By Benjamin Guggenheim

Not a subscriber? Sign up here to get this newsletter in your inbox.

• National security experts argue that Chinese open-source models are fundamentally less secure than the current U.S. models.
• The venture capitalists close to the administration, on one hand, and Anthropic and OpenAI, on the other, have an enormous amount at stake in the regulation of open-source models.
• The increased competitiveness of open-source does not bode well for the enormous valuations of Anthropic and OpenAI.

This is Benjamin Guggenheim. Welcome to WP Intelligence’s AI & Tech Brief, where we examine the transformative technology of artificial intelligence at the intersection of innovation, policy and power. Get in touch at Benjamin.Guggenheim@washpost.com or on Signal at ben_guggenheim.88.

The Lead Brief

The release of the best Chinese open-source AI model to date, Moonshot AI’s Kimi K3, has exposed a major tension between two structural trends in AI — the rise of open-source to challenge OpenAI and Anthropic’s effective duopoly on frontier AI models, on one hand, and the U.S. government’s increasingly urgent imperative to manage catastrophic risk on the other.

The tension was exposed in sharp relief over the weekend by a spat between venture capitalist David Sacks, the co-chair of the President’s Council of Advisors on Science and Technology, and Dean Ball, a widely followed AI policy analyst who is just a few weeks into his new job as the head of strategic futures at OpenAI. (The Washington Post has a content partnership with OpenAI.)

“I would guess that the Trump administration will at some point realize that their best strategy here would be to create large amounts of regulatory risk around the use of open-weight Chinese models,” said Ball in an X post, adding that this regulatory stance “needn’t be that well justified.”

This earned a rebuke from Sacks, who suggested that Ball may be advancing a “regulatory capture” strategy to shut out open-source companies from competing with OpenAI.

“The leading closed labs … want the government to eliminate their open source competition,” Sacks said on X, asserting that “regulatory decisions should always be well justified and grounded in facts, logic, and evidence.”

Department of Defense Undersecretary Emil Michael — who has been spearheading an effort to introduce venture capital-style procurement of defense tech at the Pentagon — also accused Ball on X of advancing a “deep state regulatory capture scheme.”

Ball later clarified that he was making a prediction, not asserting an argument about what should happen. He also said in a follow-up post on X that he has a “deep and abiding affinity for open-source software.”

Ball did not immediately respond to a request for additional comment.

What it means: Venture capitalists (and the start-up community they invest in) like open-source models.

Open-source AI models that are free and democratized lower the cost of building AI applications and prevent the capture of margins that happen when only two huge companies control the frontier AI resource.

Proponents of open-source also believe that Anthropic and OpenAI want to kill open-source models in the crib precisely because open-source models threaten the companies’ duopoly in the AI model layer.

“The anti-Chinese models political campaign is led by Anthropic,” wrote Nathan Lambert, an open-source advocate, in his widely read AI newsletter Interconnects. “Anthropic would gain substantial economic security in its products if the Chinese model makers they accused were banned.”

Spokespeople for Anthropic and OpenAI did not immediately respond to a request for comment.

Venture capitalist David Sacks speaks on stage the Republican National Convention in 2024. (Photo by Ricky Carioti/The Washington Post)

The regulatory question

The increased capability and competitiveness of open-source models are running headlong into the unresolved question of how the U.S. government should regulate these AI models.

The ad hoc directive, issued by the Commerce Department in June, which forced Anthropic to retract its most advanced model showed that the Trump administration is worried about cyber risks posed by these models —but has not yet developed a clear procedure or benchmark that would give the administration confidence that a model is safe to release to the public.

It is, of course, much more difficult for the U.S. government to regulate open-source models. If they come from China, the U.S. can’t dictate the terms of their release. If they’re developed in the U.S., the open-source models are still diffused much more rapidly and can be “fine-tuned” by bad actors.

There’s also another distinct threat posed by Kimi and the other leading Chinese open models, such as DeepSeek and Alibaba: that Chinese AI developers have embedded security vulnerabilities into the models’ architecture, so that they can be exploited once they become embedded in American companies and critical infrastructure.

“The Commerce Department’s [Center for AI Standards and Innovation] has put out four reports on Chinese AI models, where they include assessments of their capabilities and some of their risks, particularly prompt injection, agent hijacking, [and] security vulnerabilities writ large,” said Daniel Remler, an analyst at the Center for a New American Security.

“They make them paint a pretty compelling case that these models are significantly more insecure than the U.S. models,” said Remler, who also argued that the government should develop minimum testing standards that the models should pass before they’re integrated into U.S. infrastructure.

The national security question

Open-source is a competitive threat to the U.S. frontier AI models. There are catastrophic biological and cyber risks posed by frontier models, both open and closed. And there are unique national security risks posed by open Chinese models. The administration has not developed a substantive, transparent procedure for dealing with these risks.

But there is significant disagreement as to whether the U.S. government and the ecosystem at large are capable of making open-source models safe enough, whether through regulatory action or technical breakthroughs.

“I think it’s pretty clear that we are approaching the point I describe — the point where, absent a major technical safety breakthrough, the national security implications of frontier open-weight model distribution are simply too severe,” said Ball in a follow-up post on X. “Governments will realize these risks eventually, and when they do, they will have much lower risk tolerance.”

Not everyone disagrees with Ball’s assessment.

Kristian Stout, director of innovation policy at the International Center for Law & Economics, told me in an interview Monday that he believes, from a catastrophic risk perspective, banning open models hurts more than it helps.

Stout argues specifically that you need frontier AI defense capabilities to defend against AI-powered cyberattacks.

“I need an advanced AI agent that’s sitting on my computer that goes out and scans the dark web all day long looking for people who are trying to use my identity or compromise me,” argues Stout. “That only happens when you have more intelligence diffusing.”

The outlook: From an economic perspective, though, most AI industry analysts agree that the rise of open-source poses a very real threat to the leading U.S. labs.

If open-source survives, whether in the form of U.S. or Chinese models, the enormous valuations of OpenAI and Anthropic become more difficult to justify. Moonshot AI’s Kimi was ranked as better at coding than those other companies’ models, and most companies don’t need the very best model for everyday tasks anyway.

Ball and Samuel Hammond, the director of artificial intelligence policy at the Foundation for American Innovation, also make the point that open-source models are inherently “decelerationist”: They cut into the gross margin that frontier labs would have used to reinvest back into AI model training. That could slow down under open-source competition.

And, in the latest warning sign for the U.S. frontier labs, Chinese multinational Alibaba announced on X Sunday that it would be launching its newest model, Qwen 3.8, soon. The model has 2.4 trillion parameters and would be second only in capability to Anthropic’s Fable 5 model, the post said.

“Value shifts in the stack. It’ll shift away from OpenAI and Anthropic. It’ll shift toward value-added companies, which is what you have seen in every technical revolution where there’s been a general-purpose technology like electricity,” argued Stout.

“We live in a world now where electricity is just distributed, and the real value of electricity is what all the complement producers do with electricity,” Stout said. “I believe that is what the intelligence revolution is going to be like.”

The most prominent obstacle getting in the way of that economic future is regulation. And the size and scope of the regulation will be determined by which arguments about catastrophic and national security risk win out.

This newsletter is published by WP Intelligence, The Washington Post’s subscription service for professionals that provides business, policy and thought leaders with actionable insights. WP Intelligence operates independently from The Washington Post newsroom. Learn more about WP Intelligence.