China's leading AI companies are running into a hard ceiling: they cannot get enough advanced computer chips to train frontier models, even as the U.S. debates how to respond to their rapid progress. That constraint, revealed in leaked investor comments from DeepSeek CEO Liang Wenfeng, offers a concrete reason why export controls on semiconductors could be far more effective than other policy options being considered in Washington.

What Did DeepSeek's CEO Actually Say About China's Chip Problem?

Liang Wenfeng told investors this week that his company needs 200,000 Huawei 950 chips to train a frontier-level AI model, but Huawei can only supply 16,000 of them. Even more telling, Huawei's total expected production capacity for the entire year is just 750,000 chips, which must be divided among all Chinese AI companies competing for access. "Huawei's problem is still insufficient capacity," Liang said, adding that he expected this shortage to persist for at least three years. "This problem is currently basically unsolvable."

Liang Wenfeng

These comments matter because they come from someone who is relatively optimistic about China's long-term ability to build its own chips. If even a bullish insider sees a multi-year gap, the constraint is real. The timing is critical: if transformative AI breakthroughs happen within the next few years, as companies like OpenAI and Anthropic believe they will, a short-term compute advantage could become permanent.

Why Are Export Controls Better Than Sanctions or Bans?

The Trump administration is considering multiple responses to the recent release of Kimi K3, a powerful Chinese AI model developed by Moonshot AI. Treasury Secretary Scott Bessent has threatened sanctions against Chinese companies, and officials have floated the idea of banning foreign-made open-source models. But these approaches have significant drawbacks.

Sanctioning companies like Moonshot might reduce their access to American business partners, but it is unclear how much of their advanced capabilities come from distillation, the practice of training on outputs from American models like Claude and ChatGPT, versus their own engineering. A ban on open-source models would face fierce opposition from startups and smaller companies that rely on them as affordable alternatives to expensive proprietary systems.

Export controls on advanced semiconductors, by contrast, directly target the physical resource that Chinese companies cannot easily replace. As long as China depends on American chips to train advanced models, the U.S. has direct leverage. Two bills under consideration for this year's National Defense Authorization Act could strengthen this approach:

• Remote Access Security Act: Would prevent Chinese companies from accessing U.S. cloud computing resources, closing a major loophole where companies like Moonshot allegedly rented compute power instead of buying chips outright.
• Chip Security Act: Would mandate location verification on advanced chips to prevent smuggling, addressing the problem that Kimi K3 was allegedly trained on Nvidia Blackwell chips that were either smuggled into China or accessed remotely from Thailand.

What's the Catch With Export Controls?

The standard argument against semiconductor export controls is that they accelerate China's efforts to build its own chip manufacturing capacity. That is certainly true, but timeframes matter enormously. Even Liang, who is bullish on China's abilities, expects it to take years to catch up. During that window, the U.S. could maintain a decisive advantage in AI development.

OpenAI plans to use roughly eight times as much computing power as DeepSeek currently has access to in order to run its automated AI research "intern" later this year. If that leads to a recursive self-improvement loop, where AI systems improve themselves faster and faster, a short-term compute advantage now could translate into permanent technological dominance.

How to Evaluate Export Control Options

• Effectiveness Measure: Look at whether a policy directly constrains the physical resource that Chinese companies cannot substitute. Chip export controls do this; sanctions on individual companies do not.
• Timeline Consideration: Assess how long the constraint will remain binding. Liang's comments suggest semiconductor constraints will persist for at least three years, which aligns with the window when transformative AI breakthroughs are expected.
• Loophole Risk: Evaluate whether companies can circumvent the policy through remote access, smuggling, or alternative suppliers. The Remote Access Security Act and Chip Security Act address these specific gaps in current controls.

What Are the Competing Pressures on the Trump Administration?

The administration faces pressure from multiple directions. White House officials like Michael Kratsios, director of the Office of Science and Technology Policy, have emphasized national security concerns and accused Chinese companies of stealing American intellectual property through distillation. But Silicon Valley is pushing back hard.

A group of 179 startups wrote to the Trump administration warning that banning foreign-made open-source models would stifle competition, entrench incumbents like OpenAI and Anthropic, and function as "a tax on intelligence." Industry heavyweights including Microsoft, Mistral, Perplexity, and Meta have also called for restraint, proposing a legal and commercial framework rather than broad restrictions.

"The Kimi Panic needs to stop. As long as we don't sabotage ourselves with unnecessary rules, the U.S. will continue to win," said David Sacks, the White House's former AI policy chief.

David Sacks, Former White House AI Policy Chief

Jensen Huang, CEO of Nvidia, which manufactures the chips at the center of this debate, has also weighed in. "American companies should absolutely be allowed to use Chinese AI models," he told Axios. "These Chinese models are excellent. Open source models that are excellent should be used".

What Happens Next With Kimi K3?

Moonshot AI is scheduled to release the full model weights for Kimi K3, a 2.8-trillion-parameter model, on July 27, 2026. This open-weight release creates a critical decision point for organizations evaluating the model: whether to use Moonshot's hosted API, which routes all prompts through servers in Beijing subject to China's National Intelligence Law, or download the weights and run the model on their own infrastructure.

The window for downloading may not stay open. China's Ministry of Commerce has begun consulting domestic AI companies on export controls that could restrict future weight downloads, suggesting the political pressure is mounting on both sides.

The core question facing policymakers is whether to address the challenge through targeted semiconductor controls that exploit China's current constraints, or through broader restrictions that risk stifling competition and innovation in the U.S. tech sector. DeepSeek's CEO has inadvertently made the case for the former approach.