David Sacks called the “Pace The Frontier” initiative regulatory capture. The phrase does substantial political work. Almost none of it is technical.
I spend my working hours inside systems where the most consequential line of code is rarely the most complex one. It is the boolean that decides who may write to the ledger. A require statement. An allowlist. A threshold. In 2020 I ran ten thousand randomized liquidity shocks against Compound's V1 interest rate model in Python. The failure mode did not live in the curve's mathematics. It lived in an implicit assumption that participation would remain open and roughly uniform. Remove that assumption with a single permission and the curve inverts faster than any stress test can flag. Thresholds are where the power and the money sit. That is the substance of the Sacks argument, whether the participants intend it or not.
What was actually said, and what is missing
The reporting circulating on this is thin, and I will say so before saying anything else. David Sacks — appointed to a White House AI policy role in December 2024 — characterized the “Pace The Frontier” initiative as regulatory capture. That is the claim. What is absent is more instructive than what is present: no originating document for the initiative, no disclosure of its funding or membership, no date attached to the remark, no response from the initiative's backers, no original text. Most of the available data points trace to a single source.
I will not build a conclusion on that foundation. What I can do is what I do with any underspecified protocol: map the mechanism, locate where the trust assumptions sit, and wait for the code — or in this case, the rule text — to confirm or falsify.
Regulatory capture is not a slogan; it has a defined economic mechanism. George Stigler described it in 1971. When incumbent firms bear a lower marginal compliance cost than new entrants, regulation functions as an entry barrier, and the regulated effectively author the rules that exclude their competitors. The mechanism is symmetric, which matters here. Capture describes incumbents writing rules. It describes industries lobbying rules away with equal accuracy. The definition does not distinguish between the two directions.
The “Pace The Frontier” name itself is worth reading closely. “Pace” implies a managed rate — a governor on frontier development. If the initiative proposes pre-deployment evaluation, compute reporting, or access gating for the largest training runs, then it is a threshold proposal. And a threshold proposal is an access control list wearing policy clothing.
The threshold is a parameter, and parameters are gameable
In October 2023, US executive guidance used a training-compute figure — 10^26 FLOPs — as the trigger for reporting obligations. Whatever one thinks of that policy, its engineering form is unambiguous: a discontinuity. Below the line, no obligation. Above it, a compliance regime.
I have audited enough liquidation engines to know what discontinuities produce. A cliff creates an incentive to stand just short of the edge. This is not speculation; it is the observed behavior of every thresholded market I have examined. In DeFi the threshold is a liquidation price, and the market's response is to cluster positions fractionally above it. In compute regulation the threshold is a FLOP count, and the rational response is to fragment training runs across clusters, checkpoint and resume under partitioned budgets, or offshore the final contiguous run to a jurisdiction that does not count.
A FLOP threshold is, in security terms, a bad oracle. It measures a proxy — raw arithmetic — rather than the property the regulator actually cares about, which is capability. Proxies are trivially gamed; the block height does not lie, but a FLOP report does. Gate arithmetic and you get a rule that is simultaneously burdensome for honest actors and porous for strategic ones.
Run the cost asymmetry, because this is where the capture claim either survives or dies.
Take a frontier lab with ten researchers and a $20 million seed round. Assume annual fixed compliance costs of $2–8 million for the full stack that serious pre-deployment regimes imply: independent evaluations, red-team exercises, safety documentation, external audits, multi-jurisdiction legal review. That is 10 to 40 percent of annual runway consumed by compliance alone — and it is fixed, not marginal, so it does not scale down with a smaller model. Run the identical stack against a hyperscaler with $80 billion in annual capital expenditure. The fixed cost rounds to roughly one basis point of spend.
Three orders of magnitude separate the burden. No conspiracy is required to produce the outcome. The hyperscaler does not need to write the rule; it only needs the rule to be written in a form whose cost it can absorb and a ten-person lab cannot. That is the precise mechanism Sacks is pointing at, and it is real. Stress tests reveal the fractures before the flood, and the fracture here is fixed cost against variable balance sheets.
The part both camps are avoiding: nothing here is verifiable
Here is the blind spot, and it is the one I care about professionally.
Formal verification is the only truth in code. A contract compiles or it does not. A proof discharges or it does not. When I audited the Tezos self-amendment logic in 2017, the flaws I found were logical — provable, reproducible, citable in a patch note. When I dissected the LUNA burn mechanism in 2022, the death spiral was a deterministic sequence of function calls. The ledger remembers what the market forgets.
AI safety evaluation does not work this way. Model behavior is non-deterministic across sampling temperature, prompt framing, and version drift. An evaluation result is a statistical estimate over a moving target, not a proof. There is no assert that can be discharged. This has a consequence the policy debate has not absorbed: any rule gating frontier capability is enforced through attestation and paperwork, not through verification.
That inverts the entire framing. The question is not who writes the safety rule. The question is who signs the attestation. Whoever performs the evaluations, certifies the red-team results, and validates the compute report becomes the oracle that gates entry into the frontier market. That oracle is centralized. It is non-reproducible. It is, by construction, capturable — and unlike a FLOP threshold, it cannot be audited from the outside.
I have seen this exact architecture fail at small scale. In 2025 I audited a protocol where autonomous agents executed contracts on the basis of a language model's output. The prompt-injection vector I found did not break cryptography. It broke the assumption that a non-deterministic component could be trusted as an access control. A linguistic perturbation bypassed the gate. A governance regime that trusts an evaluation oracle is running the same pattern, one layer up.
The crypto parallel is not decorative. Stablecoin reserve attestations, sanctions screening providers, cross-chain messaging validators — each concentrates power at the attestation layer. In every case the market believed it was buying decentralization and was actually buying a new trusted third party. Verification precedes value. The AI governance debate is racing toward a design where the most valuable position in the industry is not model weights but attestation.
The contrarian reading
The comfortable conclusion is that Sacks is right — that “Pace The Frontier” is a capture vehicle — and that the correct response is to dismantle it. I do not buy that, and not because I defend the initiative.
Look at the incentive structure without the labels. Sacks spent years as a venture investor before entering policy. His stated position reduces regulatory friction on the class of companies he financed. That is not a criticism; it is a description. A position that opposes capture can itself be a capture vector, and the definition above does not exempt it. An official with an investment history declaring which safety frameworks are illegitimate is exercising exactly the authority the critique claims to distrust.
The deeper flaw runs through both camps. Neither has proposed a reduction in trust requirements. One side wants to centralize evaluation under a designated authority. The other wants to remove gates and let the market decide. Both leave the underlying architecture intact: a small number of non-reproducible attestors determining who may build. The debate is being fought over who holds the keys, with nobody asking whether the lock should exist in its current form.
Chaos is just unverified data — and so is a compliance certificate that no one can independently reproduce.
What to watch
The enforceable question for 2026 is not whether a threshold exists. It is what type the threshold is. FLOP counts are cheap to game. Evaluation regimes are non-reproducible and therefore capture-prone. Cryptographic attestation over training runs — verifiable compute, signed checkpoints — is the only variant that reduces trust rather than relocating it, and it is the variant nobody in this debate has yet proposed at scale.
Whoever ships that attestation standard will not be writing a safety rule. They will be writing the access control list for the entire frontier. Immutability is a promise, not a guarantee — and a threshold written into policy is considerably harder to upgrade than a contract.