On a Tuesday morning in late 2025, a former commissioner of the U.S. Federal Trade Commission named Alvaro Bedoya posted an argument on X that the AI policy world is still arguing about. Its logic was almost embarrassingly plain. If you genuinely believed your product was going to kill everyone, you would not petition for a regulatory carve-out, you would shut the servers, lock the door, and call the police.
I have spent eleven years watching that same rhetorical move play out in crypto, and something in my chest tightened when I read it. For two years, the loudest safety language in technology has come from the labs with the most to lose from competition, and the quietest has come from the open-weight builders who actually ship the cheaper models. That asymmetry is not an accident. It is a pattern I first learned to read during a very different crisis, and it is the reason I want to walk us through this debate the way we would audit any protocol's code, slowly, and with our eyes on incentives rather than marketing.
The story isn't in the token, it's in the trust. Every era of this industry has a moment where safety and compliance quietly stop being values and start being moats. For frontier AI, this is that moment.
To be fair to anyone who has not been following the policy beat, let me set the table honestly. Bedoya is not a neutral referee. He is a former Democratic commissioner at the FTC, known during his tenure as a privacy and antitrust hawk, and he was removed from the agency in a dispute that many observers on his side describe as political. He now stands in a natural position to criticize both the current administration and the largest technology firms, and that position carries genuine public value even as it carries obvious motive. I flag this because my training in cybersecurity taught me that the first question is never whether a source is right, but what the source has to gain.
The reporting that carried his comments came, notably, from a crypto outlet, which is itself a signal. A publication built for digital-asset readers spent its column inches on AI antitrust, and that cross-domain instinct matters, because the policy vocabulary of Web3 and AI is beginning to rhyme in ways the mainstream has not yet named. The cast of characters is familiar to anyone who has watched the last two years: Sam Altman of OpenAI, Dario Amodei of Anthropic, Elon Musk of xAI, and, hauntingly, Greg Jensen of Bridgewater, who offered the coldest sentence of the cycle, that nothing would change in AI governance until someone dies.
Bedoya's central claim is not that AI risk is fake. It is that the fear of AI risk, when amplified by the firms best positioned to profit from the resulting rules, becomes a moat. And here our history should make us sit up, because we have seen this movie. In 2019, the small circle of crypto exchanges that had the compliance departments and legal teams to absorb the FATF travel rule welcomed it publicly and quietly watched smaller competitors drown in the paperwork. When Europe's MiCA framework arrived, the briefs that praised its clarity most enthusiastically came disproportionately from desks large enough to afford a full-time regulatory staff. The regulation was real. So was the moat. The honest founder's ask and the incumbent's ask can be spelled identically and mean entirely different things.
Let me start with the economic engine, because that is where the argument actually lives. Bedoya's sharpest observation is that the largest AI labs are burning money they have not earned. That is not a rhetorical flourish, it is a description of the business model. OpenAI and Anthropic operate at enormous losses, funded by successive rounds of venture capital rather than operating cash flow. When a company's valuation depends on future profitability rather than present revenue, its most precious asset is not its model weights, it is the narrative that it will eventually be able to charge for them. And a narrative about future pricing power is, structurally, a narrative about the absence of cheaper substitutes.
That is where open-weight models enter the story like a cold draft under the door. A model you can download, fine-tune, and run on your own hardware cannot be priced by an API meter. It strips the frontier labs of their pricing power precisely where it hurts most, at the enterprise tier, where margins are widest and buyers are most price-sensitive. When Bedoya points to the price advantage of open-weight models, he is naming a competitive fact that the safety discourse has been quietly organized to bury. The cheapest path to capable AI is, for a growing slice of enterprises, the path that requires no permission from a frontier lab at all.
Now bring the two facts together and the incentive becomes legible. If an industry-wide consistent safety requirement becomes the standard, it raises the compliance cost for everyone, but the cost is borne in proportion to how thin the operator's margin already is. The lab with a dozen lawyers absorbs the new rule as paperwork. The open-weight startup with two engineers absorbs it as an existential threat. This is the mechanism economists call regulatory capture, and it is not a conspiracy so much as a gravity. Pharma lived with it for decades. Finance has never escaped it. Compliance is not the enemy of competition. Often it is the most elegant form of it. The story isn't in the token, it's in the trust that a mandate claims to protect.
I felt this gravity firsthand in 2024, when I ran a Human-Centric Crypto workshop series for a mid-sized Viennese fintech and its traditional finance clients. We onboarded two hundred institutional clients, and the feedback that reshaped our internal risk models had nothing to do with deregulation and everything to do with clarity. Conservative investors did not want fewer rules. They wanted rules they could trust, and crucially, they wanted to know who wrote them. The lesson stayed with me: the moment a standard becomes mandatory, whoever authored it has effectively chosen the winners. That is not a side effect of safety governance. For the incumbent, it is the point.
This is also why the antitrust frame matters, and why it feels new. For most of the last decade, AI competition was treated as a technology-and-safety question, handled by ethics boards and internal charters. Bedoya is relocating it to the ground of the FTC and the Department of Justice, invoking the joint policy statement on competition, and forcing the safety conversation to answer to the older, harder question of market structure. When the vocabulary shifts from alignment to moat, from responsible deployment to barrier to entry, the game changes. The labs can litigate an alignment standard. They cannot so easily litigate a market-share question.
And then there is the pacing plan, Altman's reported three-step framework, whose middle step is safety coordination. Read the word slow and ask who benefits. The leader in any race has a structural interest in collective deceleration. The challenger has a structural interest in acceleration. Altman's own framing betrays the game. On one hand he insists that nothing needs to wait for a waiver or for legislation to begin. On the other, he still wants a federal framework and international coordination. That is the posture of a player who wants to look unconstrained while retaining the power to write the constraint. Whoever defines the standard defines the finish line, and whoever defines the finish line tends to be standing near it.
Bedoya makes one more cut that deserves our attention, because it is surgically precise. He separates the motives of employees from the motives of executives. Individual researchers, he grants, may be sincere and simply mistaken, true believers in a risk model that has never been reduced to a reproducible equation. But the action of a chief executive seeking an exemption deserves what he calls deep suspicion. This distinction maps onto something I observed during my 2026 work on The Empathy Algorithm, the project where I studied how AI-driven DAOs managed their communities. The researchers who built those systems genuinely believed in their risk models. The executives who shipped them answered to a different god, the next financing round. Motive is not a moral accusation. It is a load-bearing assumption in any incentive analysis.
I want to be careful here, because the crypto reader is tempted to turn this into a clean morality play of villains and victims. It is not. The competitive landscape is more tangled than the headline suggests. Anthropic requested an exemption that even the reporting describes as narrowly scoped, a safety dialogue rather than an outright cartel charter. The gap between a clarification request, which is routine legal hygiene, and a waiver seeking exclusivity, which would justify the cartel charge, is exactly where the real evidence would live, and the reporting does not supply it. Meanwhile Google DeepMind, holder of the strongest cash position and the largest compute footprint, is conspicuously silent, neither begging for relief nor being named. That silence is louder than any statement. The player who needs no moat is the one best positioned to profit while others build them. And the open-weight camp, Meta, DeepSeek, Alibaba's Qwen line, Mistral, is spoken for throughout this debate rather than speaking. It is represented as a threat, not interviewed as a participant. A whole side of the market has been rendered an object rather than a voice.
If we zoom out, the structure of the contest becomes almost diagrammatic. On one side sit the closed frontier labs, capital-rich, cash-poor, and racing to convert a temporary capability lead into a durable rule. On the other side sits the open ecosystem, capital-light, cost-advantaged, and structurally allergic to any standard that taxes distribution. Between them hover the regulators, newly empowered, unsure whether they are referees or prize. And beneath all of it, the quiet fact that the two camps are fighting over narrative as much as product. In a market where capability converges within months, the durable asymmetry is not who builds the best model. It is who writes the rule that decides what best is allowed to mean.
The legislative signals this debate has pulled into the open deserve their own note. A superintelligence ban is, as written, close to unenforceable. You cannot regulate an artifact that does not yet exist and has no agreed definition. But unenforceable does not mean inconsequential. It raises the political risk premium on the entire sector, which is exactly how a posture becomes a price. And notice who is pushing both the restriction and the loosening: the same narrow band of actors, arguing for a boundary at one table and a carve-out at another. Both moves serve the same purpose, to set the perimeter of the field while the leaders are already standing inside it.
There is a market dimension that the policy framing tends to hide, and it is the one my institutional clients cared about most. When a company's valuation rests on future profitability, and that profitability depends on the ability to resist a cheaper substitute, then the regulatory moat is not a political footnote, it is a valuation input. Approval or rejection of a safety exemption becomes, quite literally, a repricing event. This is the part of the story a crypto audience should find uncannily familiar, because we watched the same mechanism decide which chains survived the last cycle. The tokens that endured were rarely the most elegant. They were the ones that captured a standard, an integration, a listing, a compliance posture, and then charged the rest of the market for access to it. Regulatory legitimacy is the most undervalued asset class in technology, and it is currently being minted by whoever writes the first draft of the safety standard.
And this is where the two industries I live between finally touch. I have argued for years that the vocabulary of Web3, decentralization, open weights, permissionless access, credible neutrality, is migrating into AI faster than the AI establishment wants to admit. The open-weight debate is, at bottom, the sovereignty debate we have had in crypto since 2017, wearing an inference-cost mask. Who gets to run the model, and on whose terms, is the same question as who gets to run the node, and on whose terms. The people now writing AI safety standards are, in many cases, the same archetype that wrote crypto's compliance frameworks: well-intentioned institutionalists who believe that legitimacy flows from the center outward. My work with AI-agent DAOs taught me the opposite lesson. A standard that no one below the leadership tier helped write is a standard no one below the leadership tier will defend. That is not ideology. That is what I saw when automated governance lost loyalty the moment it lost story.
There is one more thread, thinner but worth holding. The single largest line item inside the burning money Bedoya describes is compute, the GPU clusters and inference costs that turn venture capital into electricity. If financing tightens, that burn reaches the hardware supply chain within two quarters. A regulatory moat that protects the incumbents concentrates compute demand into fewer hands and flatters the cloud providers. A regulatory moat that fails scatters it. Either way, the compute layer is where the policy debate becomes a physical event, and it is where I would look first for confirmation that the narrative has turned real.
Here is where I part company with both camps, and it is the least comfortable part of this piece to write. The doomers and their critics are, in the end, fighting strawmen made of one another. Bedoya's rebuttal leans on historical induction. Humanity survived the ice age, the plague, the nuclear age, therefore it will survive superintelligence. That is a survivor's argument, and survivors are exactly the sample least equipped to estimate the odds of a novel catastrophe. If the Cuban missile crisis had resolved the other way, no one would be here to make the comparison. The nuclear analogy does not defang the risk. The doomsday clock and the thirteen days of October argue the opposite. His rhetoric is doing work his logic cannot support. That is worth naming, because I hold him to the same standard I hold the labs.
And the labs have their own evasion. The argument that safety requires us to build it first, so it may as well be us, was entirely absent from the reporting, though it is the strongest available answer to Bedoya's sincerity test. More damning still, neither side will state a probability. No frontier lab has published a falsifiable extinction estimate. No critic has answered what probability would justify slowing down. A debate with no numbers on either side is not a debate about risk. It is a debate about power, wearing risk as its coat. The silenced party is the public, neither the beneficiary of the fear nor a participant in the decision, only the audience for it. And I keep returning to the same uncomfortable symmetry: the loudest claim of care and the loudest claim of skepticism are both being made by people who will never have to live with the outcome they are arguing for.
So watch three signals, and watch them the way you would watch a token unlock. First, whether Anthropic's narrow exemption is actually filed, and how the FTC and the DOJ respond, a matter of weeks rather than quarters. Second, whether Altman's full pacing plan is published, and whether its middle step hardens into a mandatory standard. Third, whether any open-weight model closes the capability gap on the frontier's flagship tier, the quarter-by-quarter tell that the moat is either being built or being breached. If the compliance moat does form, it will not look like censorship. It will look like safety, it will be praised, and it will be the most expensive thing the market never priced. The story isn't in the token. It's in the trust, and lately, someone is trying very hard to own the word.