Trump Canceled His Own AI Executive Order — What That Decision Reveals About Silicon Valley's Power Over Washington

Introduction

On the morning of May 21, 2026, the White House had already sent out invitations. Technology CEOs had cleared their calendars. Senior officials had conducted a press briefing. The Oval Office ceremony was scheduled for that afternoon, and the event was set to mark the Trump administration's first serious attempt to put a governance framework around the most powerful artificial intelligence systems in the world. Then, hours before it was due to happen, the president canceled it.

Trump's stated reason was blunt: he did not like what he saw in the order's text. He told reporters he thought it could undermine the United States' lead over China in artificial intelligence. "We're leading China, we're leading everybody, and I don't want to do anything that's going to get in the way of that lead," he said. He added that the order "could have been a blocker." The signing was called off without advance notice. The executives who had been invited to attend found out the same way the public did — through news reports and a presidential statement delivered at an unrelated Oval Office event.

What followed was two weeks of intense, private negotiation. What emerged from those negotiations tells you something specific about where real power in the relationship between Washington and Silicon Valley currently sits — and it is not where most Americans assume.

What the Original Order Actually Said

The draft that Trump refused to sign on May 21 had been in development for approximately six months. According to people familiar with the text, it was intended to direct federal agencies to strengthen key cybersecurity defenses and to establish partnerships with leading AI companies to test advanced models before they were released to the public. A review window of up to 90 days was reportedly included, during which the government could assess the national security implications of a new frontier model. The order would also have established a national AI research cloud and created a federal AI civil rights office to address concerns about algorithmic discrimination.

By the standards of AI governance proposals in Europe or the United Kingdom, this was a restrained document. It did not propose mandatory pre-release approval authority. It did not create a formal licensing regime. It did not establish criminal penalties. But it did include a 90-day review period that, depending on how broadly it was applied, could have created meaningful friction in the release timelines of the most advanced AI systems in the world. That was the provision the technology industry found unacceptable. A 90-day review window, even a nominally voluntary one, is three months of revenue, three months of competitive disadvantage relative to Chinese labs that face no such constraint, and three months of legal uncertainty for any company that chose not to participate while a competitor did.

After the cancellation, a round of frantic negotiation began involving frontier AI executives including Sam Altman, Dario Amodei, Sundar Pichai, Satya Nadella, Elon Musk, Mark Zuckerberg, Jensen Huang, Tim Cook, and David Sacks. They lobbied to change the 90-day mandatory blocking window and reduce it to a 30-day voluntary review, salvaging the security framework while keeping the competitive edge. That is a remarkable list of names to have assembled in opposition to a single provision. It is also a remarkable outcome: a sitting president agreed to the change.

Why the Banking Sector Wanted This Order in the First Place

The original push for an AI executive order did not come from civil society groups or academic researchers. It came from Wall Street. In April 2026, Treasury Secretary Scott Bessent and outgoing Federal Reserve Chair Jerome Powell convened an urgent meeting with Wall Street CEOs, warning them about the cybersecurity risks posed by Anthropic's AI model, Claude Mythos. The meeting was assembled at short notice at the Treasury Department's headquarters. Bessent described it as an effort to ensure that banks were aware of the risks and to discuss best practices.

The order's planned debut would have come less than two months after Anthropic's Mythos Preview model demonstrated the novel ability to autonomously discover thousands of severe and critical cyber vulnerabilities, including in leading operating systems and web browsers. This was not a hypothetical capability. Mythos did not set out to become a cybersecurity weapon. The vulnerability-finding capability emerged as what Anthropic describes as a downstream consequence of general improvements in code reasoning and autonomous operation. During testing, Mythos identified thousands of zero-day vulnerabilities — flaws previously unknown to software developers — across every major operating system and web browser. The capabilities were significant enough that Anthropic chose not to release the model publicly.

Instead, Anthropic launched Project Glasswing, a controlled program giving access to roughly 50 organizations including Amazon, Apple, Google, Microsoft, Cisco, CrowdStrike, Palo Alto Networks, and JPMorgan Chase, with the purpose of giving major institutions a head start on strengthening their defenses before the capability became more widely available. The banking sector was simultaneously alarmed by Mythos's power and eager to use it. JPMorgan Chase, Goldman Sachs, Citigroup, Bank of America, and Morgan Stanley were all reportedly testing the model. The government's role, in the version of events Washington was comfortable with, was to provide a framework that managed the risk without choking off access.

What the Technology Industry Changed, and How

The version of the order that Trump eventually signed on June 2, 2026, was materially different from what had been prepared for the May 21 ceremony. The order asks AI companies to voluntarily submit their most powerful models for the government to test up to 30 days before releasing them to the public. It also directs federal agencies to develop benchmarks to assess AI models' cyber capabilities, to create an AI cybersecurity clearinghouse to review and share information on vulnerabilities, and to shore up the government's security defenses. The keyword that every technology company had lobbied for was in the text: voluntary.

The order explicitly bars the government from creating a mandatory licensing or preclearance requirement for new AI models, making the government's move a request, not a rule. The 90-day blocking window was gone. The mandatory testing period was gone. What remained was an invitation. Sam Altman publicly praised the compromise, stating that the executive order "gets the balance right." Arvind Krishna, CEO of IBM, said it hit the "goldilocks spot." Trump signed the revised order in private, without a ceremony and without the invited executives present.

Trump ended up signing it without any ceremony. The spectacle of an Oval Office event with the CEOs of the largest technology companies on earth had been replaced by a quiet signing that generated no photographs and no speeches. The substance that the industry had objected to had been removed. What remained was a framework with no enforcement mechanism, administered through cooperation that companies could withdraw at any time without legal consequence.

The Problem No One in the Room Wanted to Name

There is a structural contradiction at the center of this episode that neither the White House nor the technology industry has been willing to address directly. The same administration that canceled an executive order to avoid burdening AI companies simultaneously used emergency export control powers to shut down Anthropic's Fable 5 model for all non-US nationals within 90 minutes of issuing a directive. The same government that declared it could not accept a 90-day voluntary review window because it might slow down American AI demonstrated that it can and will pull a product used by hundreds of millions of people offline in under two hours when it chooses to.

The message to Anthropic is incoherent: it is too dangerous to trust with defense contracts, but indispensable enough that the Treasury Secretary personally phones bank CEOs to recommend its product. That contradiction is not an oversight. It reflects the reality that the Trump administration is managing at least three competing pressures simultaneously: the demands of the technology industry, which has become one of the administration's most significant political and financial constituencies; the demands of the national security establishment, which is genuinely alarmed by the cybersecurity implications of frontier AI; and the administration's own ideological commitment to deregulation, which sits uneasily with the impulse to impose any oversight on industry at all.

The Trump administration released a long-anticipated executive order aimed at exercising oversight of cybersecurity risks posed by the most advanced artificial intelligence models. The order attempts to thread the needle between the administration's anti-regulatory posture and the recognition that it cannot remain on the sidelines in the face of AI's serious risks to national security. Threading that needle produced an order with no mandatory teeth, signed without ceremony, after the industry removed the provision it found objectionable.

What the Critics Are Not Wrong About

One cybersecurity legal expert said in an interview that giving so much discretion to the NSA director was a "dangerous precedent" that could enable the government to "weaponize" the policy against companies it is clashing with, like Anthropic. That concern points at something real. A voluntary framework administered at the NSA director's discretion, with no clear criteria for which models qualify for review, hands the executive branch a tool that can be applied selectively. A company in good standing with the administration participates on favorable terms. A company in a legal dispute with the Pentagon — as Anthropic is — participates on whatever terms the NSA director decides to offer.

The Trump administration has been divided over how to approach AI. While the Biden White House pushed for federal oversight of the emerging technology, Trump has sought to minimize regulation, including at the state level, even as concerns over safety risks have proliferated. But recently, the development of more powerful AI models has spooked some federal officials, prompting the White House to reverse course and back some safety measures. The reversal produced a framework that satisfied almost no one who had been pushing for substantive oversight. Cybersecurity professionals remained skeptical of a voluntary arrangement with no enforcement mechanism. Researchers who had argued for independent testing requirements got neither independence nor requirements. What the technology industry got was preservation of its release timelines and the elimination of any binding government authority over its products.

What This Moment Actually Reveals

A president of the United States prepared, for six months, a policy governing one of the most consequential technologies in modern history. He invited the chief executives of the companies building that technology to attend its signing. Hours before the event, he canceled it — not because new information emerged, not because a court intervened, not because Congress raised objections, but because the industry had communicated that it did not approve of a specific provision. Two weeks later, with that provision removed, he signed a version the industry found acceptable, in private, without the executives who had originally been invited.

That sequence of events is not ambiguous. It describes an industry that had sufficient political leverage to halt the signing of a presidential executive order on the day of the ceremony, negotiate the removal of its objectionable content, and receive a revised order within two weeks. No other industry in the United States currently possesses that degree of access and influence over White House policy on this timeline.

The framework that resulted may still accomplish something useful. An AI cybersecurity clearinghouse, even a voluntary one, creates a channel for information sharing that did not exist before. A 30-day review window, even without enforcement authority, creates social pressure for participation that zero days of review does not. Whether those marginal gains justify what was traded away — any binding government authority over the most powerful AI systems being built — is a question that the White House, by its actions, has already answered.

Conclusion

The cancellation and renegotiation of Trump's AI executive order is the clearest illustration yet of where authority over American AI policy actually resides. It does not reside in regulatory agencies. It does not reside in Congress, which has produced no AI legislation of significance. It does not reside in the White House, which prepared a policy and then revised it at industry request. It resides, to a degree that should be uncomfortable for anyone who believes democratic governance should set the rules for technologies affecting hundreds of millions of people, with the small number of companies that build frontier AI systems.

That is not a conspiracy. It is a structural outcome of a government that lacks the technical expertise to credibly threaten enforcement, the legislative groundwork to impose binding rules, and the political independence from the industry it oversees to resist pressure when it arrives hours before a signing ceremony. What it does have is the ability to issue emergency export controls, designate companies as national security risks, and use existing statutory authority in ways that are legally contested and procedurally opaque. The result is an AI policy landscape where binding action is arbitrary and unaccountable, while structured oversight remains voluntary and toothless.

That combination is not governance. It is the absence of governance, with a logo on it.