Trump's AI Safety U-Turn and Anthropic's Unlikely Musk Deal

The administration that literally deleted the word "safety" from the US AI Safety Institute's name is now signing agreements to safety-test AI models before they ship. That whiplash is real, and it tells you everything about what actually moves the needle in Washington.

For context: when Trump took office, one of his first moves in the AI space was rebranding the Biden-era US AI Safety Institute as the Center for AI Standards and Innovation, or CAISI. The message was clear — safety talk was out, innovation was in. Voluntary testing requirements from the Biden era got quietly shelved as overreach.

Then Anthropic spooked everyone. The company announced it was sitting on a model — Claude Mythos — that it considered too dangerous to release, citing its advanced cybersecurity capabilities and the very real risk of bad actors getting their hands on something that powerful. Suddenly, the White House had an opinion on AI safety again. National Economic Council Director Kevin Hassett has indicated Trump might even issue an executive order mandating government testing of frontier models before release.

This week, CAISI formalized agreements with Google DeepMind, Microsoft, and xAI to conduct exactly those kinds of pre- and post-release evaluations. The press release from CAISI's director, Chris Fall, acknowledged the new partnerships "build on" the Biden-era policy — a quiet admission that the previous administration wasn't entirely wrong.

The substance of CAISI's testing program is actually more interesting than the politics. The agency says it has completed roughly 40 evaluations to date, and it frequently tests models with safeguards removed or dialed down, which gives evaluators a clearer picture of what a model is truly capable of when the guardrails are stripped away. A newly formed interagency task force is also feeding national security priorities directly into the evaluation process.

On the industry side, reactions have been measured but positive. Google DeepMind's VP of frontier AI global affairs called the plans encouraging. Microsoft framed government collaboration as an obvious necessity for risks at national scale. xAI, meanwhile, didn't respond to press inquiries — which is notable given the company is currently tangled up in a very public legal dispute with OpenAI over which firm's leadership actually cares about safety.

The skeptics, though, are far from satisfied. Critics are raising a structural problem that no press release really addresses: if AI companies keep their model architectures and training data largely secret, how rigorous can any external evaluation actually be? Testing a black box is a fundamentally different challenge than auditing open systems, and CAISI's 40 completed evaluations cover a tiny slice of what's being deployed commercially right now.

What this moment really represents is a government scrambling to catch up after spending months signaling it wouldn't bother. The Mythos announcement forced the issue in a way that abstract safety arguments never quite managed. Whether the resulting framework has any teeth is the question nobody in Washington seems eager to answer directly.

Earlier this year, Elon Musk called Anthropic's policies "misanthropic" and "evil" on X. This week, he announced he had spent quality time with their senior leadership and came away impressed. That is either a remarkable change of heart or a very telling sign of how badly the AI industry needs computing power right now.

The deal: Anthropic has signed an agreement to tap into xAI's Colossus 1 supercomputer, housed in a former Electrolux factory in Memphis, Tennessee. The arrangement gives Anthropic access to more than 300 megawatts of capacity backed by roughly 220,000 Nvidia GPUs — a mix of H100s, H200s, and the newer GB200 chips. That is not a rounding error. That is a serious chunk of the world's available AI compute.

The announcement came at Anthropic's annual developer conference in San Francisco, which is a deliberate choice of venue. This wasn't buried in a filing — Anthropic wanted the world to know it had solved, or at least significantly eased, its compute problem. The company has been getting hammered by developers for rate limits and service outages on Claude Code, its increasingly popular coding assistant. As more users run long agentic tasks that chew through GPU cycles for hours, the cracks have been showing.

A quick note on the entity Anthropic is actually dealing with: SpaceX and xAI completed a merger earlier this year, forming a combined company called SpaceXAI, still under Musk's ownership. So when Anthropic taps the Colossus supercomputer, it is working directly with Musk's consolidated tech-and-space empire. The awkwardness of that arrangement, given the very recent public insults, is not lost on anyone paying attention.

For SpaceXAI, the timing is equally strategic. The company is eyeing a public offering as soon as next month, and landing Anthropic as a compute customer is a credibility boost that money can't easily manufacture. Investors being pitched on the vision of orbital data centers — yes, actual supercomputers in space — will feel a lot better hearing that a leading AI lab is already a paying customer on the ground-based version.

That orbital ambition is worth pausing on. SpaceXAI's blog post about the deal noted that Anthropic has "expressed interest" in eventually tapping space-based compute capacity. Whether that ever materializes is an open question, but the signal it sends to IPO investors is obvious: this isn't just a data center company, it's a platform.

There are complications worth tracking. Colossus 1's gas turbines have drawn environmental complaints from Memphis residents, and protests have already materialized near SpaceXAI investor events. Anthropic, a company that positions itself as the safety-conscious adult in the AI room, is now operationally linked to a facility facing environmental scrutiny. That tension will not quietly resolve itself.

For now, though, the compute math wins. Claude Pro and Claude Max subscribers are the immediate beneficiaries, and a smoother Claude Code experience will matter far more to most users than the corporate drama surrounding how it got built.