Analysis
I have watched a lot of founders use "the DMV" as a rhetorical trump card, and it works because everyone hates the DMV. That is the entire argument. White House AI czar David Sacks deployed it this week against Anthropic CEO Dario Amodei, who made a rare weekend post on X defending both his company's regulatory stance and how it talks about AI risk, per Fortune's Jeremy Kahn.
Here is my problem with the line. Pharmaceuticals are regulated to the point that a molecule takes a decade and a billion dollars to reach a patient, and biotech is still one of the most venture-fundable categories in America. Commercial aviation is regulated down to the bolt, and Boeing and Airbus built a duopoly worth hundreds of billions inside it. Banking is regulated so heavily that compliance is a line item on every P&L, and fintech was the largest venture category of the last decade. Safety regulation did not kill any of those industries. In several of them, it created the trust that made the market big enough to be worth entering.
“Commercial aviation is regulated down to the bolt, and Boeing and Airbus built a duopoly worth hundreds of billions inside it.”
The genuine cost of regulation is not that it stops innovation. It is that it shifts advantage toward whoever can afford the compliance function. That is a real objection, and it is the one Sacks should be making, because it is the one that actually hurts startups. A federal safety regime designed by frontier labs will produce requirements frontier labs can meet. Anthropic, OpenAI and Google can each staff a policy team of dozens. A twelve-person team fine-tuning an open-weights model cannot, and the compliance burden lands on them at a fraction of the revenue.
Room for disagreement: Sacks is right that a badly built agency is worse than none. The DMV analogy is lazy, but the failure mode it gestures at is real -- an approval bottleneck with no throughput target and no appeals process would hand incumbents a permanent moat and slow American labs against Chinese ones that face no such gate. If you believe the capability gap with China is the binding national security question, then even a well-designed licensing regime is a self-inflicted delay. I do not fully buy that, because a model that compromises third-party infrastructure during a training run -- which OpenAI just disclosed happened in July -- is also a national security problem. But it is not a stupid position.
What I want from this debate is less analogy and more mechanism. Tell me the throughput target. Tell me the exemption threshold for models under a given compute budget. Tell me who pays for the audit. Founders can build around a rule with a number in it. Nobody can build around a metaphor.
