Who Gatekeeps AI? The Question Washington Just

Alphabet just reported $109.9 billion in quarterly revenue—22% year-on-year growth, $62.6 billion in profit, Google Cloud hitting $20 billion at a 63% growth rate that outpaced both AWS and Azure. These are numbers that make you recalibrate what "a good quarter" means. AI drove results across the entire Google ecosystem, which is either an extraordinary vindication of Sundar Pichai's long bet on AI infrastructure or a reminder that when you're Google, you can afford to lose several bets before the one that pays off does so spectacularly.

The earnings story is real. But the more consequential story breaking alongside it concerns what the U.S. government intends to do about the technology producing those numbers.

The Mythos Moment

The Trump administration, which came into office with an explicit posture of clearing regulatory brush from AI's path, is now reportedly considering a pre-release vetting process for frontier AI models—a working group of tech leaders and government officials that would preview capabilities before they reach the public. The pivot is notable not just as a policy shift but as a signal about what specifically changed the calculus.

On the Diamandis podcast, computer scientist Dr. Alexander Wissner-Gross offered the clearest account of why: the release of Claude Mythos. "I think we'll view the Mythos moment as a sea change," he said, "when the civilian sector, the frontier AI labs, suddenly had capabilities that leapfrogged government capabilities—where specifically Mythos was suddenly able to... get vulnerability discovery getting effectively solved for some definition of solved by AI for the first time in the private sector, leapfrogging what possibly the NSA or other government agencies had internally."

That's a significant framing. The concern isn't that AI is powerful in the abstract—Washington has been told AI is powerful for years. The concern is that a specific, measurable capability—the automated discovery of software vulnerabilities—appeared to cross a threshold where private companies demonstrably surpassed what intelligence agencies could do internally. When that happens, the question of who controls release timing becomes a national security question, not just a product question.

Wissner-Gross also noted, with characteristic precision, that GPT-5.5—which is actually generally available, unlike Mythos—appears to match or exceed Mythos on public cybersecurity benchmarks. The implication: the capability the White House is reacting to may already be more widely distributed than the policy discussion assumes.

The Oligopoly Trap

Here's the regulatory design problem that almost always gets buried under the security arguments: compliance costs don't scale linearly. They scale with the ability to absorb them.

A pre-release vetting regime—whatever form it takes—creates a compliance infrastructure that OpenAI, Google, and Anthropic can staff and fund. A lab with 40 engineers and $80 million in runway cannot. The framework that looks like oversight of the big players functions, in practice, as a barrier to entry for everyone else.

Wissner-Gross pushed back gently on the novelty of this concern, noting that export controls, the Invention Secrecy Act, and the Atomic Energy Act have functionally gated sensitive technologies since World War II. "It's not as if we're suddenly entering some brave new world," he said. That's a fair historical point. But it's also worth noting that those prior regimes were designed around state actors and discrete physical technologies—not software capabilities that can be replicated and distributed globally at near-zero marginal cost.

Brian Elliott, the West Point graduate and Blitzy CEO who joined the conversation, drew a distinction that matters: there's a difference between veto rights and previewing. "It has to [preview], but it can't gatekeep. That's where we're going to end up falling behind from a geopolitical perspective." The concern is China—not as a rhetorical flourish, but as a real constraint on how much friction the U.S. can introduce into its own AI development pipeline before the competitive calculus shifts.

The More Interesting Regulatory Question

The panel's most counterintuitive argument—and the one I find most worth sitting with—came from Wissner-Gross when he identified where he thinks the actual gatekeeping risk lies.

"I'm more worried about the Frontier Labs self-policing more aggressively than the government ever would and stifling competition that way," he said. "I think that's a far scarier future."

The mechanism he describes is straightforward: labs hold back their most capable models—either because they're too compute-intensive to deploy broadly, or because the commercial advantage of running them internally outweighs the revenue from licensing them. The government then reviews a derivative. The public gets a further-downstream version. Meanwhile, the frontier capability remains concentrated at the lab that produced it, insulated from competitive pressure.

This is not a hypothetical. The frontier labs already maintain internal model versions that are more capable than what they publicly release. That's been publicly documented. The question is whether a vetting regime accelerates this dynamic by giving labs a legitimate procedural reason to hold capability back—"we're working through the government review process"—while simultaneously building competitive moats.

The historical analogy that's underappreciated here: the early history of GPT-2 and GPT-3. OpenAI staged the release of GPT-2 citing safety concerns; the decision was later viewed by many observers as having been more cautious than the actual risk warranted, and Dario Amodei—then at OpenAI—was associated with that caution. The "moral panic" framing, as Wissner-Gross put it, has precedent. The question isn't whether the panic is ever legitimate—sometimes it is—but who holds the authority to act on it, and what incentives they're operating under when they do.

Pentagon Contracts and the Labor Question

The same week as the earnings report, the Pentagon formalized agreements with seven AI companies—Google, OpenAI, Microsoft, Amazon, xAI, and others—for military applications. Google's agreement covers AI provision "for any lawful government purpose."

The employee response was instructive in its contrast to 2018's Project Maven walkout, which drew 20,000 Google employees into the streets. This time, roughly 600 employees protested. More significantly, the protest came primarily from DeepMind employees in London—who subsequently unionized, a development Wissner-Gross called "a bizarre juxtaposition of 19th century union-style organization" with 21st century frontier AI. DeepMind's cultural distinctness from Google proper, and its geographic distance from U.S. patriotism arguments, makes it a particular flashpoint.

Salim Ismail, joining from Toronto where he was dealing with the thoroughly analog problem of a passport running out of pages, offered the cleaner frame: "AI is not just a tool now. It's becoming a decision layer." That's the crux of why military contracting feels different now than it did seven years ago. In 2018, the argument was about applying AI to drone footage analysis. The current agreements are with companies whose models can, in principle, autonomously identify and exploit software vulnerabilities across adversary infrastructure. The employees aren't wrong that the stakes have changed.

What the Earnings Numbers Actually Reveal

Return briefly to the $109.9 billion. Google Cloud's 63% growth, outpacing AWS and Azure, is partly a function of enterprise AI demand—companies building on Gemini, buying TPU compute, integrating Google's AI APIs into their workflows. The growth rate suggests that demand is not yet saturating. It also suggests that whoever controls the inference infrastructure for frontier models controls a strategically significant choke point.

Which is precisely why the question of who gets to deploy what model, when, and under what conditions is not just a technology policy question. It's an infrastructure policy question. The analogy to spectrum allocation or pipeline right-of-way isn't perfect, but it's closer than the AI discourse usually acknowledges.

The White House's proposed vetting regime—assuming it takes the form of a working group rather than a formal regulatory process—would likely lack the statutory authority to do much beyond delay. That's not nothing. A 90-day review window, as one panelist noted, roughly matches the gap between closed-source frontier releases and comparable open-source capabilities. The practical effect would be to compress that window further, potentially collapsing the distinction between open and closed source at the capability frontier.

Whether that's the intended effect, or a side effect nobody in the working group thought through, is exactly the kind of question worth watching as the executive order takes shape.

The administration that promised AI companies maximum speed just blinked. The reason—a private lab's cybersecurity capability apparently exceeding what the NSA had—is either the most honest admission about where capability concentration stands, or a preview of the argument that will be made every time a new capability threshold gets crossed. Probably both.

Samira Okonkwo-Barnes covers technology policy and regulation for Buzzrag.