free. Watson’s value evaporated overnight.
The 2026 Parallel: The “Thin Wrapper”
- The Product: “Chat with your PDF,” “AI Code Assistant,” or “Marketing Copy Generator.”
- The Reality: These are just API calls to OpenAI with a nice UI.
- The Sherlocking:
- Startup feature: “Upload a PDF and ask questions.” → Sherlocked by: OpenAI Canvas / File Search.
- Startup feature: “AI Agent that can code.” → Sherlocked by: GitHub Copilot / Replit Core.
- The Verdict: If your startup’s “moat” is a prompt chain, you are standing on the tracks. The platforms are the steam train.
IV. The Business Model Crisis: The Death of “Per Seat”
The entire software economy (Salesforce, HubSpot) is built on charging $X per User / Month.
- The Paradox: The promise of AI is that one employee can do the work of ten.
- The Suicide Pact: If an enterprise successfully uses AI to reduce headcount by 30%, the software vendor’s revenue drops by 30%.
- The Shift: We are watching the painful death of the “Per Seat” model. Companies are scrambling to switch to “Consumption Pricing” (pay per outcome), but this crushes margins because running AI has high COGS (electricity/compute), unlike traditional software.
Summary & Strategy
- The Macro View: The infrastructure is ready for tomorrow, but the business models are stuck in today. We are likely building too many GPUs relative to current revenue models.
- The Prediction: A period of massive consolidation. Hardware prices will crash (good for builders, bad for investors) before the “real” AI economy matures.
- The Builder’s Rule: Assume the cost of intelligence falls to near-zero. Do not build wrappers. Build Vertical (Proprietary Data + Workflow).