What does Trump's "AI self-regulation" mean? Jensen Huang achieves a major victory, leads the charge questioning Anthropic's Amodei behind the scenes

On September 30, after having lunch with about twenty AI executives in the East Room of the White House, Trump walked out and announced to reporters a voluntary “AI Accord.”

The signatories are six companies: Anthropic, OpenAI, Google, Meta, xAI, and Nvidia.

Trump described it as “almost like a constitution,” calling it “a form of protection” against the risks of AI. However, according to the Financial Times, the two-page document was hastily drafted in the East Room, and even included a spelling error.

It is worth noting that, on the surface, this is a two-page document about AI safety, lacking legal force. However, judging from the concrete requirements of the accord, the rejected alternative proposals, and Nvidia’s recent product moves, the more important signal is: the White House and leading industry companies have not chosen the path of “slowing down AI,” but aim to embed safety requirements into the ongoing expansion of AI infrastructure investments.

“Safety compliance” requirements such as AI testing, auditing, and monitoring will require more computing power, generating new demand for inference and optical interconnects, potentially making Nvidia the biggest winner. Jensen Huang has even teamed up with Zuckerberg to oppose strong regulation and openly questioned Anthropic chief Amodei about why he overemphasizes AI risk in public statements.

What does the accord say?

According to the text of the accord, the six signatories must fulfill three core obligations:

First, establish a “multi-layered control and audit” mechanism. Specifically, this requires the deployment of robust internal control systems, ongoing monitoring of cybersecurity and biosecurity threats from advanced models, and ensuring that models do not invade other systems in “unintended ways.”

Second, work with independent external auditing agencies. The signatories must proactively accept third-party assessments to verify if internal control measures are truly effective.

Third, set up an independent committee at the board level specifically in charge of supervising the above two mechanisms.

These three obligations point to a common outcome: every model release must go through a complete process of testing, monitoring, and external verification.

How much of this is actually new?

Not much, actually.

Reportedly, Anthropic, OpenAI, and Google DeepMind have already been evaluating advanced models prior to release, and apart from Nvidia, the other companies all signed similar commitments after the 2024 Seoul AI Summit. Media have commented that “the accord may force signatories who have not yet opened access for independent evaluation—especially Nvidia—to implement external audit requirements. In other words, the accord has limited implications for leading labs, but imposes new constraints on other companies.”

Mackenzie Arnold, Managing Director of LawAI’s US Legal and Policy Practice, commented directly: “It is most meaningful at the ‘perception level’. It does not authorize any agency… to actually set binding standards.”

The accord also lacks legal enforceability. Trump himself called it a “moral constraint”—which, in itself, is an admission.

Critics are even more blunt. Sacha Haworth of the Tech Oversight Project argued that the accord allows CEOs to “grade their own homework, sidestepping real, effective AI safety guardrails.”

However, conservative lawyer Joel Thayer offered another perspective: “You can’t make a public statement to do something and then not actually deliver. That is essentially a deception of consumers.” He also pointed out that FTC Chair Andrew Ferguson attended the luncheon. “These public-facing commitments... they scream FTC enforcement.”

On the day after the accord’s release, the US FTC expanded its investigations into AI companies, including Anthropic and OpenAI.

“Safety” is not the same as “slowing down”: This is what Jensen Huang really won

The instinctive market reaction to “safety regulation” is to assume a slowdown—which logically means: testing slows the release rhythm, slower releases suppress demand for computing power.

InvestorPlace analyst Luke Lango believes this logic is inverted. The White House accord does not require companies to slow model development, but instead places emphasis on internal evaluation, external audit, and risk control.

This also leads to another interpretation of “AI safety” at an industrial level: safety mechanisms could mean more testing, monitoring, and model operation, not necessarily less computational resource investment. If every model release must trigger rounds of testing, monitoring, and peer review, then “safety compliance” itself is a sustained expense in computing power—not a suppression of it.

On September 28, Nvidia partnered with more than 100 partners to launch the Open Agent Safety Platform, which combines open-source access control with a monitoring layer running on independent chips—this layer can monitor AI systems without embedding itself inside them. Jensen Huang put it simply: AI’s potential can only be realized when safety and capability are solved in parallel.

Nvidia has already promised to invest $2 billion in the optical interconnect company Lumentum to expand capacity and deepen joint R&D for data center optics—a concrete step that translates “safety narrative” directly into hardware contracts.

Lango further pointed out that the structure of demand is shifting directionally: from GPU-intensive pre-training, toward inference and testing phases. The latter rely more on data transfer efficiency and multitasking coordination than on raw computing power. He expects this shift to dominate demand trends over the next 6–12 months.

Lango believes that self-regulation will not end the AI capital expenditure cycle; it simply changes where the money goes—from pre-training GPUs to inference, testing, and interconnect infrastructure. As Amodei said after the meeting, the specific mechanisms for managing risks “are still under discussion.”

Lango highlights three main beneficiaries:

Marvell: specializing in data center network interconnects.

Arm: launched its first self-designed mass-production chip—the AGI CPU—in March, featuring up to 136 Neoverse V3 cores manufactured by TSMC on a 3nm process. Arm claims per-rack performance exceeds the x86 platform by more than double.

Lumentum: focused on data center optical interconnects. With the surge in GPU counts and increased chip-to-chip communication density, data centers are rapidly replacing copper interconnects with optical solutions.

Behind-the-scenes clashes: Jensen Huang confronts Amodei face-to-face

On camera, CEOs stood beside Trump, showing a united industry front.

Off camera, a direct confrontation took place.

According to The Wall Street Journal, after lunch, in a smaller gathering in the Roosevelt Room, several top executives—including Nvidia CEO Jensen Huang—directly questioned Anthropic CEO Dario Amodei: Why make such extreme public warnings about AI’s capabilities and risks?

Amodei responded that it is important to honestly inform the public about model capabilities and not downplay risks.

This dialogue took place while executives and White House officials were finalizing the principles of the accord.

Amodei had repeatedly criticized the White House and competitors for downplaying safety issues and warned that AI models could enable cyberattacks and cause mass unemployment, making him a lightning rod in both Washington and Silicon Valley.

During the luncheon, when Amodei once again raised safety concerns, Zuckerberg replied: the best way to address these issues is for the industry to make these principles a reality.

Zuckerberg and Jensen Huang: The real driving forces behind this accord

This White House luncheon was not a coincidence.

According to The Wall Street Journal, the starting point was last week's state dinner—Zuckerberg was seated next to Speaker of the House Mike Johnson, and they discussed AI policy. Subsequently, Zuckerberg and Jensen Huang began coordinating among CEOs, while Trump and Johnson organized the luncheon.

Zuckerberg was the key person behind the final formation of the accord. He communicated frequently with Trump and Commerce Secretary Howard Lutnick, who oversees a crucial government AI testing department.

He and Jensen Huang jointly led something even more important: they vetoed the earlier strong regulatory proposal.

Earlier this summer, Anthropic, OpenAI, and Google had advocated for an AI self-regulatory organization similar to FINRA (Financial Industry Regulatory Authority). But according to earlier reports, Zuckerberg, Jensen Huang, and Musk told Trump they worried this would vest too much power in the hands of leading AI companies, leading to the proposal being abandoned.

What ultimately materialized was this lighter, softer voluntary accord.

It is worth noting that the rejected FINRA-style proposal actually also included content similar to the final accord—stronger internal reviews and external testing mechanisms. The two frameworks were highly overlapping in technical requirements; the real dispute was over who would control these mechanisms. Jensen Huang and Zuckerberg opted for the industry itself.

Anthropic’s situation: repairing relations while preparing for IPO

Amodei told reporters that day: “If industry and government work hand in hand, we can win safely.” Trump praised him as “great.” The two had just dined at the White House on Sunday, their first face-to-face meeting.

But Anthropic’s situation is not easy.

Reportedly, the company is preparing for an IPO, targeting a valuation of about $2 trillion. Meanwhile, it faced two direct run-ins with the government earlier this year: two models were forced offline for two and a half weeks due to safety issues, and it was blacklisted by the Pentagon over AI for military use controversies.

The immediate backdrop for the accord was a series of AI cyberattack incidents triggered by Anthropic’s Mythos and other models after their release.