Goolsbee's AI Inflation Thesis: The Energy Channel Nobody Is Pricing Into Bitcoin Mining
On October 13, Chicago Federal Reserve President Austan Goolsbee called artificial intelligence an inflation issue. He offered no direction. He did not say AI raises prices. He did not say AI suppresses them. He moved AI from the technology desk to the monetary policy desk, and the relocation is the event. The wire item ran to roughly three sentences. No data. No methodology. No link to a primary transcript.
By the evidentiary standard I apply to any protocol disclosure, that is an unverified claim from a secondary source. I have spent years treating three-sentence press fragments the way I treat unaudited tokenomics: as a hypothesis, not a fact. The ledger does not lie, but the narrative does. So I will not trade the headline. I will trace the transmission channel the headline opens. That channel runs through electrical power, and electrical power is the same resource Bitcoin mining has been bidding for since 2016.
The AI-crypto convergence has been sold to retail as a synergy story. Compute meets consensus. Autonomous agents settle on-chain. Decentralized GPU markets price idle silicon. Most of that is marketing. A smaller fraction is a genuine structural collision, and the collision is over a fixed input: megawatt-hours. When a Federal Reserve governor classifies AI as an inflation variable, he is implicitly classifying the energy demand curve underneath it. Bitcoin sits on that curve. So does every large language model that now signs transactions. The convergence is real. The synergy is not.
The source material here is thin, and I will not pretend otherwise. A single official, a single sentence, a secondary transcription with no year attached. I have flagged the same weakness in my own audits: when the data is absent, the absence is the finding. Silence in the data is a confession. What Goolsbee's silence on direction confesses is that the Fed does not yet know whether AI is a demand shock or a supply shock. That uncertainty is not a flaw in the message. It is the message.
To read it correctly, you have to understand what the Fed actually controls. It does not control the price of compute. It controls the price of money, and it reacts to the price level that money buys. AI enters that equation through three doors: energy, metals, and labor. Each door is physical. Each is slow to open. That slowness is the whole point. A technology can move at software speed while the inputs it consumes move at construction speed, and the gap between those two speeds is where inflation is manufactured.
History offers one precedent, and it is instructive. In the late 1990s the Fed confronted a productivity boom it could not measure in real time. Officials debated whether the internet had raised the economy's speed limit. Greenspan eventually leaned toward yes, and that judgment helped justify the easy policy that preceded the 2000 bust. Goolsbee's AI remark is the same debate in a new costume: is this a supply-side miracle or a demand-side bubble? The Fed got the dot-com answer wrong for years because the data lagged the narrative. There is no reason to assume it will get the AI answer right on schedule.
Consider what AI consumes at the physical layer. A frontier training cluster draws tens of megawatts continuously. Inference at scale draws more, because it never stops. The demand is not digital. It lands on the grid as load, on the metals market as copper and transformer steel, and on the labor market as electricians. None of those inputs scale at software speed. You cannot compile a substation.
This is the mechanism by which a technology narrative becomes a price level. The Fed's core inflation measure excludes food and energy, which is exactly why officials keep flagging energy-adjacent demand as a risk to the headline figure. If AI capital expenditure keeps pulling power and copper into a supply-constrained market, the resulting price pressure shows up in the index the Fed is mandated to control. Goolsbee is not talking about software. He is talking about the cost of running it. The transmission is not instantaneous, but it is monotonic.
Now place Bitcoin mining on the same curve. A modern mining fleet is, functionally, a large flexible load. It buys power under contracts that assume it can curtail on demand. For a decade, that flexibility was a liability. Miners were the marginal buyer nobody wanted at the negotiating table. Since 2022, the same contracts became an asset. AI and high-performance computing operators need firm, dense, grid-connected sites with cooling and power already permitted. Bitcoin miners already hold those sites. The pivot is not a metaphor. It is a contract assignment.
The mechanics of the pivot determine who bears the risk. A Bitcoin miner typically holds a power purchase agreement with a fixed or indexed rate and a curtailment clause. That clause is the miner's hedge: when spot power spikes above the mining margin, the miner shuts off and sells nothing, avoiding a loss. An AI tenant needs the opposite. It needs uptime, not flexibility, because a training run interrupted mid-epoch is wasted capital. So the miner who hosts AI is effectively selling its curtailment option. It is monetizing flexibility and then giving it away. The cash flow looks cleaner. The optionality is gone.
This is where the Fed's framing gets operational. If AI demand is inflationary because it competes for power, then Bitcoin miners who pivot to AI hosting are not diversifying away from volatility. They are converting a hashrate business into a real-estate and energy business that is now directly exposed to the same inflation variable the Fed is watching. The revenue model changes. The incentive does not. Merges change the mechanics, not the incentives. The market is pricing the upside of the pivot and ignoring the exposure it creates.
Trace the on-chain side. Hashrate is the only metric in this sector that cannot be inflated by a press release. It is a physical output: work performed, measured, and committed to the ledger. When miners redirect capacity to AI hosting, hashrate does not vanish. It reprices. Difficulty adjusts on a fixed schedule, roughly every two weeks, and it does so mechanically, without sentiment. That is the value of proof-of-work as a data source. It is a sensor. It tells you what energy actually cost, not what a spokesperson said it cost.
Look at the numbers the sector actually reports. Global hashrate has grown roughly an order of magnitude since 2020, and the energy attached to it has grown with it, even as efficiency per hash improved. That is the Jevons pattern: efficiency gains get reinvested in more capacity rather than in less consumption. AI compute follows the identical curve. Cheaper chips do not reduce total draw; they increase the number of workloads worth running. Any inflation model that assumes efficiency automatically lowers aggregate demand has not internalized Jevons, and the Fed, to its credit, appears not to have internalized it either. Productivity is treated as a residual; Jevons treats it as a ratchet.
Here is the asymmetry I keep returning to. A mining operation is fully exposed to the spot price of power and the spot price of the asset it mines. An AI hosting contract is exposed to a negotiated rate and a counterparty's credit. The first is volatile and liquid. The second is stable and illiquid. When miners migrate from the first to the second, they trade mark-to-market risk for counterparty risk. In a tightening cycle, that looks like prudence. In a credit event, it looks like a trap. I audited the custody structures of the proposed spot Bitcoin ETFs in early 2024 and found the same pattern: a structure that reduced one risk by introducing another, sold as pure improvement. The gap between promise and proof is fatal.
There is a second channel, and it is the one I have spent the most time auditing. In 2026 I documented twelve instances where autonomous agents interacting with DeFi protocols exploited gas-fee prediction errors on Layer 2 rollups, triggering liquidations that no human operator intended. The finding was not that agents are malicious. The finding was that the standards were written for human latency and human judgment. A machine does not read a disclaimer. It reads a function signature.
This is why I now run a machine-readability audit on every protocol I review. The test is simple: can an autonomous agent interact with this contract using only its public interface, without human interpretation of intent? Most fail. Disclaimers, upgradeable proxies, and off-chain governance introduce ambiguity that a language model resolves by guessing. Guessing is acceptable for a chatbot. It is catastrophic for a settlement layer. Code designed for humans is insufficient for an AI-driven economy, and the Fed's new framing is an admission that the AI-driven economy is no longer hypothetical.
If the Fed is now treating AI as a macroeconomic variable, the same logic applies one layer down. AI agents are becoming economic actors on-chain, and the contracts they touch were not compiled for machine-to-machine trustless interaction. When an agent misprices gas, the loss is not a rounding error. It is a liquidation event settled on an immutable ledger, with no appeals process and no counterparty to negotiate with. History is written by the auditors, not the poets.
Connect the two channels and the shape of the risk becomes clear. AI drives physical demand, which drives energy prices, which the Fed treats as an inflation constraint. AI also drives on-chain activity, which drives fee markets, which the same Fed treats as a financial stability question. One technology, two transmission paths, both terminating at the central bank. Goolsbee's single sentence is the first official acknowledgment that the Fed sees both.
The most tradeable element of the statement is not what it says. It is what it contradicts. The dominant market narrative frames AI as deflationary: efficiency gains, cost compression, productivity without wage inflation. That narrative is embedded in the valuation of every AI-adjacent equity and, more quietly, in the AI-token complex that recycled itself after the 2023 washout. Goolsbee's framing inverts the sign. If the Fed prices AI as inflationary, the policy path is tighter for longer, and every asset priced on the assumption of imminent easing reprices.
The AI-token complex deserves separate scrutiny. After the 2023 washout, a cohort of projects rebranded around decentralized compute, inference markets, and agent frameworks. Their valuations now embed a specific macro assumption: that AI is deflationary and that the Fed will ease into it. Goolsbee's framing challenges that assumption directly. If the policy path is tighter for longer, the discount rate applied to those long-duration, pre-revenue token models rises, and the compression is mechanical. The narrative and the cash flow were never the same thing. The ledger does not lie, but the narrative does.
For crypto, the repricing has two directions and they are not symmetric. Rate-sensitive assets, including long-duration growth exposure and a meaningful share of the token market, face a discount-rate headwind if easing is delayed. Physical-infrastructure assets, meaning power, mining sites, and energy contracts, face a demand tailwind if AI keeps pulling load. The same statement is bearish for one basket and structurally bullish for another. Volatility is the tax on unverified consensus, and the consensus here is that AI is a one-way deflationary force. It is not.
I want to be precise about confidence. The source is a three-sentence wire item from a secondary outlet with no year, no transcript, and no directional claim. My confidence in the macro conclusion is low-to-moderate. My confidence in the mechanism, that AI and Bitcoin mining compete for the same constrained energy input and that this competition is now inside the Fed's inflation framework, is higher, because it is verifiable against grid data and mining contracts rather than against a single quote. Privacy is not secrecy; it is control. The Fed controls the narrative by controlling what it declines to specify.
Here is what the AI-crypto bulls got right, and it is the part the skeptics keep missing. If AI genuinely raises total factor productivity, the deflationary thesis is not wrong. It is early. A supply shock that lowers unit costs is disinflationary at maturity. The Fed's caution is a statement about timing, not about direction. Officials are not saying AI will never compress prices. They are saying the compression has not arrived while the demand has. That is a defensible position, and it is the reason a naive trade on AI as a one-way deflationary force has been so expensive to hold.
The blind spot runs the other way. Crypto's energy critics have spent a decade arguing that Bitcoin mining is a pure waste of power. That argument assumed the power had no alternative use. AI just gave it one. The same flexible-load contracts that made miners pariahs now make them strategic counterparties for the most capital-intensive compute buildout in a generation. The waste critique does not survive contact with a data-center shortage. It was never really about energy. It was about narrative, and narratives do not compile.
There is a harder version of the bull case, and it deserves a hearing. If the Fed is wrong, if AI is genuinely disinflationary and the demand spike is transitional, then the tight-for-longer path is a policy error, and the assets it represses are the ones that recover hardest. I do not think that case is proven. I think it is underweighted. The market has priced the Fed's caution as wisdom and the bulls' optimism as naivety. Both are bets. Only one of them is being stress-tested.
The question is not whether Goolsbee is right. It is whether the Fed's framework and the market's framework can diverge indefinitely without a repricing. They cannot. Watch three signals: whether other FOMC voters repeat the AI-as-inflation frame within a month; whether power and copper inflation prints accelerate; and whether Bitcoin hashrate growth decouples from mining revenue as capacity migrates to AI hosting. Source code is the only truth that compiles. When the Fed's framework and the market's framework stop compiling against each other, the reconciliation is the trade.