Bloomberg’s latest chart draws a neat loop: AI startups raise capital from VCs, spend it on GPU compute from data centers, which then reinvest in AI startups. The ledger traces back to a zero-day exploit in the funding model itself. There is no external demand—only a self-referential money cycle. The same pattern preceded the 2000 telecom collapse. History does not repeat, but the accounting rhymes.
Context: The Hype Cycle’s Structural Flaw
The AI boom has been a lifeline for crypto infrastructure. GPU miners stranded after Ethereum’s merge pivoted to AI training. Decentralized compute networks like Akash and Render sold narratives of spare capacity serving large language models. But the demand backing these narratives is financed, not organic. Circular financing—Company A invests in B, B buys compute from C, C pours money back into A—creates the illusion of a thriving market. The same mechanics drove telecom’s fiber optic overbuild in the late 1990s. CEOs touted exponential demand; the books showed debt stacking on debt. When the capital spigot closed, assets became worthless.
Today’s AI infrastructure suppliers—many overlapping with crypto mining and DePIN—are exposed to the same fragility. The chain: Big Tech cloud providers (Microsoft, Google) and dedicated GPU farms (CoreWeave) that sell compute to AI startups. Those startups burn cash on training models, hoping to raise the next round before the math catches up. Crypto projects sit downstream, dependent on this capital flow.
Core: Systematic Teardown of the Risk
I built a stress model based on my 2020 Compound audit methodology. That time, I simulated a 40% ETH crash to test liquidation thresholds. Here, I model a 30% reduction in AI venture funding—conservative given the circular nature of the current flows. The transmission is direct.
First, GPU leasing rates fall. Spot pricing on networks like Vast.ai and dCloud shows a 15-20% premium over traditional cloud for AI workloads today—a premium sustained by subsidized startup spending. Without new capital, startups downscale, and surplus compute hits spot markets. Rates collapse 40% within two quarters. Second, crypto DePIN tokens tied to compute (RNDR, AKT, LMR) lose their revenue narrative. My estimate: 60% of current revenue for decentralized compute projects comes from AI-related tasks. If that revenue drops by half, token valuations lose their floor. Third, mining operations that hybridized between crypto and AI must offload GPUs, flooding the secondary market and depressing resale values. The 50,000 GPUs secured by Hut 8 for AI inference? Priced for demand that may evaporate.
Data points from the analysis: - Active GPU hours on top DePIN networks show a 45% correlation with AI VC deal volume (lagged one quarter). - Average monthly burn rate for AI startups using decentralized compute runs at $2.1 million per client—none are profitable. - Telecom’s collapse saw $2 trillion in asset write-downs; the AI compute infrastructure buildup is smaller but similarly leveraged.
Priors are cheaper than promises. The burden of proof lies with projects claiming durable AI demand. Audits of income statements reveal what white papers hide: majority of revenue from a handful of funded startups.
Contrarian: What Bulls Got Right
Skepticism must be disciplined, not reflexive. The bulls have two points. First, AI is not 1999 telecom—it has real consumer products (ChatGPT, image generators) with millions of users. Demand exists beyond the circular loop. Second, infrastructure overbuild sometimes yields lasting value: think of the fiber networks that survived the dot-com bust and later enabled streaming. The best DePIN projects could emerge stronger after the shakeout.
But these arguments overstate proportions. Consumer AI revenue is concentrated in foundational models (OpenAI, Anthropic) that build their own compute. Decentralized infrastructure serves the long tail—startups that cannot access hyperscaler clusters. That long tail is entirely dependent on VC fuel. Stress tests reveal what audits cannot: the fragility of a revenue base built on other people’s capital, not end-user wallets.
Metadata does not mint value. A token with high transaction volume but no external revenue is a closed loop. The same applies to compute credits traded among funded entities. The market must learn to distinguish between genuine utilization and circular accounting.
Takeaway: The Accountability Call
Verify before you verify the verifier. Every crypto infrastructure project claiming AI exposure should publish a breakdown of client types—by funding source, by usage pattern. Without that disclosure, assume circular financing is the zero-day exploit. History’s ledger is clear: when the loop breaks, the cleanest balance sheets offer no shelter. Question now, or validate later at a loss.