DeepSeek didn't need to build its own chip to make a point. By giving Huawei and other Chinese chipmakers first-class support, it made Nvidia's moat look less automatic.
For much of last year, the cleaner story around DeepSeek was that the Hangzhou-based AI lab would eventually design its own silicon. That may still happen. But the sharper move is the one you can already see: DeepSeek has been making its models run better on Chinese chips that exist now, not waiting for a homemade accelerator that might take years to reach volume.
Reuters reported in February that DeepSeek had denied Nvidia and AMD early access to its coming V4 model while giving Huawei and other domestic suppliers a head start. That is not a small courtesy. Chipmakers normally want early access so they can tune the drivers, get the kernels sorted and adjust the software stack before a major model lands. DeepSeek gave that advantage to Huawei and Cambricon, and to other Chinese firms besides, instead of the American companies that still dominate AI training and inference.
That is the opposite of building your own chip. It is picking partners and building the software bridge that makes their chips usable at scale.
When DeepSeek released the V4 preview on April 24, the signal became harder to miss. Tom's Hardware reported that V4 arrived as a 1.6 trillion parameter model with a 1 million token context window and was optimized for Huawei's Ascend processors rather than Nvidia hardware. The same report said V4-Pro cost $3.48 per million output tokens, while the smaller V4-Flash cost $0.28. Those prices matter because cheap inference only works if the hardware supply underneath it can keep up.
Jensen Huang didn't hide his concern. Nvidia's chief executive warned that DeepSeek optimizing for Huawei's chips instead of American hardware would be "a horrible outcome" for the United States. That is unusually blunt from a CEO who usually has every reason to talk about demand rather than dependence.
The market heard the same message.
Semiconductor Manufacturing International Corp jumped 11% and Hua Hong Semiconductor rose more than 18% in Hong Kong trading after the V4 preview landed, according to Investing.com. Investors were not buying a rumor about DeepSeek's private chip lab. They were buying the idea that one of China's most watched AI companies had just pushed real demand toward domestic silicon.
DeepSeek's decision looks even less surprising when you look at the cost of the alternative. Designing a competitive AI accelerator from scratch is not like hiring a few engineers and waiting for a tape-out. You need architecture work, software tools, packaging, memory supply, foundry access and customers willing to live with the first version. Washington's export controls have already made the foundry and equipment side harder for Chinese companies. Partnering with Huawei, which has spent years fighting its own chip war with the United States, gets DeepSeek working hardware now.
There is also fresh evidence that Huawei wants more than compatibility badges. Tom's Hardware reported in June, citing the South China Morning Post and the Shenzhen municipal government, that a Huawei-linked research group post-trained DeepSeek's V4-Pro model using more than 1,000 Ascend 910C chips. DeepSeek did not comment, and Huawei did not publish the kind of benchmarks that would settle the argument. Keep that caveat. Still, full-parameter post-training on domestic accelerators is a real step, even if it is not the same as proving Huawei can handle frontier pre-training at Nvidia scale.
The funding picture points in the same direction. The Wall Street Journal reported last week that DeepSeek had raised more than $7.4 billion at a valuation above $50 billion and planned to roughly double its workforce. Reported backers included Tencent and Contemporary Amperex Technology, plus China's national AI investment fund. You don't raise that kind of money just to admire your own efficiency story from 2025. You spend it on people, compute, deployment and the hard plumbing that turns a model into an operating business.
None of this means DeepSeek has abandoned in-house silicon for good. Chinese AI companies that once bought whatever Nvidia was allowed to sell them now treat chip access as a strategic asset, not a line item on a purchase order. But for now, the company's choice is plain enough. Let Huawei, Cambricon and Hygon fight the chip fight. DeepSeek can spend its war chest making sure its models run where China's compute supply actually is.
Also read: SK Hynix's Wild Stock Swings Threaten to Reshape Its $28 Billion Nasdaq Debut • Chinese Firms Plan to Send Nearly Half Their AI Chip Budgets to Local Rivals • Illinois Becomes the First State to Mandate Annual AI Safety Audits
