From the ashes of 2017 to the fluidity of DeFi, I've learned to watch the money, not the hype. But last week, when Bloomberg's data team circulated a chart—AI startups had raised $48 billion in Q2 2025, with 72% flowing to a closed loop of cloud providers and GPU lessors—I felt a familiar chill. The same chill I felt in 2017 when ICO whitepapers promised revolution but delivered only token velocity. The same chill from the summer of 2022 when Terra's bloom turned to ash. That chart was a confession: AI's boom is not built on real demand; it's built on circular financing. And if you hold any crypto asset tied to compute—Render, Akash, even certain L1s renting GPU time—you are holding the downstream risk of a debt-fueled bubble that has yet to pop.
Context: History Rhymes in Telecom and Tokens
In the late 1990s, telecom companies borrowed billions to lay fiber optic cables across the globe. The story was irresistible: the internet would need infinite bandwidth. The debt was secured against future revenues that never materialized because the actual users—households, small businesses—didn't need that much speed yet. By 2002, 85% of the fiber laid was 'dark'—unused. The companies went bankrupt, and the investors holding the debt (and the infrastructure tokens of that era, like the bonds of Global Crossing) were wiped out.
Today's AI boom is structurally identical. OpenAI, Anthropic, and a dozen other labs raise massive rounds from venture capital, then immediately spend that capital on GPU cloud services from CoreWeave, Lambda, or AWS. Those cloud providers, in turn, use the revenue to finance more GPU purchases from NVIDIA, whose valuation triples. The money circulates within a closed ecosystem: startup → cloud → chipmaker → startup (via VC recycling). No external end-user is paying for the inference or training. The only 'revenue' is the next round of financing. This is not a market; it's a capital churn.
I saw this pattern before. In 2017, I ditched my PhD thesis on lattice-based cryptography to track ICO whitepapers. I analyzed 500 projects and found that those with the strongest 'community narratives' outperformed technically superior ones by 300%. But the narratives were backed by no real usage—just token speculation. When the music stopped, the floor price of most protocols collapsed to zero. Crypto infrastructure projects today that depend on AI demand are repeating that mistake, but with hardware leverage.
Core: The Narrative Mechanism and Sentiment Analysis
Let me walk you through the contagion path, based on my own on-chain forensics from the 2022 crash. The narrative of 'AI is the next big thing' has overflowed into crypto assets that claim to provide decentralized compute—Render Network (RNDR), Akash Network (AKT), and even some GPU-based mining operations that shifted from Ethereum to AI inference. The sentiment chart on these assets shows a classic 'hype slope': parabolic rises since early 2024, followed by a consolidation that looks eerily like the top of the 2021 NFT spike. The social volume for 'GPU DePIN' hit an all-time high in May 2025, but the actual utilization of these networks—checked via on-chain contract interactions—has been flat or declining for three months.
Here's the hard truth: the circular financing model only works as long as the capital taps stay open. If any major VC signals a pullback—say, Sequoia or Andreessen Horowitz cautions its LPs about AI valuations—the entire loop seizes. The cloud providers lose revenue, start selling their GPU assets on the secondary market, and the price of compute drops. Tokens pegged to compute value (like RNDR, whose burn mechanism relies on job demand) lose their narrative anchor. In a bear market, tokens without income die first.
I've run a sensitivity analysis using a simple stock-to-flow model for compute demand. If VC funding for AI startups drops by just 30%—a mild correction—the implied GPU demand from crypto infrastructure projects could fall by 50-60%, because most of their current 'jobs' are actually test runs or pilot programs paid for by the same VCs. The market hasn't priced this yet. Current RNDR trading at $8.50 assumes continuous growth in rendering jobs. If growth flatlines, the fair value drops to $2.00—and that's optimistic.
Contrarian: The Blind Spot Everyone Misses
Here's the contrarian twist that most analysts overlook: not all crypto infrastructure is equally vulnerable. Bitcoin mining, for instance, uses ASICs—not GPUs—so it's insulated from the AI GPU glut. And some L2s and data availability layers (like Celestia) don't depend on AI workloads at all. The blind spot is that investors are treating all 'infrastructure' as a monolith. They're not.
Furthermore, the circular financing model, while dangerous, could actually accelerate the development of truly decentralized compute if the bubble deflates gradually. Why? Because when centralized cloud providers (CoreWeave, AWS) get squeezed, they'll need to offload GPU capacity to alternative markets. That could mean lower prices for decentralized networks like Akash, allowing them to attract organic users who were previously priced out. In the 2000 telecom crash, the dark fiber that survived eventually became the backbone of the internet—but only after the debt was cleared. The survivors in crypto infrastructure will be those with the lowest operational costs and the most diversified demand (not just AI inference but also gaming, rendering, and privacy computing).
But here's the catch: the timing is everything. In a liquidity crisis, all assets fall together. Even the 'good' infrastructure tokens will drop 80% before rebounding. You need a time horizon of 18-24 months and a stomach for volatility.
Takeaway: What the Next Narrative Looks Like
From the ashes of 2017 to the fluidity of DeFi, I've developed a rule: when the Bloomberg chart becomes the narrative, sell the story. The circular financing in AI will break—maybe not today, but within the next two quarters. When it does, crypto infrastructure tokens tied to GPU compute will be the canary in the coal mine. The next narrative will be about 'real yields' from protocols that generate revenue from actual users, not from VC recycling. I'm already seeing early signals in lending protocols that accept real-world assets—that's where the attention moves next.
Ask yourself: is the project you're holding earning dollars from strangers, or is it earning tokens from other tokens in a closed loop? If it's the latter, you're not an investor—you're a lender to the telecom bubble of the 2020s.