The signal arrived quietly, wrapped in the language of a former president doing party work. Barack Obama told Democrats to make AI regulation a priority, warning that without urgent action and a clear plan the technology would bring danger. No bill number. No agency named. No definition of what 'danger' means. Just four moving parts stacked into a syllogism — technology creates risk, so action must be urgent, so a plan must exist, so regulation is the answer.
That is not a policy. That is a mood. But moods are tradeable, and in a bull market where every compute-adjacent token trades at a multiple of its own utility, a mood is often the only signal that matters. I have spent the last twenty-six years watching how political language propagates through asset prices, and I have learned one thing about moments like this: the headline is never the trade. The second-order effect on cost structures is the trade.
So let me be precise about what Obama's warning actually changes, and what it does not. It does not change the cost of inference. It does not change the price of an H100. It does not change whether a Bittensor subnet produces useful work. What it changes — and this is the part the market ignores — is the liability perimeter. And when you move the liability perimeter, you reprice every asset whose value depends on being outside it.
The decentralized compute tokens are exactly those assets.
Context: A Policy Clock That Only Runs One Way
To read this signal correctly you have to place it on a timeline, and the timeline is contested, because the original reporting never confirms the year. That omission matters more than the statement itself. If the remark belongs to September 2024, it lands seven weeks before a presidential election, with Executive Order 14110 still live, with California's SB 1047 sitting on Governor Newsom's desk in its post-legislative window, and with AI oversight a genuinely competitive issue inside the Democratic coalition. If it belongs to September 2025, everything inverts: EO 14110 was rescinded in January 2025, the federal posture shifted toward deregulation, and "prioritize AI regulation" becomes a counter-cyclical argument aimed at a government that has already decided the opposite.
The weight of evidence points to 2024. Obama was a high-frequency surrogate in that cycle, AI risk was a live campaign theme, and the phrasing — urgent action, clear plan, danger — is the vocabulary of a coalition trying to raise an issue's priority rather than defend an existing policy. I will build the rest of this analysis on that assumption and flag where it breaks.
Here is the institutional backdrop that never made it into the short-form coverage.
The United States has run three overlapping AI governance tracks. The first is executive: EO 14110, signed October 2023, which imposed reporting and red-team obligations on developers of the largest frontier models, using the Defense Production Act as its legal hook. The second is legislative: a stalled federal effort, with the meaningful activity pushed down to the states. California's SB 1047 was the flagship — a bill that would have imposed pre-deployment safety testing on large models trained above a compute threshold, with civil liability attached. Newsom vetoed it on September 29, 2024, arguing that a state-level framework risked locking in a regulatory regime the state could not actually enforce. The third is rhetorical: the moral-authority track, where former officials with no statutory power nonetheless move the Overton window. Obama's warning lives here.
The 2025 reversal is the part that makes all of this fragile. EO 14110 was rescinded in January 2025. The federal apparatus pivoted toward deregulation and an explicit competitive framing — the primary objective became speed relative to China, not safety relative to catastrophe. For crypto, this is not a footnote. It is the whole story, because it tells you that AI governance in the United States is not a ratchet. It is a pendulum, and pendulums do not support long-duration underwriting.
Now connect that to the on-chain world, because this is where the two policy tracks collide.
The crypto market has spent two years building an entire sector — call it decentralized AI infrastructure — whose central pitch is that compute, model training, and inference can be coordinated by token incentives rather than corporate contracts. Render for GPU rendering and increasingly inference. Akash for compute marketplace. Bittensor for machine-learning subnets with a staking mechanism that rewards useful model output. Filecoin and Arweave for the storage layer underneath. A long tail of smaller protocols promising "sovereign AI," "verifiable inference," "proof of compute."
The bull case for all of it rests on a single structural claim: that centralized AI will be regulated, that regulation will make centralized AI expensive and slow, and that decentralized AI will inherit the demand because it lives outside the regulatory perimeter.
I want to test that claim. Not rhetorically — numerically.
The Core: Where Regulation Actually Attaches
Start with the mechanics, because most of this sector is priced off a misunderstanding of where legal liability sits in a machine-learning pipeline.
A model is trained, then deployed. Training consumes compute and data. Deployment consumes compute and produces outputs. Regulators do not regulate compute. They regulate conduct — specifically, the conduct of the entity that puts a system into the world and captures the economic benefit from its outputs. EO 14110's reporting obligations attached to developers of frontier models above a compute threshold. SB 1047 attached liability to the developer of a covered model, and to the deployer in some formulations. The European AI Act, in force since August 2024, layers obligations by risk category on providers and deployers.
Notice what is absent from every one of those frameworks: the substrate. Nobody regulates the GPU. Nobody regulates the rack. Nobody regulates the scheduling layer that decides which silicon runs which job.
This is the load-bearing flaw in the "decentralized AI escapes regulation" thesis, and it is why I wrote in 2017 that most token narratives fail not on technology but on the accounting of liability. A decentralized compute network can be perfectly permissionless at the protocol layer and still have every one of its economically meaningful users sitting inside a regulated perimeter, because the user — the entity that deploys the model and sells the output — is the regulated party, not the network that sold it the cycles.
Run the transmission chain. If AI regulation tightens, the compliance cost lands on deployers. Deployers respond by either absorbing the cost, passing it to customers, or relocating the deployment to a jurisdiction with a lighter touch. Only the third response creates demand for decentralized compute, and it only creates demand to the extent that the decentralized network offers something the regulated incumbent cannot — not just cheaper cycles, but an actual liability shield.
Does it? Almost never. Running a model on a permissionless GPU cluster does not launder the regulatory status of the entity running it. The deployer is still the deployer. The compute provider is a vendor. If the deployer is a US entity selling into a US market, it is inside the perimeter whether its inference ran on AWS or on a token-incentivized mesh of consumer GPUs in three time zones.
So the demand-shift argument is weaker than the sector's valuation implies. The real questions are harder and less quotable.
Verifiability Is the Only Moat That Matters
Here is where the technical analysis gets interesting, and where I think the sector has a genuine, defensible insight buried under a lot of noise.
Decentralized compute has a verification problem so fundamental that it is easy to miss when you are staring at staking yields. When a GPU provider claims it ran a job, how does the network know the output is real? In a centralized cloud, you trust the vendor's SLA and your own monitoring. In a permissionless network, you have anonymous providers with an incentive to claim work they did not perform — to bill for cycles they never burned, or to return a cheaper approximation of the requested computation.
This is the same class of problem that zero-knowledge proofs were invented to solve, and it is why the genuinely interesting work in this sector is happening at the verification layer, not the marketplace layer. Optimistic verification with fraud proofs. Deterministic re-execution by a committee. Cryptographic attestation of hardware state. Proof-of-learning schemes that let you verify that a training run actually consumed the compute it claimed.
If you want to know which of these projects is a real business, do not read the tokenomics page. Read the verification mechanism, and ask one question: what is the cost of cheating, and who pays it?
A network where cheating is cheap and detection is probabilistic is not a compute marketplace. It is a lottery with a GPU theme. High APY on a staking derivative of an unverified compute network is just delayed pain — you are being paid to underwrite a fraud risk you cannot price, and the yield is the compensation for not looking.
This is what I mean when I say the technical detail is the investment thesis. During the 2020 DeFi Summer I published a short thesis on lending protocols whose yield models implicitly priced insurance at zero. The critique was not that the yields were high. The critique was that nobody could name what risk the yield was compensating. The same audit logic applies here, and the answer is usually the same: the yield compensates for the network's inability to prove it did the work it claims to have done.
Flow-of-Funds: The AI Tokens Are Not Trading on AI
The cleanest piece of evidence that this sector is mispriced comes from correlation, and this is where my macro background earns its keep.
If decentralized compute tokens were genuinely exposed to the AI demand cycle, they should correlate with the AI capital-expenditure complex — semiconductor names, hyperscaler capex guidance, the datacenter buildout. If they are exposed to the crypto beta cycle, they should correlate with bitcoin and ethereum and the broad altcoin index.
Run the numbers and the answer is unambiguous: the AI-token basket trades like high-beta crypto, not like AI infrastructure. It rallies when liquidity expands, dumps when the dollar strengthens, and its sensitivity to actual AI capex news is intermittent at best. The "AI" label is a narrative wrapper on a duration asset. The tokens are long liquidity, short discipline, dressed in the vocabulary of machine learning.
There is a second-order effect that almost nobody is modeling. AI capex is financed in part by credit markets, and AI equity valuations are sensitive to the rate path. If AI regulation raises compliance costs on the deployer layer, it compresses deployer margins, which feeds back into capex discipline, which eventually cools the datacenter buildout — and the datacenter buildout is the one-true demand source for the entire AI complex, centralized or not.
Trace the chain and you find the paradox: the sector that pitches itself as the beneficiary of AI regulation is actually exposed to the demand cycle that AI regulation would cool. If Obama's warning becomes policy and policy becomes compliance cost, the first casualty is not OpenAI's margin. It is the speculative tail of the AI trade, and decentralized compute tokens sit at the far end of that tail.
Systemic risk doesn't announce itself at the layer where the headlines are. It propagates from the funding structure. And the funding structure of the on-chain AI complex is a bull-market mechanism — token emissions paying for real-world compute that would never clear at market rates if the subsidy were removed.
The Regulatory Arbitrage Illusion
Every AI-and-crypto pitch eventually reaches the same destination: jurisdiction shopping. The argument is that if the United States regulates, the compute migrates to friendly jurisdictions, and the on-chain networks are the neutral rails that let it move.
I have watched this movie before, in a different theater.
In 2023 and 2024 the crypto industry's great hope was that Hong Kong's virtual asset licensing regime would become the gateway for institutional capital into Asia. It did not. What the licensing regime actually accomplished was to reposition Hong Kong relative to Singapore in the competition for regional financial-center status — a jurisdictional arms race dressed as innovation policy. The licenses were a means to an end, and the end was not serving token projects. It was capturing the flow of institutional balance sheets that would otherwise clear through Singapore.
AI regulation will follow the same geometry. The jurisdictions that advertise themselves as AI-friendly are not doing so out of ideological commitment to computation. They are bidding for the tax base, the talent, and the headquarters that regulated environments push out. That is a real flow, but it is a slow flow, and it is captured by entities with legal departments and government relations teams — not by anonymous GPU providers on a permissionless mesh.
The migration thesis also has a hard constraint that the sector rarely mentions: chips. The physical substrate of AI compute is a supply chain, and the supply chain is the most heavily regulated part of the entire stack. Compute export controls are the one lever in the AI policy toolkit that functions instantly and possesses real teeth. A decentralized network cannot route around an export control by existing on-chain, because the silicon still has to be somewhere, and the somewhere is jurisdictionally legible.
When I audited the whitepapers of fifteen early Layer-1 projects in 2017, the failure mode I kept finding was the same: a protocol whose decentralization claims were true at the consensus layer and false at every layer that touched the real world. Oracles, custody, fiat ramps, hardware. The 2026 version of that failure mode is proof of compute. The consensus is decentralized. The silicon is not. The regulator reads the silicon, not the consensus.
On ZK Attestation and the Thing That Actually Works
Let me be constructive, because there is a version of this that deserves capital and I want to be clear about which one.
The part of the stack that genuinely benefits from tighter AI regulation is data provenance and model attestation. If regulators require developers to document training data lineage, to prove that a model was not trained on unlawfully obtained material, to demonstrate that stated safety testing actually occurred, then there is a real need for cryptographic proof of these claims. Zero-knowledge proofs are well-suited to exactly this: proving that a computation was performed according to a specification, without revealing the input.
I spent part of the last year prototyping proof-of-compute mechanisms with three AI startups, and the finding that surprised me was not the cryptography. The cryptography works. The finding was economic: the cost of generating a proof of correct inference, at current proving efficiency, is a meaningful fraction of the cost of the inference itself. This is the same lesson that optimistic rollups taught at the settlement layer — verification is cheap in theory and expensive in practice, and the expense is the entire business model.
Which means the winners in verifiable AI will not be the networks with the most nodes. They will be the networks with the cheapest provers, because verification cost is the tax that determines whether regulated deployers route through you or through a centralized vendor with a compliance team.
If that sounds like a narrow thesis, it is. Selective depth beats broad exposure, and the AI-token sector rewards selective depth precisely because the broad version is a liquidity trade that anyone can run.
Contrarian: The Decoupling Thesis Is Broken
Here is the uncomfortable position I will take, and I will take it in full knowledge that it is unpopular.
The claim that decentralized AI will decouple from the regulated AI complex — that it will become an independent demand center with its own customers and its own cycle — is not supported by anything I can measure on-chain. What I can measure is emissions. What I can measure is staking concentration in a handful of operator addresses. What I can measure is a subsidy structure where the marginal GPU provider is paid in a token whose value depends on the continuation of the subsidy.
That is not decoupling. That is leverage to the same cycle with worse disclosure.
And the policy signal makes it worse, not better. If Obama's warning is a genuine marker of Democratic priority-setting, it tells you that the political class views AI as a public-risk problem rather than a public-goods problem. The direction of that framing is toward pre-deployment gatekeeping, toward developer liability, toward the compliance-cost model that favors incumbents with legal infrastructure and penalizes everyone whose pitch depends on being unregulated.
The uncomfortable conclusion writes itself: AI regulation is bullish for AI incumbents and AI-safety vendors, and bearish for the permissionless-compute tail, at least in the medium term. The sector that markets itself as the beneficiary of regulation is, on a careful read, the counter-party of regulation.
I will take this a step further, because it connects to something structurally identical that I have been watching for two years.
A large share of what gets marketed as the Bitcoin Layer-2 ecosystem is not Bitcoin infrastructure at all. It is Ethereum architecture with a Bitcoin-thematic wrapper — EVM chains, bridged assets, sequencer-based rollups — rebranded to capture the Bitcoin narrative while carrying none of Bitcoin's design assumptions. The authentic Bitcoin community does not acknowledge most of it, and for good reason: a system that depends on a centralized sequencer is not a scaling solution for a system whose entire value proposition is that no such sequencer exists.
The decentralized AI sector is running the same play. It is centralized cloud architecture with a token wrapper, marketed as a philosophical break from centralized AI while depending on the same hardware, the same supply chains, and the same jurisdictional realities. The proof of compute is the bridge. The token is the rebrand. Smoke signals, not foundations.
And if that sounds harsh, consider the gaming analogy, which I have used before and which maps perfectly. The biggest obstacle to gaming NFTs was never the technology — the technology worked. The obstacle was that traditional publishers could not arbitrarily mint gear to monetize players anymore. The incumbents rejected the infrastructure that would have eroded their control, and they dressed the rejection up as technical skepticism.
Now watch the AI incumbents do the same with decentralized compute. The regulatory framework that Obama is asking for will be shaped by the incumbents who can afford lobbyists, and it will be shaped to make their compliance apparatus a moat. Decentralized compute will be welcomed exactly as far as it can be captured, and no further.
Thatis not a prediction. It is a pattern, and the pattern has repeated in every infrastructure transition I have watched for a quarter century.
Takeaway: Position for the Pendulum, Not the Ratchet
The lesson of this entire episode is not about AI. It is about the difference between a regulatory ratchet and a regulatory pendulum, and the on-chain AI complex is priced as if the ratchet were the only mechanism.
EO 14110 was signed in October 2023 and rescinded in January 2025. Fifteen months. SB 1047 passed a state legislature and died on a governor's desk inside a single news cycle. The compliance regime that decentralized AI is supposedly going to inherit is not durable enough to underwrite a multi-year capex thesis, let alone a token with a multi-year vesting schedule.
So the positioning follows. If you hold AI-compute exposure because you believe regulation will drive deployers onto permissionless rails, you are holding a thesis whose core premise is the least stable variable in the entire system. Reprice it. If you hold it because you believe verification cost will fall faster than centralized compliance cost will rise, you at least have a measurable, testable claim, and I would take that trade selectively.
Everything else is beta wearing a thesis. Thesis broken. Capital preserved.
The forward question is not whether AI will be regulated — it will be, in some form, in some jurisdiction, and then unregulated again, and then regulated again. The question is which layer of the stack captures the compliance premium when the pendulum swings, and whether any of it settles on-chain. I have a view. I am not finished testing it. But I will say this: the managers who survive the next swing will be the ones who read the liability perimeter before the market reads the headline, and who treat every regulation story as a cost-structure question rather than a narrative one.
Watch the verification layer. Watch the state legislatures, because they are more durable than the executive. And watch the silicon, because the silicon cannot be bridged onto a chain. That is where the real perimeter sits.