Why the US Is Betting Open-Source AI to Beat China

The Theft Problem That Won’t Go Away

Scott Bessent didn’t mince words before the House Financial Services Committee Tuesday. Chinese AI labs are taking American closed-source models and distilling them into cheaper, capable open alternatives. His translation of that process was blunt: distillation is “a polite, scientific word for steal.” The US Treasury secretary’s prescription wasn’t tighter restrictions or export controls. It was simpler and more disruptive — make more open-source models yourself and flood the market until Chinese versions look like pale imitations.

That framing matters because it signals a fundamental shift in how Washington is approaching AI competitiveness. Instead of treating open-source AI as a benign academic exercise or a cybersecurity risk to be contained, the Trump administration is now positioning it as the primary instrument in a tech Cold War.

The Mythos Benchmark

Bessent cited Anthropic’s Mythos model as proof that open models can leapfrog in capability. Released earlier this year, Mythos was described as a “step change” in the progression toward artificial general intelligence. That reference is deliberate — it gives the administration a credible, commercially available open model to point to rather than relying on academic papers or government research projects.

The implication is stark: if Mythos-level capability can be open-sourced, then Chinese distillation pipelines become less threatening not because they stop stealing, but because the stolen goods are no longer special. When everyone has access to good models, the Chinese advantage in copying evaporates.

This is also a shot across the bow to Anthropic and its closed-source competitors. The administration is signaling that it will privilege companies willing to release capable models openly, while casting closed labs as potential bottlenecks.

Regulatory Capture as the Real Enemy

Bessent’s most loaded phrase may have been “regulatory capture.” He warned that large AI labs could use safety regulations to lock out competition — a concern that cuts both ways. If the government imposes strict pre-release reviews or safety certification requirements on AI models, closed developers with legal teams and compliance departments will absorb the cost. Open-source developers operating at the speed of GitHub will find themselves suffocated.

The letter from Nvidia, Microsoft, Meta, Dell Technologies, Palantir, Hugging Face, Mozilla, Mistral and others, signed in July, made exactly this argument. The coalition called for “targeted legal and commercial frameworks” to address unlawful distillation rather than “sweeping restrictions” that would stifle innovation.

Bessent’s testimony aligns with that position, but with a twist. The administration isn’t just defending open source as a matter of principle — it’s weaponizing it. Open-source AI is now a national security strategy.

What Happens to Closed-Model Startups

The policy pivot creates an immediate question for every startup building on closed models: where do you stand relative to the government’s new thesis?

Companies like OpenAI, which has confirmed it won’t go public in 2026 amid safety concerns, face a dilemma. Sam Altman’s caution about openness and accessibility is now structurally at odds with the administration’s competitive framework. If Bessent’s logic holds, OpenAI’s closed approach is a liability — not just commercially but politically. The company has already lost ground on the regulatory front: the Trump administration’s voluntary AI framework released in August exempted open-source and open-weight models from pre-release security reviews, focusing scrutiny instead on proprietary systems.

That exemption is a gift to open-source developers and a tax on closed labs. It means open models ship faster, iterate faster, and reach users first — exactly the conditions under which distillation becomes a losing proposition for any government.

The Distillation Arms Race

The distillation pipeline works like this: take a powerful closed model, run inputs through it, collect the outputs, and use that data to train a smaller, open model that mimics the behavior. The result is a capable open-source alternative that costs a fraction to build. Chinese labs have used this method extensively, producing models that increasingly resemble their American parents.

Bessent’s observation that many Chinese models “think they’re Mythos, they think they’re Claude” is both an indictment and an opportunity. It confirms the theft is working — but it also suggests the theft is running out of fresh material if American open models continue to improve faster than Chinese labs can copy them.

The administration’s strategy assumes that open development cycles are faster than closed ones. In a race where the winner is defined by who ships the better model first, openness wins. The question is whether that assumption holds under commercial pressure. Open-source models are cheaper to build but often lag behind closed models in raw capability. The gap narrows over time — but not always fast enough.

Who Wins and Who Loses

The winners are clear: open-weight model developers like Mistral and Hugging Face, which operate in the new regulatory bright zone. Academic labs building on top of open models. Researchers and engineers who want to inspect codebases and modify training parameters. The US government, which gains a strategic tool against Chinese AI without needing to embargo hardware or restrict talent.

The losers are more complicated. Closed-model startups face increased regulatory scrutiny and a policy environment that treats their caution as obstruction. Chinese AI labs lose their copying advantage if US open models outpace their distillation pipelines. Even some American closed labs may find themselves squeezed — Nvidia and Microsoft appear to be hedging by supporting both open and closed ecosystems, but the administration’s tilt is unmistakable.

There is also a risk that the policy underestimates the value of closed models. Security reviews, controlled access, and responsible disclosure matter for models that can be weaponized. The administration’s voluntary framework draws a bright line between open and closed that may not hold in practice — a sufficiently capable open model is, after all, almost indistinguishable from a closed one to someone who knows how to use it.

What Comes Next

The next few quarters will test whether Bessent’s thesis can survive contact with industry reality. Watch for two developments: whether the administration expands the open-source exemption beyond pre-release reviews, and whether closed-model companies mount a coordinated response to what they’ll likely call discriminatory regulation.

Anthropic will be the canary. Its Mythos release was designed to prove that open models can compete at the frontier. If subsequent open models from Anthropic and other developers fail to stay ahead of Chinese distillation, the administration’s strategy loses its intellectual foundation. If they succeed, closed labs face an increasingly hostile political environment.

The broader arc is simpler than either side admits. The US government has decided that openness is the only competitive strategy that scales against a rival willing to copy. Whether that bet pays off depends on whether open models can keep improving faster than anyone can steal them — including, eventually, the companies that created them.