On September 30, after having lunch with roughly twenty AI executives in the White House East Room, Trump walked out of the White House and told reporters he had announced a voluntary "AI agreement."

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

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

Notably, on the surface this is a two-page document about AI safety with no legal force. But judging from the agreement's specific requirements, the alternative proposals that were rejected, and Nvidia's recent product moves, the more important signal it sends is this: the White House and the industry's leading companies have not chosen a path of "slowing down AI." Instead, they are trying to embed safety requirements into continuously expanding AI infrastructure spending.

"Safety compliance" requirements such as AI testing, auditing and monitoring will demand more computing power and create new demand for inference and optical interconnect, making Nvidia potentially the biggest winner. Jensen Huang even teamed up with Zuckerberg to block stronger regulation, and confronted Anthropic's chief Amodei to his face over why he was so loudly publicizing AI risk rhetoric in public.

What does the agreement actually say?

According to the original text of the agreement, the six signatories must fulfill three core obligations.

First, establish a "multi-layered control and audit" mechanism. The specific requirement is to deploy robust internal control systems that continuously monitor cybersecurity and biosecurity threats posed by advanced models and ensure the models do not intrude into other systems in "unexpected ways."

Second, cooperate with independent external auditors. Signatories must proactively submit to third-party assessment, with outside institutions verifying whether internal controls are genuinely effective.

Third, set up an independent committee at the board level specifically responsible for overseeing the above two mechanisms.

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

How much of it is genuinely new?

Not much, in fact.

According to reports, Anthropic, OpenAI and Google DeepMind had already conducted capability assessments of advanced models before release, and apart from Nvidia, the other companies had signed similar commitments after the 2024 Seoul AI Summit. Media assessments noted that "the agreement may force signatories that had not previously opened access to independent evaluation—especially Nvidia—to follow through on external audit requirements. In other words, the agreement means limited change for leading labs, but represents a new constraint for other companies."

Mackenzie Arnold, managing director of LawAI's US law and policy practice, put it bluntly: "It is most meaningful at the 'feeling' level. It does not empower any body... to actually set binding standards."

The agreement also has no legal force. Trump himself called it a "moral constraint"—which is itself an admission.

Critics were harsher. Sacha Haworth of the Tech Oversight Project said the agreement lets CEOs "grade their own homework, bypassing genuinely meaningful AI safety guardrails."

Conservative lawyer Joel Thayer, however, offered another angle: "You cannot make a public statement saying you will do something and then substantively fail to deliver. That is essentially a deception of consumers." He also noted that Federal Trade Commission Chairman Andrew Ferguson attended that day's luncheon. "These public-facing commitments... have the distinct smell of FTC enforcement."

The US FTC expanded its investigation into AI companies the day after the agreement was released, adding Anthropic, OpenAI and others to its review.

"Safety" does not equal "slowdown": that is what Jensen Huang really won

The market's instinctive reaction to "safety regulation" is to price in a slowdown—the logic being that testing slows the pace of releases, and a slower release pace compresses computing demand.

InvestorPlace analyst Luke Lango argues that logic is backwards. The White House agreement did not ask companies to slow model development, but instead focused on internal evaluation, external auditing and risk control.

This also leads to another interpretation of "AI safety" at the industry-chain level: safety mechanisms may mean more testing, monitoring and model operation, but not necessarily less computing resource input. If every model release must trigger a round of testing, monitoring and peer review, then "safety compliance" itself is an ongoing computing expense rather than a suppression of computing spending.

On September 28, Nvidia joined more than 100 partners to launch the Open Agent Safety Platform, combining open-source access control with a monitoring layer running on a separate chip—a layer that can monitor an AI system without being embedded inside it. Jensen Huang put it succinctly: AI's potential can only be realized if safety and capability are solved in tandem.

Nvidia has committed US$2 billion to optical interconnect company Lumentum to expand capacity and deepen joint R&D in data center optics—a concrete move that turns the "safety narrative" directly into hardware contracts.

Lango further noted that the demand structure is undergoing a directional shift: from GPU-intensive pretraining toward the inference and testing phase. The latter relies more on data transmission efficiency and multi-task coordination than on raw computing power. He expects this shift to dominate demand trends over the next 6 to 12 months.

Lango believes self-regulation will not end the AI capital expenditure cycle; it merely changes where the money flows—from pretraining GPUs toward inference, testing and interconnect infrastructure. As Amodei said after the meeting, the specific mechanisms for handling risk are "still under discussion."

Three beneficiary directions Lango is watching:

Marvell: focused on data center network interconnect.

Arm: in March this year it launched its first self-developed mass-production chip, the AGI CPU, featuring up to 136 Neoverse V3 cores and manufactured by TSMC on a 3nm process. Arm says its performance per rack is more than double that of x86 platforms.

Lumentum: focused on data center optical interconnect. As GPU counts surge and chip-to-chip communication density rises, data centers are accelerating the replacement of copper interconnect with optical solutions.

Behind the scenes: Jensen Huang confronts Amodei to his face

On camera, the CEOs stood beside Trump, presenting an image of industry unity.

Off camera, there was a direct confrontation.

According to the Wall Street Journal, citing people familiar with the matter, after the lunch ended, in a smaller meeting in the Roosevelt Room, several executives including Nvidia CEO Jensen Huang confronted Anthropic CEO Dario Amodei to his face: why was he issuing such extreme warnings about AI capability and risk in public?

Amodei's response was that it is important to be honest with the public about model capabilities and not to downplay risk.

The conversation took place while executives were working with White House staff to finalize the agreement's principles.

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

At that day's luncheon, when Amodei again raised safety concerns, Zuckerberg's response was: the best way to solve these problems is for the industry to genuinely implement these principles.

Zuckerberg and Jensen Huang: the real drivers of this outcome

This White House luncheon was no accident.

According to people familiar with the matter cited by the Wall Street Journal, the starting point was last week's state dinner—Zuckerberg and House Speaker Mike Johnson sat next to each other and talked about AI policy. Afterward, Zuckerberg and Jensen Huang began coordinating with various CEOs, while Trump and Johnson organized the luncheon.

Zuckerberg was the key figure in the agreement's eventual formation. He communicated frequently with Trump and Commerce Secretary Howard Lutnick, who oversees a key government AI testing office.

He and Jensen Huang jointly led something more important: blocking the previous strong-regulation proposal.

This summer, Anthropic, OpenAI and Google jointly pushed for an AI self-regulatory organization modeled on FINRA (the Financial Industry Regulatory Authority). But according to earlier reports, Zuckerberg, Jensen Huang and Musk told Trump they worried this would concentrate too much power in the hands of leading AI companies, and the proposal was quickly abandoned.

What ultimately landed was this lighter, softer voluntary agreement.

Notably, the rejected FINRA-style plan actually contained content similar to the final agreement—stronger internal review and external testing mechanisms. The two plans overlapped heavily on technical requirements; the real disagreement was over who would control the mechanism. The answer Jensen Huang and Zuckerberg chose was: the industry itself.

Anthropic's position: mending relations while preparing for an 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 had dinner at the White House on Sunday, their first face-to-face meeting.

But Anthropic's position is not an easy one.

According to reports, the company is preparing for an IPO with a target valuation of about US$2 trillion. At the same time, it experienced two direct clashes with the government earlier this year: two models were forced offline for up to two and a half weeks due to safety issues, and it was blacklisted by the Pentagon over a dispute about military use of AI.

The immediate backdrop to this agreement is a series of AI cyberattack incidents triggered after the release of models such as Anthropic's Mythos.