Here's a sentence I didn't expect to write in 2026: the U.S. government is now deciding, on a customer-by-customer basis, which people get access to frontier AI models. Let that sink in for a moment.

Two weeks after federal authorities pulled Anthropic's Fable and Mythos models from general availability, OpenAI's GPT-5.6 is reportedly headed for the same bureaucratic purgatory. According to reporting from The Information, the Trump administration has asked OpenAI to stagger the model's release, with approvals happening one customer at a time until some undefined green light for general availability is granted. Sam Altman apparently projected a "couple of weeks" of preview limbo. Which is cute. Mythos has been stuck in preview for months with no clear path out.

So we've arrived at a peculiar moment: two of the most powerful AI labs on the planet, fierce competitors with billions of dollars and enormous egos invested in beating each other, are now sitting in the same government waiting room with the same problem. The industry chatter, predictably, has been about who's to blame—is Anthropic running a regulatory capture play? Is OpenAI trading political coziness with the Trump administration to freeze out rivals? These are entertaining theories, and plenty of people with nine-figure stakes in one company or the other are happy to promote them.

But this framing misses the actual crisis developing in plain sight.

When the Bottleneck Isn't Your GPU Cluster

AI labs have spent the last several years obsessing over the right bottlenecks: compute, data quality, inference efficiency, alignment. Nobody seriously modeled "government approval latency" as a variable in their revenue forecasts. And yet here we are, with regulatory review threatening to eat directly into the economic return on billion-dollar training runs.

Think about what that means structurally. These models aren't cheap. The compute cost of training a frontier model, the data center infrastructure, the inference at scale—the numbers are staggering. That investment only makes sense if you can monetize the capability relatively quickly after release. A months-long preview window doesn't just delay revenue; it potentially kills the business case for the next training run entirely. And if labs start pulling back on model development, the data center buildout that's been driving enormous capital expenditure across the industry starts to look a lot less certain.

The entire supply chain—chip manufacturers, cloud providers, power infrastructure—has priced in aggressive AI growth. A regulatory choke point at the model release stage doesn't stay contained to the labs. It propagates upstream fast.

The Government Doesn't Actually Know What It's Testing For

Here's the part that should concern you most: there's no coherent framework here. Testing products before public release is a perfectly normal regulatory approach—we do it with pharmaceuticals, with aircraft, with medical devices. But those frameworks work because regulators have defined the failure modes they're trying to prevent, they've built the expertise to evaluate them, and they've established clear criteria for approval.

None of that exists for frontier AI models right now. The U.S. government has neither the technical staff nor the institutional knowledge to conduct meaningful safety evaluations of a model like GPT-5.6 or Mythos. More fundamentally, it hasn't articulated what specific risks it's actually worried about. Cybersecurity uplift? Bioweapon synthesis assistance? Political influence operations? These are genuinely different threat models requiring entirely different testing methodologies. Without defining the target, you can't build a useful test—and without a useful test, "approval" is just bureaucratic theater that delays releases without making anyone safer.

GMU fellow Dean Ball laid this out clearly in a recent post, and he's right: the absence of a coherent risk framework isn't a minor implementation detail. It's the whole problem. You can't satisfy a regulator who can't tell you what they need to be satisfied about.

The Real Concerns Don't Disappear Just Because the Process Is Broken

That said—and I want to be careful here, because the temptation is to declare the entire government concern illegitimate and move on—there are actual risks worth taking seriously underneath this bureaucratic mess.

The evidence that advanced AI tools are meaningfully accelerating certain categories of cybersecurity threat is real. The concerns about biological risk are not science fiction; researchers at places like CSET have done serious work mapping the threat surface. Alignment remains an unsolved problem in ways that matter at capability levels these models are approaching. Dismissing all of this as pretextual government overreach is as intellectually dishonest as pretending the current approval process constitutes serious risk mitigation.

The question isn't whether to address these risks. It's whether locking models in regulatory purgatory with no clear criteria actually addresses them—or just limits what's available to legitimate users while doing nothing about the actual threat vectors.

What Would Actually Help

The constructive path—and I'll be honest, this requires optimism I'm not sure is warranted—involves the industry doing something it's historically terrible at: collective action in good faith.

That means getting behind independent technical bodies that can actually build evaluation frameworks, even if those bodies make decisions you don't love. It means accepting that some regulation is inevitable and fighting for the least-bad version rather than burning every bridge in the hope of avoiding all of it. And most importantly, it means OpenAI and Anthropic and Google and the rest recognizing that they have a shared interest in a functional regulatory environment—not as competitors trying to use policy as a competitive weapon, but as an industry that needs a workable framework to exist.

For a lot of people in AI, whose entire career identity is built around moving fast and breaking things, this is going to be a genuinely difficult cultural shift. The frontier model business has reached a scale where its outputs have real political consequences, and political consequences require political responses. The industry didn't ask for this moment. But it's here anyway.

Whether the major labs can actually coordinate on something this uncomfortable—rather than quietly hoping the regulatory burden lands harder on a competitor—is the question that will define the next few years of AI development. Don't bet the house on altruism. But the math on collective action is pretty hard to ignore when everyone's stuck in the same waiting room.