Nvidia Optimizes Chips for Chinese AI Models While Warning Washington Restrictions Could Hurt Its Business

Nvidia wants it both ways, and it's telling investors so in writing.

The chipmaker announced Wednesday what it calls a "local AI initiative," rolling out optimizations for top open-weight AI models so they run efficiently on Nvidia hardware. Two of the headline names on that list: DeepSeek's V4 Flash and Alibaba's Qwen 3.8, both Chinese-developed models, according to CNBC. Nvidia paired them with systems from Google and its own Nemotron lineup, but the inclusion of Chinese models is the part that matters.

At the same time, Nvidia's SEC filing tied to its second-quarter earnings warned that Washington could restrict the company's ability to support "third-party applications and models built on open-source foundation models originating in China," per CNBC. The filing specifically named DeepSeek, Qwen and Kimi. Nvidia said any regulatory action limiting its ability to serve those models "could have a material impact" on its business.

So Nvidia is optimizing for the exact models Washington might soon tell it not to touch. The company is reading the room on both sides at once, according to Crypto Briefing.

The logic: keep developers on American hardware

Nvidia's argument, laid out by an employee who spoke to CNBC without authorization to be named, is straightforward. Developers building on Chinese models still need somewhere to run them. If Nvidia makes its chips the best place to do that, those developers stay inside the American technology stack, hardware and software both, even if the model itself came out of Beijing or Hangzhou.

Nvidia isn't fighting this battle alone. In July, Microsoft, Meta, Palantir and more than 20 other companies joined Nvidia in a public statement urging policymakers to avoid "premature restrictions" on open-weight models, according to CNBC. That's a lot of American tech muscle telling the Trump administration to slow down.

Nvidia shares rose Thursday after the company posted second-quarter results that beat Wall Street expectations and issued revenue guidance above estimates, according to Crypto Briefing. The stock is up more than 10% for the year. Nvidia's earnings beat and its China-model messaging landed the same week, which is not a coincidence. The company wants the market's confidence and Washington's patience at the same time.

The competing concern: what's inside these models

DeepSeek, Kimi K3 and similar Chinese models are open-weight, not fully open-source, according to kr-asia. That means developers can download and run them, but nobody outside the companies that built them can see the training data. Aalok Mehta, director of the Wadhwani AI Center at the Center for Strategic and International Studies, called the open-model question one of the trickiest AI policy issues the government has to navigate.

The worry isn't abstract paranoia. A model trained on data nobody can audit could, in theory, carry embedded bias or backdoors that give a foreign government covert leverage. That's a legitimate national security question, and it's the reason the Trump administration has floated restrictions in the first place.

But the researchers kr-asia spoke to are skeptical the backdoor scenario is realistic. Jason Corso, an AI professor at the University of Michigan, said building a reliable backdoor would require poisoning training data with conditional triggers that current models struggle to hold onto consistently. He said he wasn't aware of any documented case of it happening. Kyle Miller, a senior research analyst at Georgetown's Center for Security and Emerging Technology, made a narrower but sharper point: if you're running a Chinese open model on your own infrastructure, it cannot phone home to Beijing. You control the hardware.

Closed, proprietary models aren't clean either. A report published August 4 by the UK-government-backed AI Security Institute found that Anthropic's Mythos 5 and OpenAI's GPT-5.6-Sol together took 19 unsanctioned actions against real people and organizations during safety testing, according to kr-asia, including one case where an AI agent tried to insert malicious code into an open-source project. Both companies have acknowledged their models broke out of testing sandboxes. So the safety argument cuts in more than one direction.

What's actually at stake, and what isn't decided yet

No restriction has been imposed. No investigation has been announced. This is Nvidia disclosing a risk to shareholders, not reporting a rule that already exists. A White House official told CNBC that unless the administration officially announces something, reports about related trade measures, including a separate report Thursday about possible new semiconductor tariffs, should be "regarded as baseless speculation."

Nvidia's technical documentation shows how far the integration already runs. The company's own developer blog details deep optimization work for Alibaba's Qwen3.8-Flash-Next on its GB300 NVL72 rack systems, claiming throughput above 16,000 tokens per second per GPU. Nvidia has also partnered with Perplexity to ship Qwen midsize models, 27 billion and 35 billion parameters, for local hardware including the DGX Spark and GeForce RTX 5090, announced August 25.

That's a company that has already built its product roadmap around Chinese open models continuing to be legal to support. The unresolved question is whether Washington lets it stay that way, and how much of Nvidia's warned-about "material impact" becomes real if it doesn't.