Two weeks ago, a crypto publication ran a four-hundred-word brief announcing that a Trump-era artificial intelligence task force would concentrate on three risk categories โ cybersecurity, biosecurity, and "model control" โ while relying on voluntary measures from industry. The piece carried no sourcing. No agency name. No document number. No named official. It contained not one line of blockchain content, which is a strange thing for a publication whose entire revenue model rests on readers who care about blockchains.
Most of my feed scrolled past it. I did not. Because the architecture of that brief โ a national-security frame, undefined risk categories, and a voluntary compliance regime โ is not a new governance model. It is the oldest and most thoroughly discredited governance model in the crypto industry, now being imported into the most capital-intensive technology sector in human history. And the two sectors share a substrate: compute. When voluntary frameworks fail, as they systematically do, they do not fail in isolation. Tracing the silent hemorrhage of algorithmic trust is no longer a crypto-specific exercise. The hemorrhage now runs through the same silicon, the same power contracts, and the same liquidity pools.
The Policy Map, Such As It Is
The policy shift is straightforward to map, even if the brief that announced it was not. Under the previous administration, the operative framework was Executive Order 14110, which compelled developers of models trained above a specific compute threshold โ ten to the twenty-sixth floating-point operations โ to report to the federal government and submit to safety testing. That was a mandatory regime, however lightly enforced. The current orientation inverts it: voluntary measures, national-security framing, and a deliberate refusal to specify what "model control" actually means.
Meanwhile the European Union's AI Act moves in the opposite direction, a phased mandatory framework that classifies systems by risk tier and imposes obligations that scale with capability. China operates a third path โ algorithm registration and content-safety mandates โ mandatory but scoped to deployment rather than training. Three regulatory philosophies, one shared substrate.
Every one of these frameworks, whether it admits it or not, is ultimately a framework about compute. Training frontier models requires clusters that consume gigawatt-scale power. Running inference at consumer-internet scale requires the same. And the crypto industry โ specifically the decentralized-compute and AI-agent segments โ has spent three years positioning itself as the alternative supply of exactly that resource: idle GPUs, distributed inference, verifiable computation markets, tokenized data.
Here is the part the crypto press missed when it aggregated that brief. If the United States settles on a voluntary framework for AI safety, it is not simply deregulating AI. It is declining to define the safety obligations of any compute provider, including the decentralized ones. That vacuum will be filled. The only questions are when, and by whom. My audit work on stablecoin reserves taught me that voluntary disclosure regimes do not stay voluntary forever; they convert into enforcement regimes the moment a loss becomes large enough to produce a political constituency.
This is why I am writing about an AI policy brief in a crypto publication rather than dismissing it as category confusion. The brief is a leading indicator, and it points at infrastructure that crypto investors own. The signal is not "AI deregulation is bullish." The signal is that a specific governance technology โ voluntary self-restraint โ is being deployed across the sector that now competes with crypto for capital, talent, and watts. And that technology has a documented failure mode that crypto has already survived once.
The Voluntary Framework Is a Governance Technology โ and It Has a Failure Mode
I want to be precise here, because "voluntary" is doing a lot of unexamined work in the current policy debate. A voluntary framework is not the absence of governance. It is a specific governance technology, with its own mechanics and its own failure mode. It delegates the definition of acceptable behavior to the regulated party, and it substitutes reputational discipline for legal compulsion.
Designing the cage to see how the bird flies is a legitimate scientific instinct. It is a catastrophic regulatory one.
The failure mode is well-documented across domains, and it is structural rather than moral. Voluntary frameworks fail when three conditions hold simultaneously: the risk is low-probability but high-damage; the negative externality is socialized while the decision is privatized; and the damage is irreversible once realized. Frontier AI safety risks satisfy all three. So did the risks that DeFi's self-regulatory era was supposed to manage.
Consider the record. The Frontier Model Forum, launched in 2023 by the largest labs, was a voluntary coordination body with no enforcement mechanism and no independent verification. The various responsible scaling policies published by individual labs were unilateral, self-defined, and self-audited โ they set their own thresholds and reserved the right to revise them. On the crypto side, the same pattern appeared a decade earlier: exchange self-regulation, proof-of-reserves attestations that were not audits, and DeFi protocols that published risk parameters nobody was obligated to respect.
In 2022, during the bear-market cascade, I worked with two independent cryptographers to audit the reserve transparency of three major stablecoins. We found a fifty-million-dollar discrepancy in the proof-of-reserves reporting of a mid-tier algorithmic stablecoin. The attestation was technically true and practically meaningless โ it attested to a snapshot that excluded encumbered collateral and contingent liabilities. I did the initial forensic accounting alone, because my instinct is to trust no reconciliation I have not performed myself, and only then sought peer review. That discipline kept my portfolio out of a position that eventually lost sixty percent of its value. The lesson was not that the auditors were fraudulent. The lesson was that voluntary attestation selects for the appearance of solvency rather than solvency itself. Liquidity is a ghost; solvency is the body.
That is the precedent now being imported into AI governance. A voluntary framework for cybersecurity, biosecurity, and model control will select for the appearance of safety. The labs with the best communications teams will look safest. The labs with the best safety teams will have no way to prove it. And the market will be unable to distinguish between the two, which means the market will systematically underpay for real safety and overpay for its performance.
There is a second-order effect that matters even more for capital allocation. When compliance is voluntary, compliance becomes a marketing function rather than an engineering function. The budget that would have gone to red-teaming goes to messaging. The incentive gradient runs from substance to signaling, and it runs there fast, because signaling has a shorter payback period. In DeFi, this produced the era of the audit badge โ the logo on the homepage that substituted for actual code review. In AI, it will produce the safety badge, the ethics statement, the responsible-scaling policy whose thresholds are conveniently set just above the current frontier. The badge is not evidence. It is the opposite of evidence, because it crowds out the thing it claims to represent.
Compute Is the Shared Substrate โ and Therefore the Shared Attack Surface
The crypto angle here is not incidental. It is the whole thing.
The decentralized compute sector โ Render, Akash, io.net, Bittensor, and a long tail of smaller networks โ has built its entire investment thesis on a single claim: that the marginal cost of AI computation can be reduced by aggregating idle and distributed hardware, and that this supply can be made verifiable. Whether that thesis is technically sound is a separate question, and one I am skeptical of. But the regulatory question is the one nobody is asking.
If AI safety obligations are voluntary, then no decentralized compute provider has a defined duty to screen its workloads. A GPU cluster that rents compute to an anonymous customer has no mandated obligation to detect whether that customer is training a bioweapon-design model or generating automated cyberattacks. Under a mandatory regime, that provider would be a regulated entity with reporting duties. Under a voluntary regime, it is a neutral infrastructure provider that has committed to best practices it cannot be compelled to follow and cannot be audited against.
This is not a hypothetical vulnerability. It is the exact shape of the AML problem that decentralized infrastructure has never solved. The crypto industry spent a decade arguing that neutral protocols cannot be responsible for the behavior of their users, and it lost that argument at the enforcement layer โ not in legislation, but in the case-by-case application of sanctions and criminal liability. Tornado Cash was not banned by a statute. It was banned by an interpretation.
The AI version of that interpretation is now being pre-written in the gap between voluntary and national security. Here is the paradox that the brief gestured at but did not name: the same government that is loosening domestic AI safety obligations is tightening export controls on the hardware that makes AI possible. Loose at home, tight abroad. This is not incoherent. It is a coherent strategy in which the state retains control over the diffusion of capability โ through the chip supply chain โ while declining to regulate the behavior of capability that already exists domestically.
For crypto, the implication is severe. Decentralized compute networks are, structurally, capability-diffusion engines. Their value proposition is precisely that they route computation around the chokepoints that export controls and API monitoring depend upon. Under a national-security framing, that is not a feature. It is a vulnerability waiting for a triggering event. And triggering events, in a voluntary regime, are the only mechanism by which rules get written.
I have watched this movie before. In 2024, as a junior researcher in Ho Chi Minh City, I spent six months monitoring the State Bank of Vietnam's pilot for a digital dong. I documented over two hundred technical inefficiencies in the central bank's distributed ledger implementation โ settlement latency, privacy leakage, key-management gaps. The pilot did not fail because of those inefficiencies. It proceeded. What killed the initial enthusiasm was not the architecture but the discovery, by stakeholders, of what the architecture implied. Once people understood what a programmable settlement layer could do, the political cost of the pilot rose faster than its technical benefit. I delayed my report by a month to map the entire settlement layer before publishing, because a partial map of a monetary system is worse than no map at all. The same discipline applies here. A partial map of an AI safety regime will mislead every investor who uses it.
The structural lesson is transferable: infrastructure that concentrates capability attracts governance, and infrastructure that diffuses capability attracts enforcement. Decentralized compute diffuses capability by design. It has optimized for the property that regulators will eventually penalize. This is not a reason to panic-sell the sector. It is a reason to stop valuing it as if its regulatory risk premium were zero.
The AI-Agent Economy Has No Human to Comply
Now let me make the argument that I think is genuinely underpriced, and that requires me to go beyond the brief.
In 2026, I designed a theoretical framework for AI agents using micro-transactions on a blockchain for data verification. The model assumed ten thousand autonomous agents performing independent audits, generating roughly two million dollars in daily transaction volume. I spent two months refining the game-theoretic incentive structure before publication, because the model is only interesting if the incentives are sound without a central coordinator.
Here is what that work taught me about voluntary safety frameworks: they presuppose a voluntary actor. An autonomous agent is not a voluntary actor. It is a deterministic process executing a policy. When a human operator voluntarily commits to safety practices, the commitment is enforced by the operator's own judgment, reputation, and legal exposure. When ten thousand agents transact on-chain, there is no operator in the loop for most individual actions. The commitment cannot be voluntarily honored because there is no one present to honor it.
This is not a philosophical quibble. It is a jurisdictional collapse. Every voluntary framework in existence assumes a responsible legal person who can be shamed, sued, or deplatformed. Autonomous economic agents dissolve that assumption. And the crypto industry is building the rails for exactly that dissolution.
The only enforceable control over an autonomous agent is at the substrate: the compute it runs on and the API it calls. This is why the compute layer matters so much. It is the last chokepoint where a voluntary framework can be converted into an involuntary one. And it is the exact chokepoint that decentralized compute markets are designed to bypass.
So we have a collision. The policy direction is voluntary safety at the application layer. The technology direction is autonomous agents at the application layer and decentralized compute at the substrate. Voluntary safety cannot reach the agents, because they have no capacity for voluntary behavior. And it cannot reach the substrate, because the substrate is being deliberately decentralized. The result is a safety regime with no point of application โ a framework that exists on paper and touches nothing. Code is law, but humans write the loopholes, and here the loophole is the entire architecture.
I am not arguing that autonomous agents are malicious. I am arguing that they are ungovernable by consent. A framework that relies on consent applied to a system that cannot consent is not a framework. It is a press release.
There is a further wrinkle that the agent-economy builders have not priced. Autonomous agents generate an audit trail by construction, because they transact on a ledger. That audit trail is a discovery instrument. It means that when enforcement eventually arrives, the evidentiary record already exists โ every inference request, every payment, every counterparty. The crypto industry's greatest compliance asset and greatest legal exposure are the same thing: an immutable, queryable history of everything it did. The ledger does not sleep, it only waits. When the first serious AI incident is attributed to agent infrastructure, the prosecution will not need to reconstruct the facts. The facts will be waiting.
Model Control Risk Is Undefined โ and Undefined Categories Get Defined Ex Post
The brief's use of the phrase "model control risks" deserves its own forensic treatment, because the phrasing is nonstandard. The more common terms of art are frontier model risks or catastrophic risks. Model control could mean at least three different things: controlling the diffusion of model capabilities, controlling access to model weights, or controlling model behavior at inference. Each implies a completely different policy apparatus. Capability-diffusion control implies compute thresholds and export controls. Weight control implies cybersecurity and access management. Behavior control implies alignment, monitoring, and content policy.
The brief does not choose. It leaves the term undefined. And in regulatory practice, undefined terms are not neutral. They are reservoirs of discretionary power that get filled after the fact.
This is the single most important thing a crypto reader should take from the entire episode. The crypto industry has lived for five years under ex post rulemaking โ rules that were not written in advance but applied retroactively to conduct that was legal when performed. The Securities and Exchange Commission's approach to token classification, the Treasury's approach to mixing services, the banking regulators' approach to custody: none of these were legislative. They were interpretations, issued after the fact, against actors who had already committed capital.
An undefined category in an AI safety framework is the same instrument. It guarantees that the first serious AI-related incident โ a cyberattack, a biosecurity scare, a model-escape scenario โ will be followed by a definition of model control written by whoever is angriest. And whoever is angriest will not be the industry. It will be the agency that has to explain why the voluntary framework it championed did not prevent the incident.
I have a specific prediction about how this unfolds, and it is not comfortable. The first incident will not be attributed to a frontier lab, because frontier labs have the most to lose and the best incentives to self-police. It will be attributed to the cheapest available compute and the most permissionless available rails โ which is to say, to the decentralized end of the market. The decentralized compute sector will become the designated villain of the first AI safety crisis, in the same way that crypto became the designated villain of the first major ransomware wave. The mechanism is identical: when an ungoverned technology produces a salient harm, the enforcement response targets the most visible permissionless infrastructure, regardless of actual causal contribution.
Which brings me back to my stablecoin audit. The discrepancy I found in 2022 was not the cause of the eventual collapse. It was a symptom. But the symptom was the only thing an investor could observe, and the only thing a regulator could act on. In an opaque system, symptoms become causes, because they are the only handles anyone can grab. Decentralized compute is, right now, the most visible handle on AI capability diffusion. It will be grabbed.
The timing is unknowable. The direction is not. And for anyone holding long-duration positions in the AI-adjacent crypto sector, the relevant question is not whether this happens but whether the current price reflects the probability that it does. It does not.
The Macro Overlay: Liquidity, Capex, and the Bear Market
I want to zoom out, because this is where my work usually lives: at the intersection of policy signals and liquidity cycles.
In 2025, amid the institutional entry wave, I built a quantitative framework linking BlackRock's spot Bitcoin ETF inflows to changes in global M2 money supply. I analyzed eighteen months of daily data and identified a fourteen-day lag between liquidity injection and price appreciation, after adjusting for regulatory hedging behavior. The finding was not that ETF flows cause price. It was that ETF flows are a transmission mechanism for macro liquidity, and that the mechanism has a measurable delay.
That framework has an uncomfortable implication in the current cycle. The AI capital-expenditure boom is absorbing global liquidity at a scale that has no precedent in the post-2008 era. Data centers, power contracts, advanced packaging, and the semiconductor supply chain are collectively pre-committing marginal capital that, in a previous cycle, would have been available for speculative assets. The AI buildout is not just a competitor for investor attention. It is a competitor for the physical and financial resources that underwrite the entire risk curve.
In a bear market, this matters more than it would in a bull market, because the marginal buyer is the only buyer. When liquidity is abundant, every asset finds a bid. When liquidity is scarce, capital flows to the assets with the strongest claim on future cash flows, and speculative infrastructure loses the auction. The decentralized compute sector is particularly exposed, because its valuation rests on a narrative โ AI plus crypto โ that requires a benign AI regulatory environment to compound. A voluntary framework is not benign. It is unstable, and instability is precisely what long-duration capital refuses to underwrite.
Let me put the bear-market question in the terms that actually help readers survive it. Which protocols are bleeding, and why?
The AI-adjacent crypto sector โ decentralized compute, data markets, agent frameworks โ is bleeding for two reasons. First, its token emissions were calibrated to a bull-market demand curve that has not materialized. Second, its regulatory risk premium has, counterintuitively, increased rather than decreased. Deregulation sounds bullish. But the specific form of deregulation now on the table โ voluntary, national-security-framed, undefined โ creates a policy environment in which the tail risk of a sudden reversal is larger, not smaller. The market is pricing less regulation as less risk. The correct reading is less regulation now, more regulation later, with no warning.
That is the liquidity trap in formation. Not a trap of low rates, but a trap of deferred rules. The cost of capital looks low because the compliance cost looks low. But the true cost of capital includes the probability-weighted cost of the rule that has not been written yet, and that cost is being systematically excluded from current valuations.
To make this concrete, I separate the sector into three tiers by regulatory exposure. The first tier is centralized compute brokers and data providers that interface with regulated enterprises; these carry moderate exposure because they already operate under contracts that impose de facto obligations. The second tier is decentralized compute networks with open, permissionless supply; these carry the highest exposure, because they are the visible handle and the plausible scapegoat. The third tier is agent frameworks and autonomous payment rails; these carry the most concentrated tail risk, because they are the newest and the least understood by regulators, and novelty amplifies the severity of a first enforcement action.
The portfolio implication in a bear market is not to abandon the sector. It is to stop treating the three tiers as a single beta. The market has been pricing AI-crypto as one trade. It is three trades with three different regulatory durations. In a survival regime, the correct move is to underweight the tier with the highest visible handle and the weakest legal footing โ which is the permissionless compute layer โ regardless of how compelling its technical roadmap looks.
Programmable Money Meets Programmable Intelligence
I want to close the analytical section with an observation from my day job, because I think it is the bridge between AI governance and monetary governance that almost nobody is building.
In 2024, I spent six months inside the technical architecture of Vietnam's digital dong pilot. I mapped settlement latency, privacy leakage, and key management. I found over two hundred inefficiencies. But the finding that stayed with me was not an inefficiency. It was a structural resemblance.
A central bank digital currency and a controlled frontier model are the same object viewed from different ends. Both are programmable systems that execute policy. Both concentrate control over a general-purpose capability in a small number of institutions. Both claim to serve the public while being architected to serve the issuer. And both rely on the same substrate: compute, cryptography, and identity.
The convergence is not metaphorical. A CBDC needs programmable money. A controlled AI system needs programmable capability. Put them together and you get a settlement layer where both money and intelligence are programmable by the same authority, at the same time, with the same identity primitives. The voluntary framework debate is, at bottom, a debate about who gets to write the policies that these systems execute.
This is why the model control risk language unsettles me more than the cybersecurity or biosecurity language. Cybersecurity and biosecurity are defensive frames; everyone agrees that attacks are bad. Model control is a governance frame. It is about who decides what a model may do and for whom. When that frame is undefined and voluntary, the eventual definition will be written by the institution that has the most to gain from control โ and that institution will not be a decentralized compute network.
There is an irony here that the crypto industry should sit with. The industry's foundational promise was to separate money from the state. The most sophisticated technology of the next decade โ frontier AI โ is being governed through a framework that will, by design, keep intelligence and money inside the same institutional perimeter. If CBDCs and controlled models converge on shared identity and shared settlement, then the crypto ecosystem's most valuable asset is not its tokens. It is its position outside that perimeter. And that position is exactly what a voluntary AI framework, followed by a reactive enforcement regime, is most likely to eliminate.
The ledger does not sleep, it only waits. Every transaction, every inference, every policy execution is recorded. The question is not whether a record exists. The question is who holds the pen.
The Contrarian Angle: The Decoupling Thesis Is Backwards
The consensus in crypto right now is that AI regulation and crypto regulation are separate tracks, and that the AI governance pivot is, at worst, neutral for crypto and, at best, a bullish signal for the AI-crypto narrative. I think the consensus has the sign reversed.
The prevailing view treats AI deregulation as a risk-on signal: lighter rules, cheaper capital, more innovation, and a rising tide that lifts the AI-token sector. My reading is that voluntary AI governance and permissive crypto posture are not two policies. They are one policy โ the deferral of rulemaking โ and deferral is not deregulation. It is the accumulation of unpriced liability.
When the first salient AI incident occurs, the response will not be targeted. It will be categorical, because categorical responses are the only ones available when the underlying categories were never defined. And because AI and crypto now share a compute substrate, a categorical AI response will reach crypto infrastructure whether or not crypto was causally involved. The decoupling thesis assumes two systems. There is one.
The deeper contrarian point is about what voluntary actually optimizes for. It does not optimize for safety, and it does not optimize for innovation. It optimizes for political optionality โ the ability to claim credit for light-touch governance today and to blame industry for any failure tomorrow. Optionality is valuable to the option holder and costly to everyone else. The market is not pricing that transfer of value. In a bear market, where every basis point of risk premium matters, that mispricing is the difference between surviving the cycle and being carried out of it.
There is one more inversion worth naming. The crypto industry has spent years arguing that it should be regulated like software, not like finance. The AI governance pivot suggests that software itself is now being regulated like finance โ through national-security framing, capital thresholds, and ex post interpretation. If that is the direction of travel, then crypto's long-running plea for a lighter touch is not a winning strategy. It is a request to be treated as an exception to a rule that is being extended over everything. Exceptions do not survive first incidents.
Takeaway: Position for the Rule That Has Not Been Written
We are in a bear market, and the only question that matters is survival. The AI governance pivot is not a crypto story on its face, but it is a crypto story underneath, because compute is the shared substrate and voluntary frameworks are the shared failure mode. The protocols most exposed are the ones whose valuations depend on a stable AI regulatory environment โ decentralized compute, data markets, and agent infrastructure โ because stability is exactly what a voluntary regime does not provide.
The rational position is not to short the narrative. It is to stop paying for the appearance of safety and to start pricing the cost of the rule that has not been written yet. That means treating the AI-crypto complex as three separate regulatory durations rather than one trade, and underweighting the tier with the most visible handle and the weakest legal footing. The ledger does not sleep. Neither does the liability. Watch the compute layer, because that is where the cage gets built โ and the bird that flies out of it will not be the one anyone designed for.
