Nvidia is pushing back on a report alleging that DeepSeek—the Chinese AI startup whose low-cost model release helped rattle markets earlier this year—is secretly training on smuggled “Blackwell” GPUs that are restricted from export to China.
The allegation stems from a report by The Information, which claimed DeepSeek has obtained Blackwell chips via smuggling networks and is using them to train an upcoming model. Nvidia, however, says it hasn’t seen evidence supporting the more dramatic parts of that narrative.
What the report alleged
Blackwell is Nvidia’s newest, most advanced AI platform, and U.S. export rules currently restrict shipping those top-tier chips to China. That’s why the Information report landed with a thud: it suggested DeepSeek was getting around the controls by acquiring Blackwell hardware through illicit channels.
Because the reporting relies on unnamed sourcing and neither DeepSeek nor Nvidia has publicly produced proof of Blackwell hardware in DeepSeek’s training stack, this remains an allegation—not a confirmed finding.
Nvidia’s response: “far-fetched,” but “we pursue any tip”
In a statement to CNBC, an Nvidia spokesperson said the company has not seen substantiation for claims involving so-called “phantom data centers” built to mislead partners and then dismantled and reconstructed elsewhere.
Nvidia added that while that kind of scenario “seems far-fetched,” it will investigate tips when they’re provided.
Why Blackwell is such a flashpoint
Export controls have turned Nvidia’s China business into a political lightning rod in Washington. The U.S. government has moved to limit China’s access to the most capable AI accelerators, and Nvidia has repeatedly had to adjust its lineup and sales approach in response.
Blackwell sits at the center of that tug-of-war: it’s exactly the kind of high-end compute that can materially accelerate AI training and inference at scale.
DeepSeek’s market shock is still fresh
DeepSeek burst into the global spotlight in late January 2025 with its R1 model, which it pitched as delivering strong results at dramatically lower costs than U.S. rivals. That triggered a broader debate about whether China’s AI ecosystem is moving faster than markets expected—and contributed to a sharp sell-off across U.S. tech stocks.
Later, in August 2025, DeepSeek suggested it was optimizing parts of its stack for “soon-to-be-released” next-generation domestic chips—fueling more speculation about how Chinese firms might reduce dependence on restricted U.S. hardware.
A new twist: Trump’s H200 comments
The situation got even more complicated this week after President Donald Trump said the U.S. would allow Nvidia to ship H200 GPUs to “approved customers” (including in China) under a framework where the U.S. government would receive a 25% fee/cut of those sales, with the Department of Commerce finalizing details.
Trump also claimed he had communicated the framework to China’s President Xi Jinping and characterized the response as positive—though that part is presented as Trump’s account, not an independently verified readout.
What to watch next
For now, there are two separate threads:
- The Blackwell smuggling claim, which Nvidia says is unsubstantiated so far
- The H200 policy shift, which appears to be moving forward, but still lacks full operational details
If regulators tighten enforcement—or if credible evidence emerges tying restricted GPUs to specific AI training runs—this story could escalate quickly. Until then, it’s a high-stakes mix of rumor, export control policy, and a very public race for AI compute.