Brian Armstrong, CEO of Coinbase, told a conference audience last week that he is personally short on leading AI companies. Nikhil Kamath, founder of India's largest brokerage, openly stated there is 'no reason' to pay current valuation multiples for private AI firms. Two billionaires, one from crypto infrastructure, one from financial markets, drawing the same conclusion: the AI investment thesis is structurally fractured.
This is not a fringe opinion from anonymous bears. These are operators who have navigated bubbles before. Armstrong lived through multiple crypto cycles. Kamath survived the Indian fintech boom. When they speak about valuation disconnects, the market should listen. But most retail and institutional capital is still pouring into names like OpenAI, Anthropic, and every startup with 'LLM' in its pitch deck. The disconnect between price and fundamental risk is widening — and I have seen this pattern before.
Context: The AI Arms Race and Its Structural Debt
The current AI boom resembles the DeFi Summer of 2020 more than any historical tech bubble. Back then, protocols like Compound and Aave issued governance tokens that were valued based on incentivized liquidity, not organic demand. I analyzed that in my risk reports — the TVL was artificially inflated by yield farming programs that would eventually dry up. Today, AI companies are valued on a similar premise: massive user growth driven by cheap or free access, but with unit economics that make no sense at scale.
Consider the numbers. Leading closed-source models cost tens of billions of dollars to train. Inference API pricing is set at a premium to recover those costs. But open-source alternatives — from Llama 3.1 to Mistral Large — are achieving comparable benchmark scores within six months of release, and their inference cost is up to 99% lower. The asymmetry is brutal: the incumbents spend billions to maintain a six-month lead, while the open-source community iterates for free. This is the same dynamic I flagged in my 2018 Parity Wallet autopsy: a missing modifier in the contract logic led to a $300 million loss. The flaw wasn't in the code itself — it was in the assumption that centralization equals security. Here, the flaw is assuming that massive training compute creates a permanent moat.
Core: Systematic Teardown of the AI Valuation Thesis
To understand why the AI bubble is fragile, we need to trace the liquidity sources. Where does the revenue come from? The top AI companies generate the majority from API calls — per-token pricing for inference. That market is highly price-elastic. Enterprises will switch to cheaper alternatives once performance thresholds are met. Open-source models now match GPT-4o on coding and reasoning benchmarks; the gap is closing at a rate of roughly three months per generation. At this pace, by mid-2027, open-source models will be on par with the best closed-source offerings. And when that happens, the pricing power evaporates.
I apply the same quantitative skepticism framework to AI that I used for crypto protocols. For a closed-source AI company, the key variables are: customer acquisition cost, cost per inference, churn rate, and R&D spend. Current data suggests that average revenue per user is declining as competition drives prices down. Meanwhile, R&D spend is accelerating — OpenAI is reportedly spending $7 billion annually just on training. That is a classic ponzinomic structure: you need ever-increasing revenue to service the rising cost base. But unlike crypto ponzis, there is no new token issuance to mask the gap.
Then there is the fragmentation risk. Kamath explicitly warned that countries will build their own AI models — 'localized tokens and energy.' This mirrors the trend I identified in the Layer2 space: instead of scalability, we got liquidity fragmentation across dozens of rollups. AI fragmentation will similarly break the global market assumption that underwrites current valuations. If India, the EU, and Japan each deploy their own open-source model, the total addressable market for a single global provider shrinks dramatically. The unit economics worsen, and the IPO window narrows.
Contrarian: What the Bulls Got Right
This is not a one-sided argument. The bullish case has real merits, and ignoring them would be intellectually dishonest. First, enterprise stickiness is real. Companies that fine-tune a model on proprietary data face switching costs. They cannot simply swap out the backend if the new model lacks their domain-specific knowledge. This creates a moat — but one that is purely operational, not technical. Second, the data flywheel: closed-source providers collect user feedback and generate preference data that can improve the model. Open-source projects lack that loop. Third, the possibility of a paradigm shift — a new architecture (e.g., reasoning-time compute) that is vastly more capable and requires proprietary hardware to train. If that happens, the open-source community might be left behind.
All three points are valid in theory. But I have seen how these arguments played out in crypto. 'Ethereum has network effects' — yet Solana captured market share through lower costs. 'Bitcoin has first-mover advantage' — yet it fails as a payments network. The caveat is that in both cases, the incumbent survived but the valuation multiples came down. The same will happen in AI. The closed-source giants will not go to zero, but the current pricing implies they will capture a disproportionate share of a global market. That assumption is the dangerous variable.
Takeaway: The Correction Timeline and What to Watch
Five years is a rough estimate for the bubble to fully deflate — Armstrong himself used that horizon. But I would argue the first cracks will appear within 18 months. Two key signals: first, the next generation of open-source models (Llama 4, Gemini Nano) should achieve near-parity on complex reasoning benchmarks by Q3 2027. Second, enterprise purchasing behavior will change as CFOs see the 10x cost differential. When a Fortune 500 company moves its internal chatbot from OpenAI to a self-hosted Llama instance, the market will reprice.
Precision is the only antidote to chaos. The math is clear: the cost of AI inference is heading toward zero, while the value of training compute is plateauing. Investors betting on perpetual premium pricing are ignoring the immutable laws of open-source economics. I dissected this same pattern in the Terra/Luna collapse — the algorithmic peg was mathematically unsustainable, but everyone assumed 'this time is different.' It was not. Logic survives the crash; emotion dissolves.
The AI bubble will not burst tomorrow. But the seed has already been planted. For those willing to look at the on-chain — or on-paper — data, the signals are unmistakable.