China's free AI models split Trump's AI camp

Capable, free Chinese AI models — above all Moonshot's Kimi — have opened a public rift inside the tech circle advising the Trump administration. As James O'Donnell reported in MIT Technology Review on 20 July 2026, the president's advisers disagree over whether Washington should back open models or restrain them on national-security grounds.

Key takeaways

Moonshot AI's Kimi is free and open, matching the paid models of OpenAI and Anthropic on capability
David Sacks, Trump's AI adviser until March 2026, accuses leading firms of wanting the government to eliminate their open-source competition
The White House introduced a pre-release review of AI models — Dean Ball called it a de facto licensing regime
In April 2026 the administration announced moves against distillation, training models on the outputs of other models
Washington loosened chip export controls, boosting Nvidia sales to China

A dispute that went public

The conflict is no longer behind closed doors. As MIT Technology Review reports, public statements from former and current officials reveal a deep split over how to respond to Chinese models, including Kimi from Moonshot AI. David Sacks, who ran AI and crypto policy in the administration until March 2026, accused the largest firms, including OpenAI, of wanting the government to wipe out their open-source competition. He described the models of Anthropic as “lobotomized" and “woke". Emil Michael, a Pentagon official, sharply dismissed one of OpenAI's strategy leads. The tone of the debate — unusual for a technology-policy fight — signals how high the stakes are.

Two visions: openness versus control

The split comes down to two positions. The first, close to Sacks, holds that unrestricted competition — including open models — serves consumers and strengthens the US position. The second stresses national security and justifies state control over what reaches the market and when. The concrete expression of the second approach is a new White House process to review models before release. Dean Ball, a former administration adviser now tied to OpenAI, called it a “de facto licensing regime" — a sign that criticism comes even from a side that could, in theory, benefit from regulation. Earlier, one Anthropic model was halted as a national-security threat.

Chips, distillation and the question over Kimi

The dispute over models overlaps with hardware policy. The administration loosened chip export controls, which boosted sales by Nvidia to China — a decision hard to square with treating Chinese models as a threat. In April 2026 Washington announced moves against distillation, the practice of training new models on the outputs of existing ones, which some US firms attribute to Chinese rivals. An open question remains which chips Kimi was trained on, given export limits and the smaller pool of compute available in China. Ben Thompson's commentary, cited by Simon Willison, adds context: Alibaba, which had kept its Qwen 3.7 Max closed, released Qwen 3.8 Max as open weights — in step with Xi Jinping's call for “open source, openness, collaboration and sharing".

Why it matters

The rift in Trump's camp shows US AI policy lacks a coherent doctrine, and that the dividing line does not run neatly between regulation and free markets. Free Chinese models undercut the assumption that US advantage rests on closed, best-in-class systems. If a free model matches paid offerings, the “better American product" argument weakens, and the temptation grows to reach for state tools: export controls, model reviews, curbs on distillation. The problem is that the same tools hit American open source, which part of the administration sees as an asset rather than a risk. The result is internally contradictory policy: loosening chip exports while warning about Chinese models. For companies this means regulatory uncertainty — it is unclear whether the next strong model will clear review or be labeled a threat. For the global market it signals that model availability is increasingly decided by geopolitics, not by benchmarks alone.

What's next

The criteria and scope of the new White House model review remain undisclosed — they will determine whether future US models ship without delay
The April 2026 moves against distillation need legal specifics — they will decide how models can lawfully be trained on other models' outputs
Further Chinese releases will test whether Washington keeps its looser chip-export stance or tightens it