Larry Summers' AI Attack Signals A Fracture In Washington's Governance Consensus

A Rare High-Level Voice Enters The AI Governance Debate With Pro-Innovation Hardline

David Saxs, co-chair of the Presidential Committee on Science and Technology Advisors (PCAST) under the Trump administration, has delivered one of the most direct public challenges yet to the AI industry’s growing emphasis on “pacing” development. His September 12 intervention on X marks a significant escalation in Washington’s internal debate over how to govern artificial intelligence, and Japanese observers are watching with unusual intensity. The timing and content of the remarks suggest a deliberate recalibration — not merely a statement of principle, but a calibrated nudge toward faster deployment cycles across the American AI ecosystem.

Saxs responded specifically to Anthropic CEO Dario Amodei’s essay “We Must Pace the Frontier” and Sam Altman’s public support for the same thesis. His message was blunt: pace what you want, but stop pretending that regulation is a prerequisite for doing so. The distinction matters because it reframes the entire discourse. Pacing, Saxs implied, is a choice — not an inevitability — and the companies that have been most vocal about it are simultaneously the ones most capable of defying it.

The timing is politically notable. Saxs, formerly the Trump administration’s special advisor on AI and cryptocurrency, left his White House role in March 2026 after reaching the limit on days permitted for special government employees. He now chairs PCAST, which includes NVIDIA CEO Jensen Huang, Meta’s Mark Zuckerberg, and AMD’s Lisa Su. With US midterm elections approaching in November, his intervention carries signals about where the administration’s tech-governance compass truly points. It also reflects Larry Summers’ broader philosophical influence — the former Treasury secretary has long championed the idea that regulatory frameworks should follow innovation, not precede it, and Saxs’s remarks read as a direct application of that thinking to the AI domain.

The 5 Things Saxs Told Companies To Stop Doing

Saxs enumerated five specific behaviors he considered unacceptable, each targeting a different pillar of the industry’s self-regulatory narrative:

First, companies should stop acting as though they need permission from others to proceed. This struck at the heart of the pacing argument — the assumption that independent safety bodies or governmental review boards must greenlight frontier model releases.

Second, they should not argue that antitrust enforcement must be suspended to form cartels. The wording was deliberately sharp. Saxs flagged that requests for regulatory suspension sounded like collaboration but functioned as coordination disguised as compliance.

Third, they should not demand regulatory approval processes that override product liability. This was the most legally significant point. Saxs suggested that outsourcing accountability to a regulatory body would effectively insulate companies from the very tort claims that might otherwise constrain reckless deployment.

Fourth, they should not treat organizations like METR — which counts investors and staff members from both Anthropic and OpenAI — as independent actors. The funding overlap between测评 bodies and the companies they assess undermines any claim to neutral evaluation.

Fifth, they should not claim that competitors who are not even at the frontier level need the same scrutiny. This pointed directly at the asymmetry in the governance debate: the companies setting the pace are also the ones writing the rules for everyone else.

The underlying point: the companies defining the frontier are Anthropic and OpenAI themselves, and their combined market share, revenue growth, and model capabilities give them near-monopoly control over frontier AI development. Any argument that they cannot proceed without regulatory frameworks looks less like public-spirited caution and more like regulatory capture dressed as concern.

The Product Liability Angle No One Else Is Making This Explicitly

Perhaps the most consequential part of Saxs’s intervention was his argument about motive. He suggested the slowdown rhetoric may not be purely altruistic — it may be commercially rational.

If a company’s own product is used in a serious cyberattack, they face enormous product liability. The market has already punished models that behave unpredictably or outside their intended scope. After the Hugging Face incident — where an open-source model was repurposed to automate phishing campaigns at scale — reducing capability in exchange for controllability became not a moral choice but a sound business decision. Call it alignment if you like; Saxs noted it may simply be giving customers what they want: safe, deniable, controllable AI that cannot be weaponized by third parties.

This reframes the entire governance debate. What looks like responsible stewardship may be risk management in corporate clothing. The companies aren’t asking for regulation because they care about public safety; they’re asking for regulation because it limits their own liability exposure and creates barriers to entry for competitors. A regulatory framework that requires approval before deployment raises the cost of building frontier models — a cost that only the well-capitalized can absorb. Startups, foreign competitors, and public-sector researchers become the collateral damage of an industry that has monetized caution.

The second-order effect is subtler still. If Saxs is correct, the pacing narrative functions as a soft monopoly mechanism. It does not explicitly exclude competitors — it makes competition structurally expensive. Every compliance step, every safety audit, every government filing becomes a toll gate. And the companies that built the gate collect the fees, indirectly, through reduced competitive pressure.

Why Japan Is Watching Closely

Japanese sources are covering Saxs’s intervention with more nuance than many English-language desks might capture. Tokyo has long balanced its AI strategy between rapid innovation and cautious governance, and the US position sets the ceiling for what global coordination is possible. Japan’s own STI (Society 5.0) framework has explicitly sought to mirror American approaches while carving out distinct regulatory space — a position now under strain.

Saxs acknowledged this reality directly. He noted that Bernie Sanders’ call to “stop everything” would actually provoke more useful intellectual debate than the current industry self-regulation framework. More critically, he stated that China is extremely unlikely to join any global consensus — a point that should temper expectations for multinational AI governance arrangements regardless of what Washington decides. Beijing has already signaled its intent to compete, not cooperate, on AI governance terms, and Saxs’s acknowledgment of this fact removes the last veil of diplomatic optimism around multilateral frameworks.

For Japan, the implication is clear: the era of expecting coordinated global AI regulation is over. The US is moving toward a pro-innovation stance, China will not cooperate, and the rest of the world must decide whether to follow Washington’s lead or chart its own course. That decision carries economic weight. Japan’s semiconductor industry, already rebuilding after the 2024 supply chain disruptions, cannot afford regulatory divergence with the United States. But neither can it afford to fall behind China’s state-directed deployment cadence. The narrowing window between these two poles is what makes Saxs’s remarks so consequential for Tokyo.

The Political Timing And What Comes Next

Saxs’s framing had an implicit political edge. He argued that if Anthropic and OpenAI want to set the frontier, the simplest way to prevent superintelligence is simply for those two companies to agree not to build it. Asking for a regulatory framework as a condition looks less like governance and more like blackmail against the public and the political system.

His advice was straightforward: execute first, earn credibility through action, then discuss the next framework. If they refuse, it will become obvious whether this is regulatory capture or election-season opinion management. The latter possibility is particularly pointed given the November midterms. A political narrative that paints AI companies as hostage-takers of public policy carries electoral weight, especially in districts where tech industry influence is resented.

The intervention also reveals a fracture within the broader pro-innovation coalition. Summers himself has not publicly endorsed Saxs’s specific claims, and some of his former colleagues in the regulatory establishment have expressed concern that the pacing critique understates genuine safety risks. The disagreement is not merely tactical — it reflects a deeper philosophical split about whether AI risk is real, imminent, and governable, or whether the greatest risk is the one posed by concentrated corporate control of critical infrastructure.

With midterm elections five months away, the intervention cannot be read as purely technocratic. It signals where the pro-innovation wing of the Trump administration stands and may pressure AI companies to pick a side before the political calculus shifts again.

The companies now face a choice. Continue the pacing narrative and risk being publicly branded as regulatory capture artists, or accelerate development and invite scrutiny on their own terms. Saxs has drawn a line in the sand; the question is whether Anthropic and OpenAI will step over it or build a fence around it. And the rest of the world — including Japan — will be watching to see which path they choose.