Nvidia's Huang: US Firms Should Use Chinese AI Models

Jensen Huang endorses Chinese open-source AI as "excellent," pushing back against Treasury sanctions threats and reframing global AI competition as a demand catalyst for Nvidia.

Lead

What Happened

Huang's remarks came in direct response to a deepening standoff between Washington policymakers and global technology executives over the status of Chinese-built large language models.

"These Chinese models are excellent," Huang said. "Open-source models that are excellent should be used." When pressed, he said companies should "absolutely" be permitted to deploy them.

The Nvidia chief also dismissed arguments that downloading Chinese AI creates a security backdoor. Enterprises can fine-tune and run these models inside air-gapped, on-premises environments, he argued, and the open nature of the weights means outside researchers can audit models for vulnerabilities — making openness a feature, not a liability.

Huang's broader case is commercial as well as philosophical: cheaper, more accessible models pull more users into AI workflows, ultimately lifting demand for the data-center infrastructure and accelerators that Nvidia supplies. More models running on more machines, in his framing, is net positive for the semiconductor stack regardless of which country built the software.

The Kimi K3 Catalyst

The timing is tied to a fresh shock from Beijing. On July 16, Moonshot AI, a Chinese startup, released Kimi K3 — a 2.8-trillion-parameter, open-weight multimodal reasoning model, the largest open-weight release in the industry to date. The model carries a one-million-token context window, an always-on reasoning mode, and pricing of $3 per million input tokens — a fraction of comparable US frontier offerings.

On independent evaluations, Kimi K3 ranks fourth among all frontier models globally, trailing only Anthropic's Claude Fable 5 and OpenAI's GPT-5.6 Sol, while edging past Anthropic's own Claude Opus 4.8. Full open weights are scheduled for release by July 27.

The announcement triggered immediate comparisons to the DeepSeek shock of early 2025, when a Chinese model demonstrated near-frontier performance at dramatically lower computational cost. Taiwan's equity benchmark fell more than 6%, Japan's market dropped 4%, and the Nasdaq slid 1.5% in a single session after Kimi K3's debut. Nvidia has now retreated roughly 14% from its 2026 high.

Bessent's Sanctions Warning

Treasury Secretary Bessent had set the confrontational backdrop just hours before Huang spoke. Appearing on Fox Business on July 21–22, Bessent said the administration is actively examining Chinese open-source AI models for signs of "distillation" — a training technique in which a less-capable model learns from the outputs of a more advanced one, a process the administration characterizes as unauthorized appropriation of American intellectual property.

"We have found watermarks of our US large language models on many of the Chinese models," Bessent said, adding that the administration would pursue sanctions if theft is confirmed. He framed the review as imminent, saying findings would emerge "in the coming days or weeks."

Bessent was careful to note that the administration "supports open-source models" in principle, but draws the line at what he termed theft of American-developed AI. The threat represents a significant expansion of the US technology-competition toolkit beyond hardware export controls into the software and model-weight domain.

Huang's Counter-Argument

Huang challenged each pillar of the administration's case. On the security question, he argued that companies deploying downloaded models in controlled internal environments face no meaningful exposure to Chinese state access. On the IP-theft question, he offered no direct rebuttal but reframed global AI competition as inherently beneficial: competition accelerates improvement, and Nvidia's business expands with every additional AI workload, wherever it originates.

The Nvidia CEO's stance carries weight beyond the rhetorical. Nvidia remains the dominant supplier of AI accelerators worldwide, and its H200 chip export pathway to China remains contested — US commerce officials stated as recently as mid-2026 that the chipmaker had not completed a single H200 sale into China despite regulatory clearance. Huang has pursued resumed Nvidia China AI sales since announcing in March 2026 that Chinese customers had placed orders and Nvidia was restarting manufacturing.

Endorsing the quality of Chinese LLM development simultaneously reframes the narrative: if Chinese models are already world-class, export restrictions on Nvidia hardware harm Nvidia's revenues without degrading Beijing's AI capabilities.

Market Reaction

Analysts have drawn a distinction between model-software economics and hardware economics: cheaper Chinese AI models may accelerate AI adoption broadly, expanding the total addressable market for chips, even as they compress margins for US AI software vendors.

Geopolitical Dimension

The Huang-Bessent exchange crystallizes a core tension in US global AI competition strategy: whether restricting access to Chinese AI software meaningfully protects American technological leadership, or whether it handicaps US businesses against foreign competitors while Beijing continues model development unimpeded.

The administration's export-control regime has already cost Nvidia an estimated $50 billion in China-market revenue. Extending restrictions from hardware to open-weight model weights would mark a qualitative escalation — policing software that, once publicly released, can be downloaded, copied, and hosted anywhere on the planet.

International partners and technology executives have argued consistently that open-source AI restrictions are difficult to enforce, easy to circumvent, and risk fragmenting the global standards ecosystem in ways that disadvantage the United States over a multi-decade horizon.

Outlook

The debate over Chinese open-source AI has moved from theoretical to operational with Kimi K3's release and now sits at the intersection of corporate strategy, national security, and trade policy. Bessent's sanctions threat signals that the administration is prepared to extend its technology-competition toolkit into model software — a domain where enforcement mechanisms remain undefined.

Huang's public endorsement of Chinese model quality is a calculated position from the executive whose company supplies the infrastructure underpinning both American and, potentially, Chinese AI ecosystems. The next material data points are the completion of Bessent's IP review, the July 27 full open-weight release of Kimi K3, and any further policy action from the Commerce Department on AI model exports. Nvidia's capacity to resume meaningful China hardware revenues remains the financial variable with the largest near-term bearing on the company's earnings trajectory.