The coffee in the basement of the Eisenhower Executive Office Building is notoriously bad, tasting faintly of paper cups and institutional exhaustion. It was late autumn, the kind of Washington D.g. evening where the damp cold settles into your bones, and a small group of civil servants sat staring at a projector screen. For six months, they had lived on three hours of sleep a night. They were drafting an executive order, a massive, Byzantine piece of regulatory architecture designed to do something no government had ever successfully managed: put a harness on artificial intelligence before the horse had completely bolted from the barn.
They felt like historical protagonists. They believed they were protecting the republic.
Then the phone calls started.
By midnight, the meticulously crafted thresholds for what constituted a "dangerous" AI model—numbers debated over countless boxes of cold pizza—began to shift. Not by a lot. Just enough to ensure that the largest software conglomerates on the planet could keep building exactly what they wanted, precisely how they wanted, while their smaller, open-source competitors bore the brunt of the compliance paperwork.
This is not a story about code. It is a story about access, whispered warnings in marble hallways, and the terrifying speed with which a multi-trillion-dollar industry can bend the knee of the state. We were told the Trump administration’s swift dismantling and rewriting of AI regulations was a populist victory for American innovation.
It wasn't. It was a masterclass in corporate capture.
The Illusion of the Empty Room
To understand how big tech got its way, you have to understand a fundamental law of Washington physics: power abhors a vacuum, but it loves a technicality.
When the conversation around AI governance shifted from the Biden administration's sweeping, safety-first executive order to the incoming Trump transition team's pro-growth agenda, the public narrative was simple. The old guard wanted to choke AI with red tape; the new guard wanted to set it free. It played perfectly on cable news. It was a clean, ideological battle.
But inside the room where the policy was actually being rewritten, ideology was a luxury no one had time for.
Instead, there were metrics. Specifically, compute thresholds.
Under the original framework, any company training an AI model using more than $10^{26}$ floating-point operations—a massive amount of computational power—was required to notify the government and share their safety testing data. It was a tripwire. The moment you built something too powerful, a flag went up in Washington.
For a brief moment, the public thought the tech giants were terrified of this rule. They weren't. They engineered it.
Consider the perspective of a mid-level policy staffer trying to understand this world. You are sitting at a desk, tasked with regulating a technology that changes every fourteen minutes. You do not have a PhD in computer science from Stanford. You do not have a supercomputer cluster in your basement. Who do you call to find out if $10^{26}$ is the right number?
You call the people who own the supercomputers.
And they were more than happy to answer. The tech giants didn't fight the idea of a threshold; they just helped draw the line exactly three inches above their own current capabilities, while ensuring the compliance costs of reaching that line would bankrupt any upstart rival. They didn't break the regulatory machine. They bought the company that manufactures the gears.
The Currency of Fear
The persuasion campaign did not rely on complex economic theories. It relied on a primal, deeply American emotion: the fear of coming in second place.
Every tech executive who walked into the transition team’s offices carried the same deck of slides. The charts varied, but the message was uniform. If you slow us down—even by a week, even by a day—Beijing wins. It was a brilliant, devastatingly effective rhetorical pivot. By framing safety checks not as a shield for the public, but as a handicap in a digital Cold War, the industry transformed itself from potential corporate villains into national security assets.
Suddenly, asking for algorithmic transparency wasn't just bureaucratic; it was borderline unpatriotic.
Let’s look at how this played out in the text of the revised executive orders and subsequent agency directives. The emphasis shifted entirely from "biosecurity risks" and "algorithmic bias" to "compute dominance" and "national defense integration." The language became martial. The state was no longer a referee monitoring a dangerous sport; it was a booster club funding the home team.
But who comprises the home team?
It isn't the twenty-something engineer in a garage in Austin trying to build a more efficient medical diagnostic tool. That engineer cannot afford the lobbying firms that swarm K Street like locusts. That engineer doesn't have a direct line to a senator's chief of staff.
When the administration agreed to lift restrictions on commercial AI development under the banner of national sovereignty, they weren't liberating the market. They were building a moat. The massive cost of computing power already created a natural monopoly; the new regulatory landscape, stripped of its public-interest guardrails but heavy on corporate-friendly procurement contracts, turned that monopoly into law.
The Open Source Sacrifice
The most tragic casualty of this quiet coup was not the federal oversight. It was the open-source community.
For decades, the best security on the internet has come from transparency. When code is open, anyone can look at it, find the flaws, and fix them. It is a messy, democratic, wildly successful ecosystem. In the early days of the AI boom, open-source models were progressing at a staggering rate, threatening to democratize access to these powerful tools.
This terrified the boardrooms of Silicon Valley. If anyone can run a powerful AI model on a home computer, how do you charge twenty dollars a month for a subscription? How do you justify a trillion-dollar valuation to Wall Street?
The solution was ingenious. The major tech firms used the government’s anxiety to launch a preemptive strike on open-source software. They argued that because open-source models cannot be recalled once they are released into the wild, they represent an inherent proliferation risk. They raised the specter of rogue actors using open-source code to engineer bioweapons or launch catastrophic cyberattacks.
It was a brilliant piece of misdirection.
The administration swallowed it whole. The revised policies introduced subtle, lethal hurdles for open-source developers. While the tech giants secured exemptions for their proprietary, cloud-hosted models under the guise of proprietary national security secrets, independent developers found themselves facing vague liabilities.
The message was clear: if you want to build AI, you must do it inside the walls of an approved corporate fortress.
The Human Cost of a Soft Document
It is easy to get lost in the jargon of compute power, algorithmic weights, and regulatory frameworks. It all sounds so bloodless. But choices made in those carpeted Washington rooms have a friction that eventually rubs human skin raw.
Imagine a woman named Sarah. She lives in a midwestern suburb, working in the billing department of a regional hospital group. She doesn’t know what $10^{26}$ floating-point operations means. She has never heard of the executives who spent millions lobbying the transition team.
But one Tuesday morning, Sarah arrives at work to find her department has been streamlined. An AI system, leased from one of the massive tech firms that helped draft the new federal guidelines, now handles the claims. The system makes errors—hundreds of them—denying coverage to patients who desperately need it.
Sarah tries to flag the errors. She points out that the algorithm is systematically misinterpreting a specific code for oncology treatments.
Under an earlier, consumer-protection-focused iteration of the AI executive order, there were mechanisms for accountability. There were mandates for regular audits, avenues for workers to report systemic flaws without fear of retaliation, and requirements that companies prove their models did no harm before deployment.
Now? Those provisions are gone, replaced by a non-binding framework of "industry-led best practices."
Sarah’s supervisor tells her to drop it. The hospital group signed a multi-year contract with the tech giant. The system is proprietary. The code is a black box, protected by federal trade secret exemptions that the tech company argued were necessary to keep their intellectual property out of foreign hands. Two weeks later, Sarah's position is eliminated entirely.
Sarah is not a statistic in a debate about national competitiveness. She is the collateral damage of a policy written by the people it was meant to regulate.
The Quiet Room
The ink is dry now. The press releases have been archived, the talking heads on television have moved on to the next crisis, and the tech stocks have enjoyed their predictable, post-regulatory bump.
If you walk through the corridors of the agencies tasked with overseeing this frontier, you can feel the shift. It is the silence of surrender. The young, idealistic lawyers who joined the government thinking they would be the ones to keep the digital titans in check are quietly updating their LinkedIn profiles. They are realizing that the most lucrative career path isn’t fighting the giants—it’s joining them to help interpret the very rules they just helped relax.
We were promised an era of unprecedented technological liberation, a grand unleashing of American ingenuity that would secure our place at the apex of the global economy.
Instead, we got something much older and far more familiar. We got a closed door. We got a handful of men in tailored suits deciding what we are allowed to see, what we are allowed to know, and who is allowed to profit from the future.
The lights are still on in the Silicon Valley campuses, and they are still on in the West Wing. The servers are humming, consuming more electricity than small nations, churning through data, growing more powerful with every passing second. There is no one left to turn them off. There is no one left who even knows where the switch is.
