Last month I tried to price out a small GPU cluster for a zero-knowledge proving service I'm prototyping โ eight H100s, nothing exotic, the kind of hardware that would have felt like a flex back in 2021. The quote came back at a number I'm genuinely embarrassed to type. The lead time was worse. So I did what any stubborn builder does: I opened a spreadsheet, multiplied hours by kilowatt-hours, added the rental premium, and watched the unit economics of my elegant cryptographic proof dissolve into a story about electricity and silicon. That same week, John Williams, president of the Federal Reserve Bank of New York, told an audience that the supply squeeze in AI would likely last beyond two years. Suddenly my little spreadsheet wasn't a personal inconvenience. It was a macroeconomic signal, and it was pointing directly at the industry I've spent a decade explaining to people.
For those who don't follow the Fed's cast of characters: Williams isn't a commentator. He's a permanent voting member of the FOMC, which means his words function less as opinion and more as choreography. When he said the U.S. central bank "may need one more rate hike this year," the phrasing matters more than the number. "One more" is not "more." It's the vocabulary a central bank uses when it wants to signal the end of a hiking cycle without committing to it โ a way of preserving flexibility while gently telling markets to stop pricing in cuts. He paired that with a second claim: that AI's demand for compute, chips, data centers, and power is itself generating inflationary pressure, and that this pressure will persist past a two-year horizon.
Read that again, because it inverts a decade of conventional wisdom. AI has always been sold to us as a deflationary force โ a productivity engine that crushes unit costs, automates labor, and bends the price curve downward. Williams is describing the near-term AI economy as a bottleneck: too much demand chasing too little fab capacity, too little grid capacity, too little transformer manufacturing. That reframing is the real story, and it's the one most crypto coverage missed.
For crypto specifically, "higher for longer" is not an abstraction. It sets the discount rate against which every token's future cash flows are valued, it raises the cost of the debt that funds mining farms and data centers, and it keeps the risk-free alternative โ Treasury yields โ competitive against yield-bearing crypto products. When Williams speaks, the entire risk curve shifts, and crypto sits at the far end of that curve.
Here's where this stops being macro abstraction and starts being an engineering problem โ and this is the part I care about, because I've watched crypto's compute economy get repriced three times now, and each repricing taught the same lesson.
Start with energy. Every serious AI workload โ training runs, inference at scale, and increasingly ZK proof generation โ is fundamentally an electricity arbitrage. When data centers bid for power in the same regions where Bitcoin miners and proof-of-stake validators operate, they don't compete on ideology. They compete on willingness to pay per megawatt-hour. I've been through this before. During the 2022 bear market, when my platform's user base collapsed 90%, I spent months in "Code & Coffee" sessions with developers who were being quietly pushed out of hosting contracts by larger buyers. The lesson wasn't that decentralization failed. It was that decentralization doesn't exempt you from the physics of supply and demand โ it just changes who gets to participate in the market.
Now layer the second-order effects. If Williams is right that AI supply tightness persists beyond two years, then the cost floor for anything that consumes compute stays elevated. That touches more of crypto than people admit. Rollups need proving hardware. ZK-verification of AI-generated content โ the entire thesis behind the "Verifiable Truth Initiative" I'm currently building with partners โ needs cheap proof generation, or it becomes a luxury good for well-funded media outlets. Oracle networks need low-latency, high-redundancy infrastructure, and they need it everywhere at once. Trust the process, but verify the code โ and also verify the power bill, because the power bill is now part of the security model.
The oracle layer deserves its own warning. I've argued for years that oracle feed latency is DeFi's real Achilles' heel, and the AI squeeze makes it worse, not better. Every price feed that depends on off-chain infrastructure depends on machines that now compete for the same scarce compute and bandwidth as AI inference. When the physical substrate gets expensive, the weakest redundancy gets cut first โ and in DeFi, the weakest redundancy is often the difference between a liquidated position and a catastrophic one. Chainlink didn't solve this by decentralizing node operators; it concentrated the real cost in a handful of well-capitalized ones. That's a supply chain wearing a network's clothes.
Rollups face a similar reckoning from the other direction. Post-Dencun, blob space felt infinite โ cheap data availability, gas fees collapsing to fractions of a cent. But blob space is provisioned by the same hardware supply chain that AI is currently exhausting. My working estimate is that blob demand saturates within two years, and when it does, the fee floor for rollup users doubles, then doubles again. The AI squeeze accelerates that timeline, because proving systems and data availability layers both bid for the same constrained silicon. Anyone pricing a rollup's economics on today's near-zero fees is modeling a world that is already ending.
Here's the piece I haven't seen anyone connect: the AI supply squeeze is a validation of DePIN-style compute networks and a stress test of them at the same time. On the optimistic side, if centralized hyperscalers can't absorb demand fast enough, decentralized GPU marketplaces get a genuine wedge โ real demand, not narrative demand. On the pessimistic side, those networks are only as good as their utilization guarantees, and I've audited enough of them to know that "idle GPUs on a token incentive" is not the same thing as "reliable compute." The gap between a whitepaper and a served job is where most of these projects die. I've watched it happen with Lightning routing โ seven years of promises, and the routing failure rates never got good enough to make it more than a niche. The same trap waits for decentralized compute: beautiful incentives, brutal last-mile reliability.
And the energy channel runs straight back into monetary policy. Higher-for-longer rates raise the cost of capital for every capital-intensive infrastructure play โ which is exactly what AI data centers and large-scale mining operations are. Williams is, in effect, describing a world where the Fed keeps the discount rate high partly because AI is inflating input costs, while that same high rate makes it more expensive to build the infrastructure that would relieve the bottleneck. That's a genuine policy paradox, and it's not lost on anyone building physical infrastructure in this space. The rate path and the compute path are now coupled, whether policymakers admit it or not.
There's a longer cycle underneath all this that the crypto industry keeps misreading. What Williams is describing โ surging demand for compute, power, and chips โ looks less like a spike and more like the early phase of a capital-formation cycle, the kind economists used to call a Juglar wave. Equipment investment begets capacity, capacity begets supply, supply eventually begets deflation. But crypto tends to mistake the inflation phase for the whole story. During the 2017 ICO boom, I watched hundreds of projects price their entire thesis on a demand curve that never materialized, because they forgot that supply catches up. The AI compute boom will follow the same arc โ expensive now, cheap eventually โ and the projects that survive will be the ones that build for the cheap phase while surviving the expensive one.
Then there's the verification problem, which is my actual obsession. If AI floods the world with synthetic content, the demand for cryptographic provenance โ signed attestations, ZK proofs of origin, on-chain content registries โ grows with it. But provenance is compute-hungry. So the same supply squeeze that makes AI expensive also makes proving what AI produced expensive. The irony is sharp: the technology we need to keep truth verifiable in the AI age is itself constrained by the AI boom's own appetite for silicon and power. Every price signal is a confession, and this one confesses that our verification infrastructure is not yet cheap enough to be universal.
Now the part that makes me skeptical, because optimism without a bug report is just marketing.
Williams's AI-supply narrative could be exactly what it looks like โ a genuine structural observation. But it could also be post-hoc rationalization. Central banks decide they want to hold rates high, then reach for a story that justifies it. AI is the perfect story in 2026: it's popular, it's visible, it's plausibly inflationary, and it lets a hawkish stance sound forward-looking rather than stubborn. I've seen this pattern in crypto a hundred times โ a token finds a narrative after the price moves, not before. So I'd treat "AI causes inflation" as a hypothesis under audit, not a settled fact. The falsification test is simple: if AI-related prices โ chip spot rates, data center rents, power contracts โ start falling while core inflation stays sticky, then Williams was rationalizing, not diagnosing.
The crypto industry has its own blind spot here. We love the "AI x crypto" thesis because it's narratively gorgeous. But the bottleneck isn't clever tokenomics โ it's fabs, transformers, and transmission lines. You cannot decentralize your way out of a transformer shortage. If the next bull cycle's story is "we'll solve AI's compute crisis with a token," I want to see the served-job metrics, the SLA data, and the uptime history before I believe a word of it. Hope is not a hedge.
Williams gave crypto a rare gift: a macro signal that doubles as a technical roadmap. The AI supply squeeze is real enough to reprice compute, energy, and verification for years. The winners won't be the projects with the best narrative โ they'll be the ones whose unit economics survive a high-rate, high-wattage world. Build for that world. Audit the power bill. And keep verifying, because the truth is about to get expensive to prove.
