Hook
Somewhere in the wreckage of the previous administration's AI executive order sits a number that mattered more than any speech: 10^26 floating-point operations. That was the training-compute threshold above which a frontier model had to be disclosed to the federal government and subjected to adversarial testing. On the order of a handful of systems on Earth ever cleared it. When the order was rescinded in January 2025, the disclosure obligation vanished with it โ and so did the only binding federal reporting requirement attached to what is now the largest capital-expenditure program in industrial history.
I read a short piece on this from a crypto vertical, headlined around the President "dismissing AI safety concerns." It got the direction right and the mechanism wrong. There was no quoted statement in it. No date. Five of six claims carried no attribution whatsoever. That is not a dataset; that is a headline with a mood attached.
So let me audit the trade that followed, not the politics. The question for anyone carrying crypto risk is not whether a politician believes in AI alignment. It is which cash flows get repriced when a compliance constraint disappears โ and which ones merely pretend to.
Context
The rescission itself ran about eleven lines. Executive Order 14179, signed in the first week of the second term, revoked the October 2023 framework that had required developers of frontier models to report training runs above the 10^26 FLOP threshold, submit red-team results, and share safety protocols with federal agencies. The threshold was never a rounding error. A GPT-4-class model sits around 2ร10^25 FLOPs; the systems then in pre-training were designed to clear 10^26. The rule was written to catch exactly the models that did not yet exist, and it was deleted at precisely the moment they started training.
What replaced it was not a lighter rule. It was the absence of a rule, dressed as a strategic posture. The personnel signature is consistent: a single appointee now holds the AI and crypto portfolios, and the safety contingent that had occupied positions at NIST and parts of the State Department lost its forward ground. In February 2025, at the Paris AI Action Summit, the American delegation declined to sign a joint declaration containing language about "inclusive and sustainable" AI, and criticized European overreach from the podium.
Here is the first place the coverage misleads. "Dismisses AI safety concerns" is an editorial verb, not a quotation. Practitioners of politics rarely say they do not care about safety; they say they will not let regulation strangle leadership. The framing is a priority ranking, not an existential claim. The distinction determines what actually got repriced.
I want to be explicit about the epistemic state of the source. No date. No attributed quote. A crypto-specific outlet covering an AI governance story for which it has no beat reporters. That is a cross-domain paraphrase with a high failure rate, and my confidence in any specific claim drawn from it sits one grade lower than it would from a primary document. What I can defend is the direction: the federal posture moved from precaution to acceleration between 2023 and 2025. Everything downstream is inference, and I will label it as such.
Now the liquidity map, because that is where this stops being a policy story. Global private capital formation is being absorbed by AI infrastructure at a rate with no post-war precedent. A single announced joint venture โ OpenAI, Oracle, SoftBank, MGX โ carried a headline figure of $100 billion committed and $500 billion aspirational over four years. Hyperscaler capex guidance for 2025 alone crossed $300 billion across the four largest cloud operators. That capital does not come from nowhere. It is drawn from the same pool that funds every other duration-sensitive asset, including digital assets, and it is being deployed into a physical supply chain with three binding constraints: advanced packaging, high-bandwidth memory, and electricity.
Core
The transmission channel from AI deregulation into crypto is not sentiment. It is the power market, and I can trace it.
When compliance friction falls, the start date of the next training run moves forward. That is the whole mechanism. A lab that no longer files a report, no longer waits for a review window, and no longer staffs a compliance function pushes its cluster into service weeks earlier. Multiply that across every frontier lab and every well-funded challenger, and the marginal demand for accelerators, for co-location capacity, and above all for firm power steps up. Grid interconnection queues in the major US data-center corridors already ran past three years in several jurisdictions before any of this. The queues do not shorten when demand accelerates. They lengthen.
That is where Bitcoin miners enter the analysis, not as an AI proxy but as the marginal seller of a scarce input. A mining site is, functionally, an interruptible power contract with a substation attached. Post-halving, with the block subsidy at 3.125 BTC, the economics of converting raw electricity into hashrate have compressed hard: hashprice has spent most of the cycle near or below the cash cost of production for all but the most efficient fleets. Meanwhile an AI operator will pay a multiple of the mining gross margin for the same megawatt, on a long-dated contract, with investment-grade counterparty risk. Several listed miners have already signed or negotiated exactly this conversion โ hosting agreements and HPC leases that value the site on contracted power and interconnection rights rather than on ASIC count.
The insight most crypto holders have not internalized is that AI deregulation is a negative cash-flow event for pure-play hashrate and a positive re-rating event for power-interconnected infrastructure. Both assets trade under the same ticker universe. The deregulation did not create a new subsidy for crypto. It raised the opportunity cost of using a megawatt to guess a nonce.
Follow the second-order effects further. Conversion removes hashrate from the network. Difficulty adjusts downward, which mechanically improves unit economics for the survivors โ and improves them most for the operators with the cheapest power and the newest silicon, which is to say the ones who were already winning. Consolidation accelerates. The network hashrate becomes less a measure of miner optimism and more a measure of how much stranded, non-convertible power exists. That is a structural change in what the metric means, and most dashboards have not caught up.
Now the part of the crypto complex that marketed itself as the direct beneficiary, and which in my assessment has the weakest claim: decentralized compute marketplaces. The thesis was elegant. Accelerators are scarce, allocation is controlled by a few hyperscalers, therefore a permissionless marketplace that clears idle GPU-hours at spot should capture a scarcity premium. Render, Akash, io.net, and the training networks built on similar primitives all priced some version of that argument.
Then January 2025 happened. A Chinese lab published a reasoning model that approached frontier performance with a training budget reported in the single-digit millions of dollars and a fraction of the accelerator-hours the consensus assumed were mandatory. The engineering was the interesting part โ mixture-of-experts routing, mixed-precision arithmetic, aggressive communication compression โ but the market read it as a hardware-demand shock, and repriced accordingly. The correct reading is narrower and more damaging to the decentralized-compute thesis than a simple selloff: if software efficiency substitutes for raw FLOPs at the margin, the scarcity premium on distributed compute narrows in exactly the regimes where it was supposed to widen.
I conducted a similar autopsy once before. During the 2020 DeFi summer I spent three weeks pulling emission schedules and liquidity-mining contracts apart, and the conclusion was arithmetically unavoidable: the headline yield was a transfer from a token treasury, not a return on capital. The same audit logic applies here. When a compute marketplace pays suppliers in its own token for GPU-hours, the supplier's revenue is a function of the emission schedule, not of end-user demand for inference. Strip the subsidy and you can measure real utilization. In several of these networks, real paid utilization has historically been a small fraction of reported capacity, because reported capacity is a marketing input and paid inference is the actual product. Unaudited yields are not income; they are risk.
The service-level problem compounds it. Enterprise inference buyers do not purchase raw FLOPs. They purchase determinism: latency percentiles, data residency, confidentiality, an indemnity clause, and a phone number to call when it breaks. A permissionless marketplace of anonymous providers can win on price for batch jobs and research workloads. It cannot currently win a regulated bank's production inference, which is where the margin lives. That is a distribution problem, not a cryptography problem, and no amount of decentralization changes the procurement conversation.
Where the AI build-out does create genuine, non-speculative demand for crypto rails is settlement. Autonomous agents transacting continuously cannot use correspondent banking. Card networks price per transaction in a way that makes a two-cent micropayment uneconomic, and they impose chargeback regimes designed for human fraud patterns. Programmable stablecoin rails solve this: instant finality, negligible marginal cost, composable authorization logic, and no counterparty that can unilaterally freeze a machine's working balance. Payment-intent standards aimed at HTTP-native settlement have started appearing for exactly this reason, and the design space is converging on stablecoins as the unit of account for machine-to-machine value transfer.
This is the substantive divergence between the two competing digital-dollar architectures, and it is worth stating without euphemism. A retail central bank digital currency is a surveillance instrument with a payment interface bolted on: programmable constraints on who may hold, spend, and where. A permissionless stablecoin on a public chain is the opposite construction โ the issuer is a chokepoint, but the rail itself is not. These are not two implementations of the same idea at different maturity levels. They are incompatible designs with mutually exclusive political premises, and the AI agent economy, which requires high-frequency machine settlement, will be the first commercial arena where the choice becomes visible to non-specialists.
Stablecoin net issuance remains the cleanest risk-appetite gauge I have. I built a crude version of it in 2017 while manually tracking whale flows across Ethereum and early EOS, and the correlation between issuance spikes and subsequent risk-asset expansion held well enough to call the January 2018 peak with tolerable accuracy. The instrument is more crowded now and the signal is noisier, but the direction of travel still beats price momentum as an input.
One more vector, quieter and structurally worse. Governance. Delegation already concentrates DAO voting power in the hands of a small set of recognizable addresses, because the cost of researching a proposal exceeds the value of a single vote for almost every holder. Introduce agentic delegates โ an LLM with a public voting log, a stated policy function, and a better track record than any human delegate โ and the concentration accelerates, because the scarce resource in governance was never capital. It was attention. Agents do not have an attention budget; humans do. The equilibrium is a handful of model-driven delegates controlling quorum, and the accountability mechanism for a delegate whose reasoning is generated is not obvious. Code is law, but incentives are the reality.
Contrarian
The market's working assumption is that crypto is a high-beta expression of AI risk appetite, and that AI deregulation is therefore a crypto catalyst. I think that is a liquidity artifact misread as a fundamental.
When the discount rate moves, everything long-duration trades together. A 0.7 correlation between digital assets and the AI-heavy equity complex in a given quarter tells you that both are discount-rate instruments held by the same marginal buyer, not that they share a business model. The underlying cash flows are unrelated. AI capex is a claim on electricity, silicon, and cooling. Bitcoin is a claim on a monetary network with a fixed issuance schedule. A stablecoin rail is a claim on transaction volume. None of these businesses acquires customers from the others.
The policy re-rating is similarly unearned. Removing a reporting obligation saves a handful of labs a rounding error relative to their capital budgets. It does not lower the cost of a training run, add a megawatt to a constrained grid, shorten an interconnection queue, fabricate high-bandwidth memory, or produce a person capable of doing alignment research. So any token that rallied on "AI deregulation" rallied on a narrative, and the arithmetic underneath it was unchanged. I have watched this pattern enough times to be unsentimental about it: the price movement was real, the causal story was fabricated after the fact.
The unpriced tail is the pendulum. The original framework existed for fifteen months before it was revoked. Anything whose valuation depends on a policy that can be reversed inside one electoral cycle deserves a discount, not a multiple. The specific trigger to watch is not a court ruling or a congressional hearing. It is an incident โ a deepfake that moves an election, a model that causes measurable harm, a breach exposing training data โ after which the reporting requirement returns with punitive terms attached. The probability is moderate. The impact on the affected cohort is severe. That asymmetry is currently free.
Takeaway
Stop reading the press conference and start reading the interconnection queue. Watch the spread between hashprice and delivered power cost in the major deregulated grids; that spread, not a policy headline, decides which mining sites become AI campuses. Watch stablecoin net issuance as the honest read on risk appetite. And watch whether the AI complex and the crypto complex keep sharing a beta, or whether they start sharing a bill of materials. The question for the next cycle is not whether AI wins. It is who ends up paying for the electricity โ and whether they knew they were bidding.
