The Word "Pace" and the Document Nobody Has Read

David Sacks called the "Pace The Frontier" initiative regulatory capture. The economics say he's probably right. The politics say he's the wrong messenger. And the crypto industry already ran this exact experiment — and lost it.

I. An Artifact, and What It Hides

There is a verb buried inside the name of the initiative, and nobody in the coverage has stopped to read it.

Pace. Not block. Not freeze. Not ban. Pace. To pace something is to concede that it moves, and to insist only that it move at a velocity a human hand can still steer. It is the vocabulary of a marathon, not a wall. Whoever drafted that name understood something the people now shouting about it have forgotten: the fight was never about whether the frontier advances. It was always about who gets to hold the stopwatch.

Last week — or last month; the briefing handed to me carries no publication date, which is itself the loudest fact in this entire story — David Sacks, the venture capitalist turned White House AI and crypto policy lead, described that initiative as "regulatory capture." Four of the five data points in the summary I received are Sacks's own framing. There is no text of the initiative. No list of signatories. No response from whatever body drafted it. No venue, no transcript, no date.

Five hundred words of accusation, and not one page of the thing being accused.

I have spent enough years inside crypto's regulatory wars to recognize this silhouette immediately. During the long gray months after Terra, I ran a project I called the Post-Mortem Anthology — thirty protocols, fifty interviews, one obsessive question: when something collapses, what was the document that didn't exist? Almost every time the answer rhymed. The failure was rarely in the code. It was in the gap between the artifact everyone argued about and the artifact nobody had read.

So let us do the unfashionable thing. Let us start with the gap, and then work outward from it — tracing the ghost in the machine back to the people who built the machine, and the people who want to meter it.

II. Context: A One-Sentence Theory of Everything

Regulatory capture is not a conspiracy theory. It is an economics paper.

George Stigler published it in 1971, and the mechanism is almost boring in its elegance: industries do not merely endure regulation, they demand it, because a rule that imposes a fixed cost on every participant is a rule that hurts small participants most. A compliance department is a fixed cost. A legal retainer is a fixed cost. A forty-page model-evaluation dossier is a fixed cost. Fixed costs are the natural enemy of the entrant and the natural ally of the incumbent, because the incumbent amortizes them across a revenue base the entrant does not have.

That is the whole engine. Everything else — the lobbying, the revolving doors, the industry working groups that draft the standards they will later be measured against — is downstream of that single asymmetry.

Now apply it to frontier AI, where the fixed costs are not modest. A serious pre-deployment evaluation regime implies red-teaming, third-party audits, interpretability reports, incident disclosure obligations, compute reporting above a training threshold, and legal review of every one of those artifacts. For a laboratory with nine figures of annualized revenue and a standing safety team, this is a rounding error with a press release attached. For a twelve-person startup training a competitive model on rented GPUs, it is an existential tax paid in the currency startups have least of: time, and the attention of their founders.

The accusation Sacks is making, stripped of its framing, is that "Pace The Frontier" is precisely this engine — a rulebook that presents itself as a safety instrument while functioning as a moat. And on the economics alone, that is not a fringe claim. It is the default prediction of the literature.

But here is where the honest analyst has to stop and put the pencil down.

I cannot evaluate whether "Pace The Frontier" is a moat, because I have not read it. Neither, as far as the record shows, has anyone reporting on it.

The summary I was given contains five information points. Four of them are Sacks's opinion. That is not a news article. That is a press release with a byline, and the discipline of the thing requires me to say so before I spend two thousand words analyzing a document that exists, for the purposes of this piece, only as a name.

What I can do — and what I think is the actually valuable work here — is map the terrain. Because whatever "Pace The Frontier" says, it will be forced into one of a very small number of shapes. And the crypto industry has already lived through every one of them.

III. Core: The Metering Problem

Let me start with the technical substrate, because the policy debate floats above it like oil on water and nobody bothers to look down.

The regulatory architecture the United States built over the past several years — and, crucially, the architecture that a successor administration would either inherit, rewrite, or hollow out — is anchored on a number. Not a capability. Not a behavior. A number.

Training compute above a threshold, widely reported as on the order of 10^26 floating-point operations, triggers a set of reporting and evaluation obligations. Below it, you are, in the eyes of the framework, unremarkable.

Sit with how strange that is for a moment.

We do not regulate automobiles by the horsepower of the assembly line. We do not regulate pharmaceuticals by the tonnage of the reactor. We regulate outputs and behaviors — crash test results, clinical endpoints, adverse events. Yet for the most consequential technology of the decade, the trigger is a resource input, chosen because it is measurable.

And that choice is not neutral. It is a choice made because measurement is easy, and because the thing we actually care about — whether a model will do something catastrophic — is not measurable in advance at all. Frontier AI governance is not fundamentally a safety problem. It is a metering problem wearing a safety problem's coat.

Once you see that, everything downstream becomes legible.

If the trigger is compute, then the compliance burden falls on whoever buys the most compute. That is a surprisingly clean proxy for "the biggest labs," which means the thresholds get set — through ordinary, well-meaning technical consultation — at levels that the largest players clear comfortably and the smallest do not. Nobody has to intend capture. The capture is a byproduct of choosing the variable you can measure over the variable that matters.

This is the same mechanism I watched chew through crypto's own compliance layer, and I want to be precise about the parallel because it is not decorative.

The crypto rehearsal

Between 2021 and 2024, the digital-asset industry fought a war over whether tokens were securities, and the most revealing document of that entire period was never a statute. It was the informal guidance — the speeches, the enforcement actions, the "come in and register" line that everyone repeated and nobody could operationalize.

What actually happened was simpler than the discourse admitted. The fixed costs of registering, disclosing, auditing, and legal-reviewing a token offering were, for a mid-cap project, survivable. For a two-person team shipping an experimental protocol, they were fatal. The result was not that the industry complied or that it fled. The result was that the compliant tier consolidated, and the experimental tier went offshore or went quiet.

The industry's loudest voices spent three years arguing about the principle of the rule while the structure of the rule quietly did the sorting.

I have watched the same sorting happen to the layer-2 ecosystem, though for different reasons — dozens of rollups now chasing a user base that has not grown proportionally, liquidity sliced thinner and thinner across bridges that all claim to be the future and none of which can afford to be the present. Nobody regulated that into existence. But the fragmentation logic is identical: a structure that looks like expansion while functioning as division.

Apply the lens to "Pace The Frontier." Whatever its text says, the operative questions are brutally few:

• Where is the threshold, and is it denominated in a resource or a capability?
• Are open-weight models in scope, and if so, who is the regulated party — the trainer, the publisher, or the downloader?
• Are the obligations self-certified or third-party audited?
• Who sits on the body that writes the evaluation standards, and what is their commercial exposure to the answers?

Every one of those is a fixed cost. Every one of them sorts the field. And not one of them appears anywhere in the summary I was handed.

The FLOPs threshold is a bad instrument, and it is the only instrument we have

I want to be careful here, because it is easy to write a clever paragraph dunking on regulators for regulating the wrong variable. The dunk is cheap. The alternative is not.

If you want to govern a technology whose harms are rare, diffuse, and possibly civilizational, you have three options. You can regulate outputs — but the outputs you fear most are precisely the ones that have not happened yet. You can regulate actors — but that becomes a licensing regime, which is the most capture-prone instrument in the entire toolbox. Or you can regulate inputs, which are measurable, auditable, and already recorded by every cloud provider on earth.

We chose inputs. Not because it is correct. Because it is administrable.

And here is the second-order consequence nobody prices: once the input threshold exists, the input becomes the target. There is a documented, entirely rational incentive for laboratories to structure training runs so that they land on the favorable side of a reporting line. I have seen this pattern before, and not in AI. I saw it in the Ethereum ecosystem during the years when the only thing that mattered to a project's legitimacy was whether it could plausibly describe itself as decentralized enough — and so every project described itself as decentralized enough. The definition was the incentive. The incentive shaped the description. The description became the reality that regulators later tried to regulate.

Follow the thread from code to culture and you arrive somewhere uncomfortable: any threshold written into law immediately becomes a design constraint, and design constraints are always satisfied before they are debated.

IV. Core: The Two Camps, and Why Both Are Right

The AI policy debate has collapsed into two coalitions, and I have covered enough of crypto's factional wars — maximalists versus pragmatists, layer-1 tribalism, the endless blocksize nostalgia — to recognize the choreography.

On one side sits the safety coalition: researchers, parts of the major labs, a meaningful slice of the civil service, and the international coordination apparatus. Their position is that frontier capabilities carry tail risks that cannot be insured against after the fact, and therefore require prior restraint. Their preferred instruments are evaluations, thresholds, disclosure, and standing bodies with the authority to say not yet.

On the other side sits the acceleration coalition: growth-focused investors, open-source advocates, parts of the political right, and a substantial share of the developer community. Their position is that prior restraint is a moat in disguise, that the tail risks are speculative while the stagnation costs are certain, and that the correct instrument is post-hoc liability — punish the harm, protect the attempt. Sacks belongs here, and so does most of the capital that funds the industry's newest laboratories.

Here is what the framing hides.

Both coalitions are describing a real failure mode, and neither has an instrument that avoids it. The safety coalition is right that unchecked frontier deployment externalizes risk onto people who never consented to bear it. The acceleration coalition is right that prior-restraint regimes reliably become entry barriers, and that a moat justified by catastrophe is the most durable moat ever constructed, because opposing it looks like opposing safety.

That symmetry is the reason this fight is unresolvable by argument. It is not a disagreement about facts. It is a disagreement about which failure you would rather explain to your grandchildren.

And it is why Sacks's intervention is so effective, and so suspicious.

V. Contrarian: The Capture of Capture

Here is the part I have not seen anyone write, and it is the reason I wanted to write this piece at all.

The "regulatory capture" accusation has become the cheapest rhetorical move in technology policy. Deploy it and you do not have to argue about substance at all — you simply assert that your opponent's rulebook exists to serve your opponent's business, and the burden shifts to them to prove a negative. That is not analysis. That is a procedural kill shot, and it works because capture is real and therefore always plausible.

But there is a deeper problem.

If you define capture as "a rule that benefits incumbent firms," then opposition to regulation is itself a form of capture. Deregulation is a policy output. Policy outputs are produced by actors with interests. The venture capital community that funds frontier laboratories has a material interest in a world where those laboratories face fewer prior obligations, and it has spent real money to bring that world about. That is not a moral failing — it is the ordinary operation of a pluralistic system. But it means the accusation and the accused are standing on the same floor.

Which brings me to the thing I actually believe, and it took me a long time inside crypto to see it clearly.

The beneficiary of a compliance regime is almost never the firm you are looking at. It is the layer beneath it.

When the digital-asset industry finally got its regulatory frameworks — a comprehensive regime in Europe, a market-structure bill grinding through the United States, a proliferation of registration categories and disclosure templates — the firms that celebrated were not the exchanges. Or rather, they were, but they were not the ones whose margins changed shape.

The ones whose margins changed shape were the auditors, the blockchain-analytics vendors, the custody providers, the compliance-tooling companies, the law firms, and the consulting shops that sprang up to translate statute into workflow. An entire industry exists today that did not exist a decade ago, and its revenue is a direct function of the complexity of the rulebook. Complexity is its product. Simplicity is its extinction event.

I have sat in rooms where this was said out loud — not cynically, but as ordinary business planning: we need the guidance to be detailed enough that firms cannot self-serve.

That is where the real capture lives. Not in the model labs, which are too few and too visible to hide. In the compliance-industrial complex, which is diffuse, unglamorous, bipartisan, and structurally invested in a rulebook that never gets shorter.

Tracing the ghost in the machine leads you, again and again, to the invoicing layer.

This reframes "Pace The Frontier" entirely. The question is not does this initiative advantage the big labs. The question is who gets paid to prove compliance, and whether that class of vendors was in the room when the thresholds were chosen. That is an audit question, answerable with documents. It is also the question that neither camp wants asked, because the safety camp needs compliance infrastructure to exist in order for its rules to have teeth, and the acceleration camp needs to keep the accusation pointed at labs rather than at the machinery, because the machinery is boring and boring arguments do not fund think tanks.

VI. Core: The Coordination Trap, and the Variable Nobody Names

Buried in the summary is a second claim, easy to miss because it arrives attached to the first: that global cooperation on AI governance is not feasible.

This is a much bigger statement than the capture accusation, and it deserves its own examination.

The logic is straightforward. Regulatory coordination requires that the coordinating parties share a threat model, a legal vocabulary, an enforcement capacity, and — critically — a willingness to accept symmetric constraints. In AI, none of those conditions hold cleanly. The United States, the European Union, and China do not share a threat model. They do not share a legal vocabulary. And at least one of them would be accepting a permanent constraint in exchange for a temporary courtesy.

So the practical outcome is fragmentation: three regulatory universes, each internally coherent, mutually incompatible at the edges. And fragmentation has a cost structure that is well understood, because crypto has been paying it for years.

I have written extensively about what regulatory fragmentation does to a protocol that wants to serve a global user base, and the lesson generalizes. A firm operating under a single regime budgets compliance once. A firm operating under three regimes budgets compliance three times — and budgets the reconciliation work in between, which is where the real spending hides. Different disclosure templates. Different evaluation standards. Different thresholds denominated in different units. Different definitions of the same word.

Fragmentation does not merely raise costs. It raises costs regressively, because reconciliation work scales with headcount and headcount is what small firms do not have.

The Brussels effect — the theory that a large market's rules become the global default because multinationals would rather comply once — only functions when the rules are interoperable. When they contradict each other, you get the opposite: a permanent duplicative layer that punishes anyone without a dedicated policy team, which is to say, everyone except the largest.

Which means the acceleration camp's preferred outcome — no coordinated rulebook — produces exactly the consolidation it claims to oppose. Refusing to write a rule does not prevent a rule from existing. It guarantees that the rule is written by whoever has the most lawyers.

I keep coming back to the same image. In 2017 I ran a scrappy newsletter tracking the beacon chain, and I spent months convinced that the interesting action was in the specification — the elegant, contentious, technically gorgeous debate about how consensus should work. It was not. The interesting action was in the standards bodies, in the client teams, in the mundane question of who implemented which version first. The specification was the argument. The implementation was the outcome.

The same is true here. The public argument is about whether frontier AI should be paced. The outcome will be decided by whoever shows up to the working groups that translate pace into a number.

VII. Core: The China Variable

I want to name something the coverage conspicuously does not, and I will flag it plainly as inference rather than reporting.

When a policy figure argues that global AI governance cooperation is not feasible, there is almost always a second, unstated clause. The full sentence is rarely cooperation is impossible. It is usually cooperation with the parties we would have to cooperate with is impossible, and the constraints would bind us more than them.

In a competitive framing, coordination is not a neutral good. It is a commitment device that locks in the current distribution of capability. If you believe you are ahead, coordination locks in your lead. If you believe your lead is fragile, coordination is a leash handed to your competitor.

So the "cooperation is not feasible" claim is not primarily an empirical claim about institutions. It is a strategic claim about relative position, dressed as an empirical claim about institutions. That is not an accusation of bad faith — it is what every serious state does, and I would be a very naive analyst if I pretended otherwise after a decade of watching jurisdictions compete to attract digital-asset activity through the blunt instrument of regulatory generosity.

The consequence for anyone building in this space is concrete, and it is the same consequence crypto operators have lived with since 2017: your regulatory risk is not a single-country risk. It is a correlation risk. A policy shift in one jurisdiction moves the others, sometimes in the same direction and sometimes in the opposite, and the second-order effects arrive before the first-order ones have finished.

VIII. Core: The Beta Nobody Prices

Markets, meanwhile, are sideways, and I want to be precise about what that means for this story rather than gesturing at it.

In a consolidation regime, price stops transmitting information about fundamentals and starts transmitting information about positioning. That is what chop is for. It is the market's way of transferring conviction from the impatient to the patient at a rate slow enough that neither notices. Over the past seven days, the AI-adjacent token basket in my tracking sheet has behaved exactly as it always does when Washington clears its throat — a modest drawdown, a wide dispersion, and a volume profile that says nothing about the news and everything about the leverage that was sitting there waiting for a reason to unwind.

The numbers in my sheet are approximate. The structure is not.

What is actually happening is a repricing of regulatory beta, and almost nobody models it explicitly.

Consider how the transmission works. A credible move toward prior-restraint obligations raises the expected compliance cost of every firm in the sector, but it raises it unevenly. The firms with the strongest balance sheets absorb it and gain relative share. The firms with the weakest absorb it and vanish. Since public-market and token-market exposure is heavily weighted toward the survivors, a tightening regime can look like a sector rally in index terms while it is quietly executing a cull of everything below the index.

That is the mechanism, and it explains an observation that has confused a lot of people in this cycle: periods of intensifying regulatory pressure have repeatedly coincided with rising valuations for the largest names in a sector, even as the underlying industry contracts. The index is not a measurement of the industry. It is a measurement of the survivors of the industry, which is a different thing entirely.

I want to add a caution here that comes directly out of my own experience and is not comfortable to write. In 2017 I chased the beacon-chain narrative hard, launching three parallel threads and a newsletter on the strength of a story I believed in, and I prioritized the excitement of the frame over granular verification. The audience grew. Some of the calls aged badly. What I learned from that period is that *narrative momentum and analytical accuracy are correlated at low frequencies and negatively correlated at high frequencies* — the moment the story gets exciting is precisely the moment the error rate rises, because excitement is a signal that you have stopped checking.

The AI governance story is now extremely exciting. Adjust accordingly.

IX. Contrarian, Part Two: What the Crypto Comparison Actually Proves

I want to close the loop on the analogy, because I have been building it for two thousand words and I owe you the honest verdict rather than the flattering one.

The seductive version of this argument goes like this: crypto fought a regulatory war, the incumbents won, the small players got sorted out, and therefore the same will happen to AI. It is a clean story. It is also too clean, and the disanalogies matter more than the similarities.

First, crypto's regulatory fight was fought over a category question — is this a security? — which is a binary that resolves through legal classification. AI's fight is fought over a threshold question — at what capability level do obligations begin? — which is a continuum that resolves through measurement. Binary questions produce winners and losers. Continuum questions produce a permanently contested boundary, which is a worse equilibrium for everyone, because it never ends.

Second, crypto's incumbents had a structural problem: they could not stop new protocols from being deployed, only from being listed. The permissionlessness was real at the base layer and gated only at the edges. In AI, the gating happens much closer to the base — at the compute layer, the cloud layer, the model-release layer — and those layers are far more concentrated than any blockchain has ever been. The AI industry's chokepoints are physically better positioned for capture than crypto's ever were. Cloud compute is three or four firms. Chip supply is one or two. Model training at the frontier is maybe a dozen organizations. That is a much easier thing to regulate, and therefore a much easier thing to capture.

Third — and this is the part that makes me genuinely uneasy rather than analytically interested — crypto's failure mode when captured was that the industry built things that did not need the regime. It routed around it. Decentralized exchanges, self-custody, offshore venues, permissionless deployment. The regime won the argument and lost the market.

AI has no equivalent escape hatch. You cannot route around a compute chokepoint. You cannot self-custody a training cluster. There is no offshore jurisdiction with a critical mass of H100s and a credible legal shield. If the frontier gets captured, it stays captured, and the escape route is not a fork — it is a different country, or a different decade.

Which is why I find the "acceleration versus safety" framing so sterile. Both camps are arguing about which set of institutions should hold the stopwatch. Neither is asking the question that the crypto decade actually answered: what would it look like to build something that did not need to be paced?

That question is not rhetorical, and it is not naive. It is the question that produced the entire open-weight movement. It is the question that produces small, efficient, locally runnable models instead of trillion-parameter monoliths. It is the strategic answer to a metering regime that keys on compute — because a technology whose governance trigger is resource consumption has an obvious, rational escape: consume less.

Nobody in the current debate is treating efficiency research as a governance strategy. They should. It is the only move on the board that is simultaneously a technical advance and a policy position — artifacts of a new digital renaissance that double as arguments about who gets to regulate them.

X. What I Would Actually Watch

The honest analyst's output is not a verdict. It is a short list of things that would change the verdict.

First, and before anything else: the actual text of "Pace The Frontier." Its signatories, its funding, its drafting process, and whether it is a voluntary commitment or a proposed instrument. A voluntary safety pledge and a mandatory licensing regime are separated by an order of magnitude in consequence, and both can be described with the same four words. Any analysis that does not distinguish them is noise.

Second, the threshold specification. If the framework keys on training compute, the number matters more than the philosophy. A threshold set at the current frontier sorts everyone below it into a different regulatory universe, and the population below it is where nearly all experimentation lives. I have seen this exact dynamic in the digital-asset world, where a reporting obligation calibrated to the largest issuers was, in practice, a definition of which issuers counted at all.

Third, the composition of whatever body writes the evaluation standards. Not the politicians. The standards body. This is the room where pace becomes a number, and it is the room where the compliance-industrial complex I described earlier does its most durable work. A standards body staffed predominantly by people whose employers sell evaluation services is not a scandal — it is a Tuesday. It is also the single most consequential fact in the entire story, and it will receive the least coverage.

Fourth, whether the open-weight question is answered explicitly. Ambiguity here is not neutrality. Ambiguity is a cost imposed on the largest population of developers, borne silently and unaccountably, and it is the kind of cost that shows up three years later as a capability gap nobody can explain.

Fifth, the market's own answer. Watch the dispersion within AI-adjacent exposure rather than its level. If a tightening regime genuinely benefits incumbents, the correct trade is not the sector — it is the spread between the top of the sector and the bottom, and that spread will widen before the headline number moves at all. Mapping that dispersion is more informative than any statement from any official, and it is measurable today.

XI. Takeaway: The Question Behind the Accusation

David Sacks called an initiative regulatory capture. The economics say the mechanism he described is real, and it is the default outcome of any fixed-cost regime — which means his accusation is not bold, it is routine. The interesting fact is not that someone said it. It is that this time, the person saying it holds the levers.

That is the symmetry worth sitting with. A former venture capitalist, whose portfolio once depended on the regulatory environment he now helps define, is the loudest voice against the capture of rulemaking by incumbents. Both things can be true. That is not hypocrisy; it is the ordinary condition of a system in which everyone with standing to object has standing because of something.

The deeper problem is that the accusation is unfalsifiable in the form it has been deployed, and unfalsifiable accusations are the end of policy analysis. To escape that, someone has to do the boring work: read the initiative, count the thresholds, name the signatories, and ask who gets paid to certify compliance.

Unearthing the human story behind the hash rate taught me that infrastructure debates always look technical and are always about who gets to decide. Semiconductors. Compute thresholds. Evaluation standards. Signature requirements. None of these are neutral. All of them are written by someone, in a room, with an interest.

The frontier will be paced. It is already being paced, by capex cycles, by power contracts, by the availability of transformers, by the speed at which humans can label data and organizations can absorb change. The only live question is whether the pacing happens visibly, in a document we can read and argue about, or invisibly, through the quiet accumulation of advantages nobody names.

So here is the question I would put to both camps, and I do not think either has an answer: if you stripped the word safety from one side's vocabulary and the word innovation from the other's, would the rulebooks they propose look any different?

Until someone publishes the text, that is all any of us have — a name with a verb in it, and a race whose finish line is a definition.

This analysis reflects my reading of the industry as of early 2026. The specifics of "Pace The Frontier," the current status of David Sacks's role, and the precise thresholds under discussion should be verified against primary sources — which, in this case, remain conspicuously unpublished. That absence is the finding.