Chinese Open-Weight Models Threaten to Become the Default AI Stack

Michelle Giuda argues that Silicon Valley and Washington are talking past each other. The visible dispute is whether open-weight AI models should remain broadly available. Her narrower concern is that Chinese open-weight models are, in her view, currently among the most accessible and cost-effective options for businesses building AI systems—and that the United States lacks a comparably robust domestic alternative.

The July 24 letter, Open Weights and American AI Leadership, makes the affirmative case for keeping such models available. It defines open weights as models that organizations can download, inspect, modify, and run on their own infrastructure. The letter argues that startups, universities, businesses, and public institutions can build on advanced models without training from scratch or paying frontier-model prices for every task. Its economic logic is to reserve frontier-scale systems for “genuine frontier problems” while using efficient, specialized models across a far larger set of routine tasks.

The letter’s broader standard for American leadership is diffusion rather than ownership of a single leading model: whether the United States develops an open ecosystem that reaches “factories, hospitals, farms, classrooms, and main street businesses.”

Giuda does not dispute the commercial case. Cheaper models that are “good enough,” she says, have become central to how companies assemble their AI stacks, particularly as the cost of the AI buildout draws scrutiny. She cited a recent $890 billion technology-market wipeout in describing concern over whether the industry can sustain its spending.

For Giuda, the practical question is therefore not whether companies value open weights. They do. It is whether low-cost, deployable American models can reach them before Chinese models become embedded in the systems they build around.

Moonshot AI’s Kimi K3 launch, Giuda says, brought that gap into sharper view. In her framing, the policy challenge is to get an American alternative into the market quickly enough that price and availability do not steer businesses toward Chinese models by default.

Ed Ludlow frames one concern behind the letter as fear that Washington could overregulate open-source or open-weight systems and weaken American competitiveness. The displayed posts from Satya Nadella and Nvidia’s Jensen Huang state the pro-open-model position plainly.

Nadella wrote that open-weight models are “essential to a healthy AI ecosystem” and could strengthen American competitiveness and expand economic opportunity while protecting national security. Huang wrote that open models strengthen safety and cybersecurity, accelerate innovation and diffusion, and enable sovereignty; in his formulation, the world needs both frontier closed models and frontier open models.

Giuda accepts that openness has a strong innovation and commercial case. But she says the calculus changes when the least expensive and most available options come from China. Businesses, she argues, have spent much of the past decade trying to decouple and derisk from China over supply-chain, financial, corporate, and national-security exposure; embedding Chinese technology in the foundation of American companies’ AI stacks would reverse that direction.

She cited U.S. actions involving Huawei and TikTok, alongside legislation she said was moving through the House on Chinese connected vehicles, as examples of a broader pattern of concern about Chinese technology. The policy choice, in her view, is not simply whether to ban models or preserve openness. It is whether trusted U.S. models can be made cheap enough, capable enough, and available enough to compete for adoption.

That standard extends beyond domestic access. Giuda calls for PCAST, which she described as a presidential advisory board including technology leaders, to move quickly on American open-weight AI and to bring allies into the effort. But she identifies no specific financing, regulatory, or technical mechanism for making those models more cost-effective or deploying them at scale.

Two July 22 posts displayed during the discussion add to the immediate backdrop. Michael Kratsios, identified on screen as a White House technology policy official, wrote that Moonshot AI had distilled Anthropic’s Fable in developing K3 through what he called a sophisticated internal platform for large-scale distillation against U.S. models. In another post, he wrote that Moonshot had acquired GB300-equipped servers and accessed GB300s in Thailand, likely to train its models.

Michelle Giuda does not present restriction of Chinese models as sufficient to produce a durable U.S. advantage. Even if the United States limited their use domestically, she says, it would still face a strategic problem if the rest of the world ran on a Chinese AI stack. Her objective is for American open-weight AI to diffuse both inside the United States and across allied markets.

Her reference point is Huawei. Drawing on her time at the State Department, Giuda says the lesson was that the United States would not prevail simply by making a principled case against a Chinese technology provider. It needed alternatives that companies and countries could afford to choose.