Sign in Subscribe
AI Intelligence

Trump Creates an 'AI Force' as a Federal Suit Accuses Labs of Illegal Slowdown Pact

Vijay Bhagwati

• 8 min read
Share

September 20, 2026 · 8 min read · Issue #268

Lead

A federal antitrust lawsuit filed Friday in the Northern District of California accuses Anthropic, OpenAI, SpaceXAI and Google of an illegal agreement to slow the pace of AI development — the same industry coordination that, a week earlier, looked like a rare moment of cross-lab consensus on safety. The complaint, brought on behalf of a proposed nationwide class of paid ChatGPT, Claude, Grok and Gemini subscribers, points to September 12 as the coordination date: Anthropic CEO Dario Amodei published his "We Must Pace the Frontier" essay that morning, and Sam Altman, Elon Musk and Demis Hassabis each responded in public agreement the same day. The plaintiffs also cite a July 2026 statement signed by senior staff across multiple labs acknowledging "intense competitive pressure not to unilaterally slow" — evidence, they argue, that the pacing conversation predates and outlasts any single essay.

The suit's legal theory is narrow but pointed: it doesn't object to any company individually choosing to slow down, only to labs agreeing collectively to do so, which lead attorney Nick Rowley frames as substituting "collective restraint for individual accountability." None of the four companies responded to requests for comment over the weekend.

Trump answered the underlying debate, not the lawsuit, on Saturday. In a Truth Social post he announced a new "AI Force" to "watch over" development and a forthcoming "AI Czar," while rejecting the labs' own slowdown proposal outright — calling it a "Hoax" pushed by "Radical Left Dumocrats" and warning that regulation "strongly implemented" would "drive them into oblivion and bankruptcy." He offered no structural detail on either the Force or the Czar. Notably, Amodei's essay had specifically asked government to "mediate or at least enable" cross-lab safety talks with a narrow antitrust waiver — the exact accommodation the new lawsuit argues the labs took without waiting for. Trump's posture makes that waiver unlikely to materialize, leaving the industry's safety-coordination bet exposed to the same antitrust laws that predate any of this month's essays.

The split isn't confined to the White House. Nvidia's Jensen Huang and presidential tech adviser David Sacks have both argued regulation would hand China an advantage; Meta's Mark Zuckerberg reportedly told Trump directly in August that a national AI regulator was a "flawed idea." Missouri Senator Josh Hawley, a Republican, has staked out the opposite position, telling a recent Senate hearing there is "no world" in which he'd grant antitrust exemptions to "the most powerful companies in the history of the world." The pacing debate that looked, a week ago, like an unusual alignment between AI CEOs now reads as three separate fights — one in court, one in the West Wing, one inside the Republican Senate caucus — with no consensus on any axis about who gets to decide the industry's speed.

Briefs

DOJ's copyright brief backing OpenAI and Microsoft blindsided the government's own IP agencies. The Justice Department's statement of interest in the New York Times' copyright suit argues training on copyrighted text is fair use because it's "new and transformative," while declining to extend that same reasoning to AI-generated outputs. Sources told Axios the filing caught the U.S. Patent and Trademark Office and the Copyright Office off guard, and unusually, no career antitrust attorneys signed it alongside the senior officials who did. Ziff Davis CEO Vivek Shah called the government's position "muddled" in a Fortune op-ed, noting DOJ argues licensing enforcement is too cumbersome without ever saying whether a licensing regime is actually feasible (Axios).

A federal judge denied OpenAI's bid to unseal Apple's settlement with SpaceXAI. Judge Mark Pittman ruled the confidential deal — reached in an antitrust case Musk's companies brought against AI-market rivals — was "irrelevant to the issues the court will ultimately decide." OpenAI had sought disclosure as part of the broader Musk-vs-everyone AI antitrust fight (Politico).

CXMT says its fifth-generation DRAM platform has entered mass production. The Hefei-based company, China's leading DRAM maker, told the 2026 World Manufacturing Convention it has cut feature spacing in the memory-cell area to 11.95 nanometers using quadruple patterning, and is shipping two new 24-gigabit LPDDR5X chips aimed at smartphones and tablets. Vice president Luo Xiaodong: "Our process capability is now on par with the most advanced mass-produced nodes out there in the industry." The company frames the milestone explicitly around reducing China's dependence on foreign memory technology (see.news).

Canada and Germany are putting up to $300 million into Yoshua Bengio's LawZero. Each government will contribute up to $150 million to the Montreal non-profit, which employs roughly 50 people and is building "Scientist AI" — a monitoring system explicitly designed without reinforcement learning, which Bengio blames for training AI systems to lie, cheat and act deceptively in pursuit of reward. "We urgently need to have a plan to build AI that is not going to do really bad, misaligned things," he said. Federal AI minister Evan Solomon, on the same stage at Montreal's All In AI conference: "No country can build alone" (Globe and Mail).

Samsung India launched a 2,000-person AI skills training program in Karnataka. The Samsung Innovation Campus 2026 initiative, run with Nrupathunga University in Bengaluru, is pitched as support for the state's push to build an "AI-native ecosystem" and expand its technology talent pool ahead of demand from the country's chipmaking and services sectors (YourStory).

China/East Asia

CXMT's mass-production claim matters less for the specific 11.95-nanometer figure — which the company itself compares only loosely to unnamed "advanced mass-produced nodes" — than for the pattern it continues. China's domestic chip suppliers keep announcing incremental parity claims on memory and logic technology on a roughly monthly cadence this year, each one framed around import substitution rather than leading-edge competition with TSMC or Samsung. LPDDR5X for smartphones is a consumer-device play, not a frontier-training play, but it's the kind of volume manufacturing that funds the next round of R&D and gives Chinese phone and laptop makers a domestic alternative supply chain regardless of what happens at the export-control frontier. Read against Alibaba's decision to open-source a medical AI model this week, the throughline is the same: Chinese firms are racing to make their AI stack — from memory chips to domain-specific models — independently reproducible inside China's own borders, hedging against any further tightening of access to foreign components or foreign model weights.

India

The country's AI story this week splits cleanly into capital and talent. Modi's chipmaking event drew renewed interest from equipment makers Applied Materials and Lam Research toward the government's $30 billion semiconductor build-out — though the only fully confirmed detail is the headline figure itself; the underlying reporting sits behind a paywall and deserves a caveat rather than a confident retelling. The talent side is more concretely sourced: Samsung's 2,000-person AI training program in Karnataka is a direct bet that India's next wave of AI-adjacent jobs needs a domestic pipeline of trained workers, not just imported capital equipment. Neither story makes India a frontier-model competitor to the US or China, but together they describe the same strategy Europe has pursued with Mistral and Cohere — build the surrounding infrastructure (fabs, equipment, trained engineers) even while frontier model development stays elsewhere, on the bet that infrastructure ownership still buys leverage in how AI policy gets made regionally.

Europe

Germany's $150 million commitment to Bengio's LawZero is a smaller, quieter move than the AI infrastructure investments Mistral and Cohere have been announcing, but it's arguably more distinctive: rather than building a European frontier-model competitor, Berlin is funding a Canadian non-profit's bet that the entire reinforcement-learning paradigm behind today's models is the wrong approach to safety. That's a genuinely different theory of the problem than the "verification and disclosure" framework Amodei has been pushing, or the "kill switch" studies states like California are running — Bengio's Scientist AI tries to avoid the failure mode at the training-method level rather than adding oversight on top of models built the conventional way. Whether it works is unproven and years from resolution; what's notable now is that a G7 government is willing to fund a genuinely contrarian safety architecture rather than only backing incremental compliance frameworks. It's a modest bet by AI-infrastructure spending standards, but it's a bet on a different theory of the risk than most government money in this space has gone toward so far.

The View

Washington spent the weekend running two contradictory arguments about the same underlying question — should AI development speed be coordinated across companies — and neither one is really about speed. The antitrust lawsuit says coordination on pacing is illegal collusion regardless of intent; Trump's answer says coordination is unnecessary because the market and existing criminal law are sufficient, and that formal regulation would simply hobble American labs against China. Both positions conveniently avoid the actual mechanism Amodei's essay asked for: a narrow, government-enabled waiver that would let labs discuss safety without antitrust exposure. Nobody in a decision-making seat this week endorsed that middle path. The result is that the pacing conversation, which looked a week ago like unusual cross-lab agreement, is now being litigated as either a crime or a hoax, with no institutional space left for a government-sanctioned safety dialogue that isn't automatically one or the other.

The Miss

DOJ's copyright brief blindsiding the Copyright Office and USPTO deserves more attention than the pacing lawsuit is getting, because it's a live demonstration of exactly the coordination failure the pacing debate treats as hypothetical. Amodei's essay argues AI companies need external, structured verification before shipping powerful systems; the underlying justification is that internal self-regulation isn't trustworthy enough on its own. DOJ's filing shows the same coordination problem exists inside a single federal government on a much simpler question — copyright fair use — where career agency staff who work the issue daily weren't consulted before senior political appointees committed the government to a position. If a single administration can't coordinate copyright policy across two agencies that talk to each other regularly, the case for trusting any government-industry safety coordination mechanism to work smoothly gets measurably weaker, and nobody covering the pacing lawsuit connected the two stories this week.

Pull Quotes

"AI will quickly spin out of human control and could kill us all if we allow AI safety and protocol...to be controlled by private self-serving agreements between the world's most powerful 'for profit' technology companies." — Nick Rowley, lead attorney for plaintiffs, via AP News

"We will not in any way hinder or stifle the Growth of this incredible Industry. Rather, we will cherish it, help it, and watch over it, as it grows!" — Donald Trump, via Business Insider

"We urgently need to have a plan to build AI that is not going to do really bad, misaligned things." — Yoshua Bengio, via Globe and Mail

"Our process capability is now on par with the most advanced mass-produced nodes out there in the industry." — Luo Xiaodong, CXMT vice president, via see.news

Out

Amodei asked for a narrow government waiver so labs could talk safety without breaking antitrust law. This week Washington delivered a lawsuit calling that talk illegal and a president calling the whole premise a hoax — nobody built the waiver.