Williams Says AI Supply Stays Tight for Two Years: The Crypto Compute Trade and the Rates Trap

On a Tuesday morning in lower Manhattan, John Williams said the quiet part loud. The New York Fed president told an audience that the supply squeeze from artificial intelligence will outlast the next two years, and that the Fed may need one more rate hike before this cycle closes. Two sentences. The market heard a novel. Within minutes, the front end of the Treasury curve twitched, growth equities sagged, and a specific basket of crypto tokens, the ones that sell GPU compute, bandwidth, and storage, went bid. I have watched this exact reflex before. It is the same reflex that fired in 2021 when the Fed first whispered 'transitory' and the market answered by levering into anything with a narrative. The difference now is that the narrative has a physical supply chain behind it. AI does not run on sentiment. It runs on silicon, copper, and megawatts. And those inputs are scarce. This is the anomaly worth trading: the Fed is simultaneously telling you money will stay expensive for longer and that the reason is a technology cycle that crypto claims to be building rails for. The rates channel says sell risk. The AI supply channel says buy compute. Most desks will pick one. The trade is in understanding why both can be true at once, and where the seam between them sits. Start with the Williams remarks read as a single unit, not two headlines. First: the supply side of the AI buildout is constrained for at least twenty-four more months. Second: policy is near, but not at, its terminal rate. Read together, they describe a regime where the cost of capital stays elevated precisely because a productive technology is bottlenecked, not because demand is running hot in the traditional sense. This is a meaningful reframing. For three decades, the consensus treated technology as a deflationary force. Moore's law pushed unit costs down. Software ate margin. Every productivity wave, from personal computers to the internet to mobile to cloud, arrived as a disinflationary gift to central banks. Williams is arguing that the near-term version of the AI wave inverts that. Training frontier models and serving inference at scale consumes electricity, advanced packaging, high-bandwidth memory, and copper at a rate the physical supply chain cannot currently match. The result is a bottleneck that behaves like inflation: prices for compute, power, and the equipment that delivers them rise faster than the general price level. For crypto, this matters along two distinct channels. The first is the discount-rate channel, which is mechanical and brutal. Crypto is the longest-duration asset class in existence. A token with no cash flow is a pure claim on a future state of the world, and its present value is exquisitely sensitive to the risk-free rate. When the front end of the curve reprices higher, the entire long-duration complex, unprofitable tech, venture-stage equity, and most of the altcoin market, takes a valuation haircut. This channel does not care about your thesis. It does not read your whitepaper. The second channel is the fundamental one, and it is where the trade lives. If compute is scarce for two years, then anything that produces, routes, or prices compute has a structural bid. Crypto has spent a decade building exactly this: decentralized physical infrastructure networks, or DePIN, meaning GPU marketplaces, storage networks, bandwidth protocols, and the token shells around them. Most are mediocre. A few are real. The Fed just told you the demand side of their profit and loss statement is underwritten for at least two more years. Now add the institutional layer. Options desks do not trade narratives; they trade the distribution of outcomes. And the distribution shifted with Williams's sentence. If the Fed is genuinely near the terminal rate but committed to holding it there, higher for longer with a supply-side justification, then the volatility surface should steepen at the front and stay elevated at the back. That is a specific, tradeable shape, and it is the shape I am going to price out below. Now the mechanics. I want to be precise about how a hawkish Fed with an AI supply story transmits into crypto order flow, because the naive read, rates up and crypto down, is wrong in the details, and the details are where the money is. Start with the discount-rate channel, quantified. In my options work I run a simple sensitivity. Take the crypto complex, bucket it by cash-flow distance, and measure its beta to the two-year Treasury yield. The pattern is stable. Bitcoin's beta to front-end yields is the lowest in the complex, because it carries a partial store-of-value bid that cushions it. Ethereum sits in the middle. The long tail, the layer-one tokens, the DePIN names, the AI-themed assets, carries the highest beta to rates, because their entire value is a terminal value discounted from far in the future. When Williams says one more hike, he is applying a haircut to that tail. That is why the reflex sell hits alts hardest and Bitcoin least. But here is the wrinkle, and it is the part the perma-bears miss. The AI supply squeeze is not symmetric across the crypto complex either. It is a tailwind for a narrow slice and a headwind for the rest. The slice is compute. If you produce GPU hours, you are selling into a market where the incumbent providers, the hyperscalers, cannot meet demand and are rationing. Your marginal revenue is rising while the Fed is trying to slow everything else. That is a rare combination: a pro-cyclical revenue line inside a counter-cyclical policy regime. I have seen this movie in the physical commodity complex. When a supply bottleneck is genuine, the spot price decouples from the futures curve, backwardation deepens, and the people who own the physical, not the paper, capture the rent. In crypto, the physical is the GPU, the rack, the megawatt. The token is the paper. The question every DePIN investor must answer is whether the token actually captures the rent from the physical asset, or whether it is just a claim on someone else's margin. Most token designs fail this test. They subsidize supply with emissions, the emissions dilute the very cash flow they claim to distribute, and the whole thing collapses into a scheme with a whitepaper. When the code bleeds, the ledger keeps the truth, and the ledger says most of these networks are selling compute below cost, funded by token inflation, which is a slow-motion insolvency dressed as a growth story. That is the filter. But the filter is not the same as a no. A minority of these networks run real GPUs, bill real customers, and settle in real stablecoins. For those, the Fed's two-year supply squeeze is a gift that keeps repricing. The trade is not buy DePIN. The trade is separate the rent-capturers from the rent-substituters, and the Fed just widened the spread between them. Now the leverage layer, which is where I live. Higher-for-longer hits crypto's borrowing costs directly, because crypto's borrowing costs are downstream of the dollar. Let me be specific about DeFi's rate mechanics, because this is where most traders are structurally blind. Aave and Compound do not discover an interest rate the way a bond market does. They set it with a formula, a piecewise-linear or kinked curve that maps utilization to a rate. Below a target utilization, the rate is gentle. Above it, the rate spikes to force repayment. That curve is a policy choice. It is calibrated by governance, not by supply and demand in any market sense. The market rate you pay to borrow USDC on-chain is an arbitrary function that a handful of delegates can move by voting to shift a kink. Read that against Williams. The Fed sets its policy rate with a committee, minutes, dot plots, and a communication apparatus. DeFi sets its rate with a governance vote and a curve someone chose years ago. Both are administered rates. Only one admits it. When the Fed goes higher-for-longer, the dollar leg of every crypto carry trade gets more expensive, and DeFi's administered curve has to adjust, usually badly, usually late, usually by ratcheting the kink until borrowers are force-liquidated. The result is that the leverage available to crypto traders is not a smooth function of the Fed's rate. It is a step function punctuated by liquidation cascades. That is the transmission mechanism, and it is why a single hike headline can trigger a twenty percent altcoin drawdown that looks disproportionate to the macro move. Then the volatility surface. Here I will be concrete about what a hawkish surprise does to crypto options, because this is the institutional bridge most retail readers never cross. Crypto has a term structure, implied vol across tenors, and a skew, the price of downside puts relative to upside calls. Before a hawkish Fed event, the market typically buys front-end protection. Short-dated implied vol richens, skew steepens toward puts, and the front of the curve inverts against the back. If the event is a genuine surprise, the market had priced a pause and got a hike, the front-end vol collapses after the print, but the back-end does not. The back-end stays bid because the higher-for-longer story is a persistent regime, not an event. That produces a specific shape: a term structure that stays elevated at the back even as the front normalizes. If you are long the back-end and short the front-end, a calendar structure, you are expressing the view that the regime, not the event, is what matters. That is the Williams trade in options form. I ran a version of this in 2024 with a Python harness pulling on-chain options data from Deribit, implied vol, realized vol, and the spread between them, refreshed on a loop. The edge was not in any single print. The edge was in the persistence of the spread between implied and realized during macro regimes. When the Fed enters a higher-for-longer communication phase, realized vol in crypto tends to overshoot implied on the day and undershoot it over the following weeks, because the market prices the event and then forgets the regime. That asymmetry, event overreaction and regime underreaction, is a structural arbitrage, and it is the closest thing I have found to a repeatable edge in macro-crypto. Arbitrage is just violence disguised as math. You are extracting rent from other participants' emotional mispricing, and you should never confuse the elegance of the equation with the fact that someone on the other side is losing money. Now the token-flow evidence, because this is where the black box of on-chain data opens. When a hawkish Fed headline hits, the first thing that moves is not price. It is funding rates on perpetual futures. Perps are the leverage instrument of crypto, and their funding rate is the real-time price of being long. In a healthy bull market, funding sits modestly positive, longs pay shorts a small carry. When a hawkish surprise lands, funding can flip negative within minutes as leveraged longs get flushed, then snap back positive as dip-buyers lever in. That oscillation is the signature of a market that cannot decide whether it is trading the rates channel or the AI channel. Here is the specific tell I watch. On a hawkish Fed day, if funding on the AI and compute token complex stays positive while funding on the broad altcoin complex flips negative, the market is telling you it believes the supply story. It is rotating out of generic beta and into the narrow fundamental slice. That divergence, positive funding in compute and negative funding in the tail, is the cleanest real-time confirmation that the AI channel is winning against the rates channel. If instead funding goes uniformly negative, the rates channel is dominant and the whole complex is a short. The signal is not the price. The signal is the dispersion of funding across sectors. Code does not lie, and neither does a funding rate, because it is a payment that must actually be made every eight hours. Layer in the stablecoin supply. Higher-for-longer raises the yield available on dollar instruments off-chain, which competes directly with on-chain stablecoin yield. If the Fed's terminal rate holds, the opportunity cost of holding USDC in a DeFi pool rises. That should, over time, drain stablecoin deposits from low-yield pools into either Treasuries or the few on-chain venues that can compete, which, in a higher-rate world, means the ones offering leveraged yield, which means the ones taking the most risk. This is a slow variable, but it is the one that determines whether crypto has dry powder to bid the AI trade. Watch net stablecoin supply as a leading indicator of the complex's capacity to absorb a hawkish shock. And the governance layer, because it is not decorative. When the Fed signals a regime, crypto's response is mediated by governance, the votes that set the rate curves, the emissions schedules, the treasury deployments. Here is the uncomfortable mechanics. Most token holders do not vote. They delegate. And delegation concentrates. A handful of delegates, often the same venture funds and influencers who hold the largest positions, end up controlling the parameters that determine everyone's borrowing costs and emissions. So when I say DeFi's rate is an administered rate, understand that the administrators are a small, self-selecting group whose incentives are not identical to the depositors'. The decentralized rate is as centralized as the committee Williams sits on, just with worse disclosure. That is a structural risk a hawkish Fed amplifies, because higher rates make the parameters matter more, and parameters mattering more makes the delegate concentration matter more. Now the contrarian turn, because everything above is consensus-adjacent and the money is in the blind spot. The blind spot is this. The market is treating Williams's AI supply narrative as a fundamental fact. I am treating it as a hypothesis with an agenda. Central bankers are not neutral observers of the economy. They are participants who must justify a policy path. Higher-for-longer is a hard sell if the only argument is sticky services inflation and a labor market that is already cooling. It is a much easier sell if you can point to a technology boom that is structurally inflationary for two more years. The AI supply story gives the Fed exactly what it needs: a reason to hold rates high that is not about demand, and therefore not about the Fed's own past mistakes. Read the causal arrow carefully. It is entirely possible that the sequence is not AI supply squeeze causes higher rates. It is possible the sequence is the Fed wants higher rates, and the AI supply squeeze is the post-hoc rationale. I have audited enough protocols to know the difference between a mechanism and a narrative that describes the mechanism. The mechanism has to close. The narrative just has to be plausible. And a two-year supply squeeze is conveniently unverifiable in real time, long enough to justify a policy stance, short enough to be quietly dropped if the data turn. The second blind spot is on the crypto side. The DePIN and AI-token complex has taken Williams's sentence as a demand guarantee and bid itself accordingly. But a supply squeeze for compute is a squeeze on inputs, and inputs have costs. If your token's business model is to buy GPUs at spot and resell compute, a supply squeeze raises your cost of goods as much as your revenue, and if your emissions are subsidizing the gap, the squeeze accelerates your insolvency rather than your growth. The market is pricing these tokens as if the squeeze is pure margin. For most of them, it is pure cost. The ones with pre-negotiated supply, owned hardware, or locked-in power contracts are the rent-capturers. The rest are rent-substituters, and higher-for-longer is a slow execution of the difference. The third blind spot is regulatory, and it connects to the first. When the Fed frames AI as a strategic supply bottleneck, it is also framing AI compute as a strategic asset. Strategic assets get regulated. Watch the direction of travel: export controls on advanced chips, licensing regimes for large-scale compute, disclosure requirements for AI capital expenditure. Projects that have spent years preaching decentralization while their foundation wallets and team allocations sit in plain sight on-chain are exactly the ones that will discover that decentralized is a compliance posture, not a legal shield. The governance wrapper that was supposed to provide deniability becomes the first thing a regulator pierces, because the wallets were always traceable and the community was always three multisig signers. The chain remembers. That is the whole point of it. So the contrarian position is not ignore the AI trade. It is do not let the Fed's narrative do your diligence. The supply squeeze may be real. The policy justification may be opportunistic. Both can be true, and the trade is in knowing which part you are actually being paid for. So here is what I am watching, with levels and triggers rather than opinions. First, the funding dispersion. If the compute complex holds positive perp funding through a hawkish Fed print while the broad tail flips negative, the AI channel is real and I want to be long the rent-capturers and short the rent-substituters, a market-neutral spread that isolates the supply story from the rates story. If funding goes uniformly negative, the rates channel dominates, and the trade is to short the tail outright and wait for the cascade. Second, the options term structure. A persistent bid at the back end, even as the front normalizes after the event, confirms that the market is pricing a regime rather than a print. That is the calendar structure I described, long back-end vol and short front-end vol, and it is the cleanest expression of the Williams thesis in an instrument that actually settles. Third, net stablecoin supply. If higher-for-longer drains stablecoins out of DeFi and into Treasuries, the complex loses the dry powder to bid its own AI narrative, and the whole trade unwinds regardless of how good the story is. Stablecoin supply is the oxygen. Watch it before you watch price. Fourth, the verification signal on the supply story itself. Watch the actual prices of the physical inputs, high-bandwidth memory, advanced packaging, power contracts. If those prices keep rising for two more years, Williams is right and the compute trade has a fundamental floor. If they roll over, the narrative was a convenience and the Fed will quietly stop citing it. The deeper question is not whether the Fed hikes once more. It is whether a technology that was supposed to be deflationary has become, for this cycle, an inflation source, and what that does to the pricing of every long-duration asset, crypto included, for the next two years. If the AI supply squeeze is real, then the higher-for-longer regime has a floor under it that no amount of cooling demand can remove. And if it is not real, then a lot of people are about to discover that they were long a narrative the Fed needed, not a mechanism that existed. Either way, the ledger will keep the score. It always does.