The number that decides who gets to build frontier AI in America is 10^26. Floating-point operations. That's the training-compute threshold US policy has used as its regulatory trigger — cross it and you owe the government a safety report, a red-team dossier, an evaluation packet, and a paper trail that takes a general counsel to produce on schedule. Stay under it and nobody asks you anything at all.
David Sacks looked at an initiative called "Pace The Frontier" and called it regulatory capture. Predictable from him. But he's not wrong about the mechanism — he's only wrong about who gets to complain.
I've been watching this exact movie since 2017, when I sat in a University of Lagos dorm room and manually verified a presale contract address on Etherscan before anyone else had the tab open. A decade of rule-makers deciding who is allowed to be dangerous and who is merely allowed to be small. The fight over "Pace The Frontier" is the same fight in better suits, in a bigger room, with more money on the table.
One disclosure before we go further. The reporting I'm working from is a headline and a handful of bullet points. No text of the initiative. No date on the quote. Four of the five information points trace back to a single source — Sacks himself. No response from whoever drafted the thing he's attacking. Treat everything downstream of this paragraph as analysis, not confirmed fact.
Who's talking. David Sacks — PayPal mafia, Craft Ventures founder, and since December 2024 the White House's point person on both AI and crypto policy. His résumé is the story. A venture investor steering the rulebook for the sector his portfolio lives in. That's not an accusation, it's a job description, and it cuts in every direction at once.
What he's attacking. "Pace The Frontier" — on its face, a governance framework whose name advertises its own thesis. Pace implies a speed governor. Frontier implies the thing being governed is the capability edge itself. As reported, it pairs frontier capability with safety guardrails and treats international coordination as a live possibility. Sacks' counter is that this framing suppresses innovation and tilts the competitive field toward whoever can afford to comply.
Why a crypto desk is writing about AI policy at 2 a.m. Because the boundary between these two industries dissolved somewhere around 2023 — same capital, same VC syndicates, same compliance departments, same token structures bolted onto GPU clusters. Every AI governance rule that lands is a de facto crypto rule, because decentralized compute networks, inference markets, agent tokens, and verifiable training all sit inside whatever perimeter gets drawn.
And because "regulatory capture" is not a novel accusation in our house. It's the founding grievance. The 2018–2020 era of "come in and register" was capture performed as courtesy. Every American token issuer that got a Wells notice was told the rules were clear, then watched them get enforced against whoever couldn't fund a defense. The pattern was consistent: rules written loosely enough to be selectively enforced.
So when a sitting policy official uses our vocabulary against a rival governance camp, the honest move is not to cheer. It's to check whether the mechanism is real, who it actually prices out, and whether we're about to become the incumbents who benefit from it.
A quick note on what single-source reporting does to a reader's calibration. Four of the five information points here are Sacks' positions. The initiative's own framing is absent from the record. Whoever wrote "Pace The Frontier" gets a news cycle built entirely from their opponent's description of them. That isn't a media critique so much as a description of how policy narratives form — the first party to name the frame usually wins the week, and Sacks is very, very good at naming frames.
Now the technical translation, because "capture" is not a moral category. It's a pricing mechanism.
Compliance is a fixed cost. Fixed costs are the moat. A ten-person lab and a ten-thousand-person lab both need a safety report, a red-team assessment, an evaluation suite, an audit trail, and legal review before shipping a frontier model. The large lab amortizes that across a revenue base a thousand times bigger. The small lab pays the same absolute number against a seed round that's already half spent. Identical rules, wildly asymmetric cost.
That is the entire Stigler mechanism, and it requires nobody to be corrupt — only for the rules to be written by people who can afford the lawyers.
Think of the FLOPs threshold the way you'd think about a validator set. If the rule only binds above a fixed compute level, then the entire market below that level is unregulated by design and the entire market above it is regulated by invitation. That's not a safety perimeter. That's a membership tier with a mission statement.
Here's where it gets relevant to us: the compliance cost function in AI looks almost exactly like the data-availability cost function in rollups. Both are metered inputs. Both are cheap right now. Both are one auction mechanism away from repricing hard.
I've been tracking blob economics since EIP-4844 went live. The pitch was simple — proto-danksharding gives rollups cheap blobs, blobs decongest calldata, L2 fees collapse, and the rollup thesis finally pays off at the settlement layer. And it worked. For a while. Blob space was effectively free, every rollup's margin looked spectacular, and an entire generation of L2 business models got underwritten on the assumption that it would stay that way.
Post-Dencun blob data will saturate within two years. When it does, every rollup that built its fee model on cheap blobs reprices — several of them by 2x. Not because anyone did anything wrong, but because a metered resource that everyone underwrites on being cheap eventually stops being cheap. I've watched this cycle in gas markets, in block space, in every scarce resource crypto has ever auctioned. The first time is education. The second time is negligence.
Translate that to AI compute. Today's GPU supply looks like post-Dencun blob space. Accelerators are abundant, inference is subsidized, VC dollars and hyperscaler capex cover the spread, and a dozen decentralized compute networks are selling cycles below cost. Everyone's building unit economics on the assumption that compute stays cheap and rules stay light.
Every subsidy looks like an architecture until the subsidy ends.
Which brings me to the second place the analogy actually earns its keep — the token-incentivized compute networks.
I've audited a handful of these. Not their contracts — their numbers. The pattern is identical to DeFi yield farming circa 2020, which I covered live from a Lagos apartment while a flash loan drained a lending protocol in real time and I was posting transaction hashes into a Telegram group of four hundred strangers.
Liquidity mining APY is the project paying for the TVL number, not the market saying the product is good. Stop the incentive and the real users evaporate before the epoch ends.
Swap "TVL" for "decentralized training FLOPs" and the sentence holds word for word.
The audit method takes about two hours if you know where to look. First, separate paying jobs from incentivized jobs — most of these networks publish utilization dashboards that quietly count self-dealing. Check whether the largest compute providers share downstream ownership with the largest buyers. Second, lay the token emission schedule against claimed compute; if emissions per FLOP sit above spot-market GPU rental, you're not looking at a market, you're looking at a subsidy with a whitepaper attached. Third, find the one number nobody publishes — retained paid utilization thirty days after an incentive epoch closes.
If you cannot find that third number, you have your answer.
None of that kills the sector. It reframes it. Verifiable inference and permissionless compute are real primitives — genuinely novel, genuinely useful, and I will defend them in any room against anyone. But the flywheel currently being sold to retail is a marketing artifact riding on cheap capital, and cheap capital is the first thing that reprices when the cost function turns. In the void, we found our value in the noise — not in the emission schedule.
Third thread: fragmentation and arbitrage.
Sacks' position reportedly includes the claim that global AI governance cooperation isn't feasible. Read that twice. It is not a prediction. It is a strategy — and it lines up with the accelerationist camp that has argued against multilateral coordination since the Techno-Optimist Manifesto made it fashionable. If cooperation is dead, the outcome is arithmetic: a race to the bottom between jurisdictions, each competing on how little they'll ask of frontier labs.
Crypto already ran that experiment. Swiss foundations, Estonian e-residency, Wyoming DAOs, Singapore family offices, Bahamian exchanges that worked fine until they didn't. Regulatory arbitrage is not a strategy; it's a delay. DeFi was not a bug; it was a feature of chaos — and chaos is a great place to build and a terrible place to compound.
What the fragmentation argument keeps missing is where demand actually originates.
I've watched crypto payments succeed in Lagos and fail in San Francisco, and the difference was never ideology. The real driver of crypto payments in developing countries isn't blockchain philosophy — it's local currency inflation forcing people to find survival alternatives. A trader in Yaba doesn't care about disintermediation as a concept. She cares that the naira bought less this quarter than last quarter and the dollar rail is closed to her.
Apply that lens to AI and the policy debate inverts. The jurisdictions that will adopt these systems fastest — Nigeria, Kenya, Indonesia, Brazil — are not going to pick a stack based on whether Washington and Brussels agree on red-teaming standards. They'll route inference to whoever is cheapest and wherever the API answers. If open-weight models are accessible and American frontier access is gated behind compliance paperwork, the deployment map draws itself. Governance frameworks designed in capitals are downstream of deployment decisions made in markets — and the markets are in the global south, where nobody is waiting for a communiqué.
Now the part nobody's publishing.
Sacks calling something "regulatory capture" is not a neutral act by a neutral observer. He is both a policy official and a member of the investor class whose returns are most sensitive to regulatory friction. An anti-capture argument advanced by a capital-owning policymaker is still a capture argument — just aimed at a different set of incumbents. Deregulatory capture satisfies Stigler's definition as cleanly as regulatory capture does. The beneficiary changes. The mechanism doesn't move an inch.
That's not a gotcha, it's a symmetry. Which means the crypto industry's reflexive instinct — deregulation good, regulation bad, cheer for whoever says "innovation" loudest — is the wrong reflex, and it's about to cost us.
Because we are no longer the insurgents. There's a spot Bitcoin ETF with BlackRock's name on it. Stablecoin legislation is being drafted with input from firms that once wanted the SEC abolished. Market-structure bills are being negotiated by people who'd prefer the rules be finished before their competitors arrive. The industry is now large enough to be the incumbent in its own categories — which means "safety" and "standards" framing is coming for us next, and it will be written by the biggest players in the room.
If you cheer when a compliance moat gets torn down for AI labs, you had better notice when someone builds the same moat around chain abstraction, cross-border stablecoin rails, or validator-level compliance. The blind spot in this entire debate is that both camps agree on the load-bearing assumption — that the rules get written in Washington, Brussels, and Beijing at all. Neither side is arguing for the actual alternative crypto represents: verifiable, permissionless, unstoppable computation whose rules are enforced by code and checked by anyone running a node.
The story isn't in the chart. The story is in the pulse — and right now the pulse is two governance philosophies fighting over who gets to be the gatekeeper, while the permissionless option sits off-stage, unrepresented in the argument that decides its fate.
Watch the paper, not the podcast. Three things tell you whether "Pace The Frontier" is a governance framework or a moat with a mission statement.
One: get the actual text. If the threshold is training compute above 10^26 FLOPs, the rule is a size filter wearing a safety badge, because only a handful of organizations on Earth train at that scale. Two: the signatory list. If frontier labs sponsored it, expect open-weight redistribution to receive the tightest language — that's where the competitive threat lives. Three: Sacks' first concrete action. Rhetoric is free. A rescinded executive order or a defunded safety institute is not.
The trade this creates is boring and reliable: third-party compliance, model evaluation, red-teaming, and audit services. Same shape as crypto's audit firms after 2020 — an entire cottage industry born the moment the rules became mandatory and nobody trusted self-reporting. Regulatory friction is a market. It always has been.
And the question worth sitting with, past the price: if the rules deciding who may build intelligence are written by the people who fund intelligence, what exactly is being governed — and on whose behalf?