DeepSeek’s V4 Will Run on Huawei Chips — and That’s Either a Geopolitical Triumph or the Most Convenient Cover Story in Tech
DeepSeek confirms its V4 model will run entirely on Chinese silicon, cutting out Nvidia and AMD. But a senior Trump official claims it was secretly trained on banned Blackwell chips.
TL;DR
- DeepSeek’s V4 — a multimodal, trillion-parameter model expected within weeks — will run entirely on Huawei chips, per The Information, marking a deliberate break from Nvidia and AMD.
- A senior Trump administration official is separately claiming V4 was actually trained on banned Nvidia Blackwell chips inside mainland China, with DeepSeek planning to publicly credit Huawei as cover.
- Whether the Blackwell accusation holds or not, China’s AI ecosystem has evolved to a point where its biggest lab can credibly claim hardware independence — and that alone changes the calculus for US sanctions policy.
The Information reported Friday that DeepSeek’s forthcoming V4 model will run entirely on chips designed by Huawei Technologies — making it the first major Chinese frontier model confirmed to operate without American silicon at inference time. The Hangzhou-based lab spent months working directly with Huawei and domestic chipmaker Cambricon to rewrite pieces of V4’s underlying code and optimize it for their latest hardware. Nvidia and AMD were cut out entirely, denied the pre-release access they would normally receive under long-standing industry convention.
The timing is pointed. Washington has spent four years building an escalating system of chip export controls designed to keep China’s AI labs dependent on American hardware. DeepSeek just told its most important suppliers they are no longer needed for the launch.
What DeepSeek V4 actually is
V4 is shaping up to be the most technically ambitious model DeepSeek has built. Leaked benchmarks describe a roughly one-trillion-parameter Mixture-of-Experts architecture — with around 37 billion parameters active per token — combined with a one-million-token context window and native multimodal generation across text, images, and video. If those specs hold, it would put V4 in the same tier as the best closed-source models from OpenAI and Anthropic.
The model incorporates what DeepSeek calls Engram, a conditional memory architecture meant to solve the long-context retrieval problem that has plagued large language models — the gap between technically “having” a million-token window and actually using relevant information buried deep inside it. Projected benchmarks suggest 90% on HumanEval and above 80% on SWE-bench Verified, roughly matching Claude Opus at a fraction of the cost.
A “V4 Lite” variant appeared on DeepSeek’s website on March 9, suggesting an incremental rollout. The full release is expected within weeks. Chinese tech giants Alibaba, ByteDance, and Tencent have already placed bulk orders for Huawei’s upcoming chips totaling hundreds of thousands of units, citing five people with direct knowledge of the purchases, according to The Information. The ecosystem is pre-positioned.
The Blackwell question nobody can answer cleanly
There’s a wrinkle. A senior Trump administration official told reporters this week that V4 was not trained on Huawei chips at all — that it was trained on Nvidia’s Blackwell architecture inside mainland China, a likely violation of US export controls. The official further claimed DeepSeek plans to publicly credit Huawei to mask that dependency and avoid scrutiny.
DeepSeek has not responded publicly to the allegation. The claim is unverified, and the administration has a clear political incentive to characterize Chinese AI progress as dependent on stolen or smuggled American technology rather than as a genuine domestic achievement. That motive doesn’t make the accusation false. It does make it difficult to assess from the outside.
What’s certain: Huawei’s Ascend 910B, 910C, and 910D chips are themselves controversial. The US Commerce Department’s Bureau of Industry and Security assessed in May 2025 that Huawei produced those chips in violation of American export controls, and warned companies globally that using them could expose them to sanctions risk. DeepSeek running V4 on Huawei hardware doesn’t automatically mean it’s running cleanly — it just means the silicon has a Chinese logo on it.
How the DeepSeek V4 Huawei chips story got here: the sanctions timeline
The chip export control story begins seriously in 2022, when the Biden administration cut China off from Nvidia’s most advanced training chips and the equipment needed to manufacture chips at leading-edge nodes. The controls were tightened repeatedly through 2023 and 2024, with successive rounds adding dozens of Chinese entities to export restriction lists and eventually blocking even the downgraded H20 chip that Nvidia had specifically designed to comply with earlier rules.
The theory was simple: AI capability scales with compute. Deny China the compute, and you deny it the models. Washington later drafted rules that would require Nvidia to apply for a license before selling to anyone, anywhere on earth — a global licensing regime for AI compute. The scope of that ambition tells you how seriously the US took the threat.
Then DeepSeek R1 arrived in January 2025 and made the theory look naive.
R1 reportedly cost around $6 million to train — 27 times cheaper than OpenAI’s o1, which it matched or beat on several benchmarks. It was built on hardware that was already restricted, using algorithmic innovations that extracted more capability from less compute than anyone thought possible. On January 27, 2025, Nvidia lost roughly $600 billion in market capitalization in a single session. The underlying assumption of the entire export control regime — that raw compute translates reliably into AI capability — had been publicly stress-tested and found fragile.
“DeepSeek shook confidence in two key assumptions underlying the AI boom,” the Peterson Institute for International Economics wrote afterward. “First, that US export controls limiting China’s access to computing power would hand US firms a large, durable lead. Second, that surging demand for AI would keep driving durable demand for chips and data centers.” Both assumptions survived, but neither survived intact.
China’s domestic chip ecosystem, a year later
The R1 shock accelerated something that was already underway. By early 2026, at least nine Chinese chip companies had exceeded 10,000 units in order shipments. Four domestic startups, dubbed the “four dragons” in Chinese financial press, had either gone public or filed to do so, attracting capital from investors who once looked exclusively at Nvidia’s supply chain.
Huawei’s Ascend line remains significantly behind Nvidia’s best products in raw performance and memory bandwidth. SMIC — the only Chinese foundry capable of building advanced chips at scale — remains stuck around 7nm due to export controls on lithography equipment. One analysis put China’s effective total AI computing power at one to four percent of the US’s, once quality differentials are factored in.
The counterargument is that DeepSeek keeps demonstrating you don’t need to close the hardware gap if you’re efficient enough. The lab’s willingness to rewrite inference code from scratch for Huawei and Cambricon rather than waiting for plug-and-play compatibility suggests a level of vertical integration that goes beyond what most Western labs attempt. That’s not a coincidence — it’s a deliberate bet that algorithmic efficiency can substitute for raw silicon.
US AI companies have their own exposure here. The quiet dependency US developers have on Chinese open-source research has been documented in uncomfortable detail. Chinese labs have dominated open-source leaderboards since R1, releasing competitive models faster and cheaper than their American counterparts. Top US universities have adopted Chinese models. Venture capital that once looked exclusively stateside has started following the research wherever it leads.
What experts expect from V4 — and what comes after
Experts who track the US-China AI gap generally place China around three to six months behind the US frontier — meaningfully behind, but no longer in a different tier. The assessment has been remarkably stable even as the export control regime has tightened, which is itself an argument about the limits of hardware restrictions as a strategic tool.
The US response has been fractured. The Biden-era AI Diffusion Rule — a tiered global licensing regime on all advanced chip exports — was rescinded by the Trump administration in late 2025. In its place came a more flexible review process. In December 2025, Trump approved the more powerful H200 for sale to approved Chinese customers at a 25% revenue share, a commercial concession that sent a confusing signal about how seriously Washington intended to enforce its own controls.
Google released Gemma 4 this week under an open Apache 2.0 license — explicitly framed as a competitive response to Chinese open-source leadership. It’s a data point about how the race is now understood inside American labs: not as a classified military competition but as an open leaderboard fight, visible to anyone who checks benchmarks.
V4’s launch will be the next data point. If it performs as leaked benchmarks suggest, running on hardware that the US has spent years trying to prevent China from building, the argument that export controls are buying meaningful time faces its most serious test yet. If the Blackwell accusation has substance — if V4 is Huawei on the outside and Nvidia on the inside — the story becomes different: not Chinese independence, but a reminder that controls are only as strong as the enforcement behind them.
Either way, DeepSeek has made a statement. It just hasn’t said which one.
Frequently Asked Questions
What is DeepSeek V4?
DeepSeek V4 is the next-generation large language model from Chinese AI lab DeepSeek, notable for being designed to run on Huawei’s Ascend AI chips rather than NVIDIA GPUs. This makes it one of the first frontier AI models trained entirely on Chinese domestic hardware, sidestepping US export restrictions on advanced semiconductors to China.
Why does DeepSeek V4 running on Huawei chips matter?
US export controls banned the sale of NVIDIA’s most advanced AI chips to China. By training on Huawei Ascend processors, DeepSeek demonstrates that Chinese AI labs can continue developing frontier models without American hardware. This has geopolitical implications—it suggests export restrictions may slow but not stop China’s AI progress, and could accelerate the development of a separate Chinese AI chip ecosystem.
How does DeepSeek V4 compare to models trained on NVIDIA hardware?
Full benchmark comparisons aren’t yet available, but DeepSeek’s previous models (V3, R1) were competitive with leading Western models despite hardware constraints. Huawei’s Ascend 910C chips are less mature than NVIDIA’s H100/B200 in software ecosystem and interconnect performance, but DeepSeek’s engineering workarounds have closed much of the gap. The real test will be V4’s performance on standard benchmarks when released.
