Washington Opted Out of AI Regulation. Crypto Already Paid This Bill Once.

Washington Opted Out of AI Regulation. Crypto Already Paid This Bill Once.

Ninety-nine to one.

That was the Senate's margin in July 2025 when it stripped a ten-year moratorium on state AI laws out of the budget package. One senator. Out of a hundred.

That number is the entire policy signal. Washington did not simply decline to write federal AI rules. It told the market, with bipartisan near-unanimity, that it cannot stop the states from writing their own. The federal framework is not delayed. On this political cycle, it does not exist.

Crypto operators have sat through this vote before. They watched a regulator construct an eight-year securities regime on the back of a 2017 investigative report and a single district court opinion. They watched the resulting legal spend migrate an industry offshore, one compliance department at a time. They watched four lenders fail inside eighteen months and the reconciliation land at roughly $8 billion of missing customer funds.

The AI sector now occupies the position crypto occupied in 2017: no statutory framework, explosive capital formation, and a working assumption that clarity arrives before anyone needs it.

The on-chain record says otherwise. The vacuum does not get filled by a bill. It gets filled by litigation, insurance underwriting, and procurement paperwork — three mechanisms that price risk far more slowly, and far more expensively, than legislation ever would.

What Actually Happened

The word "unregulated" is carrying more weight than the evidence supports. What happened is narrower and considerably more consequential.

Executive Order 14110, signed October 30, 2023, established the only quantified federal transparency anchor the AI industry has ever had. Models trained above 10²⁶ floating-point operations had to report training details and safety-test results to the government. That anchor held for fifteen months. Executive Order 14179, issued in January 2025, revoked it and instructed agencies to remove barriers to American AI leadership.

What replaced it was not a thinner rule. At the federal level it was no rule, layered on top of every rule that already applied.

And a great deal already applied. Financial institutions deploying models still answer to SEC and FINRA supervision, including SR 11-7 model risk management. Clinical software still answers to the FDA. Health data still answers to HIPAA. Credit decisions still answer to the Fair Credit Reporting Act. Hiring algorithms still answer to EEOC guidance and Title VII. Deceptive capability claims still answer to Section 5 of the FTC Act — a statute the Commission used in September 2024 to bring five simultaneous enforcement actions under the banner Operation AI Comply.

The states are moving independently, and far from uniformly. Colorado's AI Act imposes impact-assessment duties on high-risk systems. Texas passed TRAIGA. California has been advancing frontier-model transparency obligations. Each framework uses different thresholds, different exemptions, and different definitions of what counts as high risk.

Europe adds a second layer of gravity. The EU AI Act's obligations for general-purpose AI models took effect on August 2, 2025, and they apply to any provider serving EU users regardless of where the model was trained. A US lab that never files a federal disclosure may still owe documentation to Brussels.

So the accurate description is not "unregulated." It is "un-unified." No single federal standard, fifty state standards in various stages of activation, a foreign standard with extraterritorial reach, and a set of sector regulators who never left the room.

The asymmetry matters just as much. While the application layer loosened, the compute layer tightened. Advanced semiconductor export controls have been revised repeatedly, and the direction has been restrictive, not permissive. Washington is not deregulating AI. It is deregulating domestic deployment while continuing to constrain foreign capability. Those are two different policies sharing one headline.

Crypto's AI complex grew up inside that ambiguity. Decentralized physical infrastructure networks, tokenized GPU marketplaces, inference brokers, and agent frameworks now represent tens of billions in aggregate token value, most of it launched between 2023 and 2025. None of it was designed around a federal framework, because none existed. That was treated as permission. It was only absence.

The 2017 Template, Executed in Code

Crypto ran the controlled experiment. The results are worth reading in order.

July 2017: the SEC published an investigative report concluding that a particular token sale was a securities offering. No statute followed. No rulemaking followed. What followed was enforcement — case by case, subpoena by subpoena.

2018 through 2019: registration actions, settlements, and a slow migration of token issuers toward jurisdictions with clearer rules. The United States did not lose the industry. It lost the marginal, compliance-adjacent part of it, which turns out to be the part that files taxes and hires lawyers.

2020 through 2022: DeFi Summer. A genuinely new financial architecture — automated market makers, on-chain lending, composable leverage — built entirely without a regulatory perimeter. Total value locked crossed $100 billion at the peak. No federal rule was written during the entire period. The infrastructure was built on the assumption that none was needed.

2022: the invoice arrived. Terra's algorithmic stablecoin failed in May. Three Arrows Capital followed. Celsius, then BlockFi, then FTX in November, its shortfall ultimately landing near $8 billion in customer funds.

I worked that collapse in real time. In the first twenty-four hours after FTX's balance sheet became public, my team traced the commingled flows — specific USDC transfers, lending protocol exposures, the interlocking positions between Alameda and exchange customer accounts. We published a granular breakdown before most outlets had confirmed the size of the hole. The lesson was never that the breakdown was hard to produce. The lesson was that the data existed the entire time, and nobody had a legal obligation to disclose it until the company was already insolvent.

2023 through 2025: enforcement, bankruptcy, and finally legislation. Stablecoin rules arrived in 2025. Market structure legislation cleared one chamber the same year. Eight years from the first enforcement action to a statutory framework.

Now map that timeline onto AI. The withdrawal of the compute threshold disclosure puts the sector in crypto's pre-2017 position: capability information exists, and no one has to publish it. Voluntary commitments — the 2023 White House pledges from fifteen companies, the 2024 Seoul frontier safety commitments — occupy exactly the position that self-regulatory token standards occupied in 2018. They are real, they are sincere in places, and they have no verification mechanism.

Under competitive pressure, unverifiable commitments dilute. That is not a claim about character. It is an observation about incentives.

Where the Load Actually Goes

If AI agents become significant on-chain actors, the first system to feel it will not be the court system. It will be the fee market.

The precedent is Solana, April 2024. At the peak of that episode, roughly three-quarters of non-vote transactions were failing.

Solana's congestion was not caused by human users. It was driven by automated transaction flows — bots competing for the same blocks, paying priority fees that ordinary wallets could not match. The network survived. The fix was a set of engineering decisions: a new transaction forwarding protocol, stake-weighted quality of service, and a more aggressive local fee market that let validators price contention directly.

Read that sequence again, because it is a preview. Automated agents do not experience congestion the way humans do. They experience it as a cost variable. When priority fees rise, a human wallet abandons the transaction and tries tomorrow. An agent re-prices and resubmits, up to its budget constraint, inside the same block.

The distributional result is predictable. Agent-driven demand does not merely raise fees. It raises them in a way that systematically prices out the participants with the lowest willingness to pay. Fee markets are efficient. They are not fair, and they were never designed for a world where the marginal transaction originator has no fatigue and no bedtime.

The second-order problem is architectural. Most of the throughput that AI-adjacent activity would run on today sits behind a single sequencer. Arbitrum has one. Base has one. Optimism has one. The operator is a company, and that company can order, delay, or censor transactions. This has been true for years, and the phrase "decentralized sequencing" has been a slide in a deck for at least two of them.

A centralized sequencer is a single point of policy failure. If AI agents become the dominant fee payers on a chain, whoever operates that chain's sequencer becomes the de facto regulator of that agent population — with no statute, no due process, and no appeal.

That is not a hypothetical about jurisdiction. It is a description of where the switch currently sits.

Add the extraction layer and the picture sharpens. Latency is the only moat in MEV. Models that read the mempool and price inclusion are structurally faster than any human searcher, and they get faster every quarter. The mempool's congestion becomes an input to a model, and the model's output becomes a bid that humans cannot match. The competitive equilibrium of MEV is an arms race measured in milliseconds, and the equilibrium of an arms race is that the fastest participant captures most of the value. Human searchers have been exiting that market for two years. Agent searchers will finish the job.

Compute Is the Binding Constraint, Not the Statute

The regulatory debate is downstream of a physical one, and the physical one cannot be legislated away.

Hyperscale capital expenditure crossed into the $350-400 billion range on a combined basis in 2025. That number is driven by model capability, inference demand, and competitive positioning. Deregulation did not create it. Deregulation only removed the tail risk that a policy change interrupts it.

The constraint that does not bend is power. PJM, the largest US grid operator, cleared its capacity auction at roughly nine times the prior year's price — a jump driven substantially by data center demand. That is the market saying something a policy statement cannot. The electrons are the scarce input, and no executive order produces a transformer.

Tokenized compute markets ran into the same wall from the other side, and their failure mode is instructive. In April 2024, a decentralized GPU network suspended operations after metadata spoofing inflated its reported supply of available hardware. Listed capacity and verifiable capacity turned out to be two different numbers wearing the same label.

I spent 2021 auditing metadata infrastructure for NFT marketplaces and found the identical structural flaw in a different costume: roughly 40% of supposedly permanent digital assets resolving to centralized servers that could be taken down at will. The lesson transfers cleanly. In any market where supply is self-reported, the reported supply converges on whatever figure maximizes the token narrative, and the gap between that figure and the physical figure is where the losses live.

For compute markets the gap is measurable. A GPU is not listed until someone benchmarks it, attests to its identity, and can be held to a service-level agreement under sustained load. Everything short of that is a marketing number in hardware clothing.

This relocates the real bottleneck. Congress can decline to audit models. It cannot make electricity cheap, transformers fast, or verifiable compute abundant. The infrastructure-first reading of the AI trade is blunt: stop watching the Federal Register and start watching interconnection queues. The base layer's congestion in 2027 will be measured in megawatts, not gas.

The Pricing Error in the AI-Agent Trade

Now the bear market lens, because this is where money is actually made and lost.

AI-adjacent tokens — agent frameworks, inference markets, data networks — traded through 2024 and 2025 on a straightforward thesis: AI is coming, crypto is the permissionless layer for it, and no regulator will stand in the way.

Two of those three claims hold. The third is where the pricing error sits.

Start with the revenue base. An agent framework's on-chain fee revenue comes from the transactions it routes. In a risk-off market, routed volume falls faster than token price, because volume responds to opportunity and price responds to narrative. Trailing fee revenue on most agent tokens does not support the valuations they carried at the top, and in a drawdown that gap widens rather than closes. Bear markets do not compress narratives evenly. They compress the ones with the weakest cash conversion first.

Then separate the two claims that get bundled together. There is a real business in verifiable compute, inference routing, and settlement infrastructure. There is a much larger market cap in tokens whose product is a mascot and whose economics are an emission schedule.

I have written this before about DeFi, and it applies here without modification. An incentive program does not create activity. It rents activity, and the rental clears at a price set by the emissions rather than by demand. Stop paying and the activity stops. The only open question is whether anything durable got built while the meter was running.

The trap in AI tokens is subtler than in DeFi. Even a genuinely useful agent still settles its transactions in stablecoins, holds assets through custodians, and relies on exchanges for liquidity. Every one of those rails received a regulatory framework in 2025 — stablecoin legislation signed, market structure legislation advancing.

The agent may be unregulated. Its plumbing is not.

That inversion is the trade nobody is positioned for. The AI token market is priced on the assumption that the rail is permissionless and the rulebook is empty. In reality the rail is becoming the most regulated part of the stack, and the rulebook is being written one state at a time, by regulators with subpoena power and no obligation to read the whitepaper.

Litigation Is the Regulator of Record

When the legislature declines, the court system inherits. It is a slower, costlier, and less predictable regulator, and it operates strictly after the loss has already been booked.

The mechanism is already visible. In February 2024, an employee of the engineering firm Arup was deceived into transferring roughly $25 million during a video call populated entirely by deepfaked colleagues. The fraud was real. The meeting was not. No statute specifically addressed it.

Legislative response to that class of harm followed a now-familiar pattern. It arrived narrowly, after a visible incident, and only where the harm was graphic enough to build a coalition. The TAKE IT DOWN Act, signed in May 2025, addressed non-consensual intimate imagery. It did not create a general federal AI liability framework.

So the defaults govern. Product liability, negligence, deceptive trade practices, and state attorney general enforcement fill the space a federal statute would otherwise occupy. Copyright questions accumulate as case law rather than as legislation, which means the ruleset is discovered one docket at a time, at a cost of roughly $10-30 million per contested matter.

For a builder, that is the worst available configuration. A statute tells you the boundary before you build. Case law tells you the boundary after you have already shipped, sold, and scaled — and the judgment arrives with interest attached.

The cost does not stop at the courtroom. It stops at the underwriter. AI liability insurance remains an undeveloped market precisely because there is no federal liability framework to price against. Without agreed standards for audit, testing, and disclosure, an insurer cannot set a premium. It either prices high and narrow, or it declines the risk outright.

That is the binding constraint on high-value deployment, and it is almost entirely absent from the policy conversation. A hospital will not run a diagnostic model inside clinical workflow, and a bank will not route credit decisions through one, if the liability sits unhedged. You do not need a federal statute to slow AI adoption in the sectors where harm is expensive. You need an insurance market that cannot price the exposure and therefore will not write the policy.

Crypto learned this from the other direction. Its legality was adjudicated in courtrooms for years before it was legislated, and the industry carried that uncertainty as a permanent line item in its cost of capital. The AI sector is about to run the same tab, and it has no idea what the meter reads.

The Compliance Option Nobody Is Pricing

The coverage misses one mechanism entirely: federal procurement.

The government is the largest single buyer in the American economy, and its purchasing rules function as a de facto standard regardless of what Congress does. When federal contracts demand specific documentation, evaluation methodology, and reporting from AI vendors, those requirements propagate. Vendors build to them because the alternative is losing the contract. Enterprise buyers then adopt them because they are the only benchmark available.

This is how standardization actually happens in the absence of legislation. It arrives through procurement language rather than statute, and it stays invisible to anyone reading only the Congressional Record.

The investment implication is positional, not directional. In a market with no federal rule, the value of a verified compliance capability does not go to zero. It becomes an option. Companies that built auditable evaluation pipelines, reproducible safety testing, and documented model governance did not waste the money. They purchased the right to sell into regulated procurement and the right to survive the first crisis-driven statute without a rebuild.

The historic analogy is unflattering to the industry. The 2023 White House voluntary commitments and the 2024 Seoul declarations were genuine efforts, and they were also unverifiable. There is no audit trail, no reporting cadence, and no consequence for non-compliance. Unverifiable commitments dilute under competitive pressure, and they dilute fastest in the quarters immediately preceding a capability race.

The rational posture is not to wait for the framework. It is to assume the framework arrives within two to three years, triggered by a specific incident, and to hold compliance capability as a call option rather than a cost center.

The Contrarian Read

Here is the angle the AI-versus-regulation coverage systematically misses.

The assumption buried in every "Washington opts out" headline is that crypto is the laggard still waiting for its turn at legitimacy. The sequencing suggests the inverse.

Crypto received its first real statutory framework in 2025 — stablecoin legislation signed into law, market structure legislation advancing through Congress. AI received nothing except the removal of a disclosure threshold. For the first time since 2017, crypto is closer to having a legal perimeter than the technology sector that was supposed to leapfrog it.

That inverts the trade. If autonomous agents eventually need to transact, hold value, and settle obligations, they need rails. The rails have rules now. An agent operating on a licensed stablecoin network is transacting on regulated infrastructure with a defined compliance perimeter. An agent routed through unlicensed intermediaries is not.

The visible market is buying agents. The durable business is selling them the substrate: regulated settlement, custody with an audit trail, and attestation services that let a buyer prove to an underwriter that the model behaved as documented. None of that is glamorous. All of it is billable, and unlike an emission schedule, none of it requires the token price to hold.

There is a second blind spot worth naming. The industry keeps debating whether regulation will be heavy or light. The more useful question is whether the regulator has the capacity to be either. Assessing a frontier model requires scarce technical talent that government does not currently retain at scale, and that talent is being paid multiples elsewhere. The binding constraint may not be political will at all. It may be that the state cannot hire the people who would write the rule.

If that is true, then voluntary frameworks, procurement standards, and insurance underwriting are not interim measures. They are the durable governance layer, and they will decide what ships long before any statute reaches a floor vote.

What to Watch

The next twelve months hand you three signals and no summary.

Watch the first nine-figure AI tort judgment. That is the event that converts an absence of legislation into a crisis-driven statute. It is the AI sector's version of a lender collapsing on a Friday afternoon with the withdrawals already frozen.

Watch interconnection queues and capacity auction clearing prices. Physical scarcity outlasts political cycles, and no executive order has ever moved a megawatt.

Watch whether on-chain agent volume shows up as sustained fee revenue after emissions taper, or disappears alongside them. That single number separates a business from a token.

Washington voted 99-1 to leave the question open. The question does not stay open on its own. Something closes it — and the only remaining uncertainty is whether you find out from a legislator, or from a subpoena.