DeepSeek, the Chinese AI lab that wiped nearly $600 billion off Nvidia's market cap in a single January 2025 trading session, has shut both Nvidia and AMD out of early access to its upcoming V4 model, according to a Reuters report citing two people familiar with the matter. Huawei and other domestic Chinese chipmakers got several weeks of advance optimization time instead.
What actually happened
The standard playbook in AI goes like this: before releasing a major model, you hand pre-release builds to the big chipmakers so they can tune drivers, run benchmarks, and make sure everything runs smoothly on their hardware. DeepSeek used to do exactly that with Nvidia's engineering teams.
Not this time. For V4, expected around the Lunar New Year holiday, Nvidia and AMD got nothing. Both companies declined to comment. DeepSeek and Huawei didn't respond to requests for comment either, which tells its own story.
Ben Bajarin, CEO of research firm Creative Strategies, tried to talk the impact down. The practical damage to Nvidia and AMD "is minimal," he told Reuters, since most enterprises treat DeepSeek as a benchmarking tool rather than production infrastructure. He also noted that AI coding tools are compressing hardware optimization timelines "from months to weeks." But then Bajarin said something more revealing: the move is likely part of a broader Beijing strategy to "keep US hardware and models disadvantaged" in China.
That's the more interesting read. If Bajarin's right, this isn't just one company making one product decision. It's policy.
The Blackwell problem
Here's where the story gets uncomfortable. Days before the V4 lockout was reported, a senior Trump administration official told Reuters separately that DeepSeek's latest model was trained on Nvidia's most advanced chip, the Blackwell, using a cluster in mainland China. That would be a direct violation of US export controls, which currently bar Blackwell shipments to China.
The US official went further: DeepSeek may try to scrub technical indicators showing it used American silicon, then publicly claim the model was trained on Huawei hardware. The Information had previously reported that DeepSeek smuggled chips into China for training. The government's confirmation that those chips are likely Blackwells, clustered at a data center in Inner Mongolia, is new.
So the company allegedly trains on smuggled Nvidia chips, then locks Nvidia out of optimization access and prepares to credit Huawei instead. If that sequence is accurate, it's a level of strategic misdirection that goes well beyond typical vendor politics.
Does it matter for Nvidia?
Bajarin's assessment that the near-term financial impact is minimal is probably correct, though consider who benefits from that framing. Nvidia and AMD still dominate the global AI hardware market, and most paying enterprise customers worldwide run their models on Nvidia GPUs regardless of what DeepSeek does with its pre-release access.
AMD's China-specific MI308 inference chip pulled in $390 million in its most recent quarter, a figure that speaks to real demand for US hardware in the Chinese market. The question is whether that demand holds if Beijing keeps pressuring domestic labs to optimize for, and publicly endorse, Chinese silicon.
DeepSeek's models have crossed 75 million downloads on Hugging Face since the company emerged in early 2025, and Chinese open-source models now account for more downloads on the platform than those from any other country. If those models increasingly ship pre-optimized for Huawei's Ascend processors and not for Nvidia, the optimization gap could start mattering in markets where DeepSeek has traction.
Saif Khan, who served as director of technology and national security at the White House National Security Council under Biden, put it bluntly in the earlier Reuters report: reliance on smuggled Blackwells "underscores their massive shortfall of domestically produced AI chips." That's true today. Whether it stays true depends on how fast Huawei can close the gap, and whether decisions like this one accelerate that process.
DeepSeek is among several Chinese AI firms expected to release new models this month. US authorities allowed Nvidia's H20 and AMD's MI308 to resume shipping to China in 2025, though licenses for more advanced processors remain restricted. Whether DeepSeek has approval to purchase even those chips remains unclear.
