DeepSeek’s V4 runs entirely on Chinese chips, challenging US export control strategy
New frontier model demonstrates that semiconductor restrictions may have accelerated architectural innovation rather than slowing China's AI development.
DeepSeek released its V4 series on 24 April 2026, featuring two models with 1-million-token context windows running exclusively on Huawei’s Ascend processors—the first frontier-class AI built entirely on Chinese domestic semiconductor infrastructure. The release tests a core assumption underlying US export controls: that restricting access to advanced chips would meaningfully slow Chinese AI development.
The V4 series includes two mixture-of-experts models, per Hugging Face. V4-Pro contains 1.6 trillion total parameters with 49 billion activated per token. V4-Flash uses 284 billion total parameters with 13 billion activated. Both support context windows of 1 million tokens—roughly 750,000 words—a technical threshold that positions DeepSeek alongside OpenAI and Anthropic in handling extended documents and conversations.
27%
10%
62.0%
The Huawei Integration
V4’s exclusive deployment on Huawei silicon marks a departure from prior releases. Earlier DeepSeek models relied on Nvidia A100 and H100 chips, despite export restrictions that limited Chinese access to cutting-edge hardware. The shift reflects both constraint and opportunity. US semiconductor controls, imposed in October 2022, aimed to starve Chinese labs of the compute necessary for frontier research. Instead, the restrictions appear to have accelerated development of a parallel hardware ecosystem.
“V4 is the first frontier-class demonstration that this parallel ecosystem is functional at scale,” an analyst told Ogun Security. The architectural choices reflect adaptation to hardware limitations. DeepSeek engineers developed compressed sparse attention and heavily compressed attention mechanisms that reduce computational overhead, according to a technical report from Hugging Face. At 1 million tokens, V4-Pro requires 27% of the single-token inference compute and 10% of the key-value cache compared to DeepSeek’s V3.2 model.
“The day that DeepSeek comes out on Huawei first, that is a horrible outcome for [the U.S.]”
— Jensen Huang, CEO of Nvidia
That day arrived this week. Nvidia CEO Jensen Huang had warned of this scenario in recent months, quoted by Fortune. The concern centres on decoupling: if Chinese AI development no longer depends on American semiconductors, export controls lose their strategic leverage.
Pricing and Market Positioning
DeepSeek’s pricing undercuts US competitors by an order of magnitude. V4-Pro costs $3.48 per million output tokens. V4-Flash costs $0.28 per million tokens. OpenAI charges $30 per million output tokens for comparable context windows. Anthropic charges $25. The gap reflects both lower hardware costs—Huawei chips trade at a discount to Nvidia equivalents—and aggressive market positioning.
| Model | Provider | Cost |
|---|---|---|
| DeepSeek V4-Flash | DeepSeek | $0.28 |
| DeepSeek V4-Pro | DeepSeek | $3.48 |
| GPT-4 (extended context) | OpenAI | $30.00 |
| Claude 3.5 Sonnet | Anthropic | $25.00 |
DeepSeek is in funding talks with Tencent and Alibaba that would value the lab at $20 billion, Fortune reported. The valuation reflects investor confidence in China’s ability to compete despite semiconductor constraints. CSIS projects that domestic chips will capture 50% of China’s AI market in 2026, up from negligible share two years prior.
Benchmark Performance
On the CorpusQA benchmark, which tests retrieval accuracy across long contexts, V4-Pro reached 62.0% at 1 million tokens. That beats Google’s Gemini 3.1 Pro, which scored 53.8%, per FellowAI. The results come from DeepSeek’s own technical report. Independent third-party validation remains limited as of 25 April 2026, given the model’s recent release.
The performance gap between Chinese and US frontier models has fluctuated. DeepSeek’s R1 release in January 2025 triggered what some called a ‘Sputnik moment’ for US AI, with Nvidia losing $600 billion in market value in the largest single-day decline in US stock market history. Chinese models on average still trail leading US releases by roughly seven months, though the gap narrowed briefly during R1’s launch before widening again.
US semiconductor export controls began in October 2022 with restrictions on advanced graphics processing units. The policy aimed to prevent China from accessing chips capable of training frontier AI models. The controls expanded in 2023 to cover chip manufacturing equipment and cloud computing services. China responded by stockpiling older-generation chips, investing in algorithmic efficiency, and accelerating domestic semiconductor development through Huawei’s Ascend line and Cambricon Technologies.
Implications for Export Control Policy
V4’s existence raises questions about the durability of chip-based containment. If frontier models can be built on hardware outside US supply chains, export controls may redirect Chinese development rather than prevent it. Wei Sun, principal analyst at Counterpoint Research, told CNN Business that V4 “allows AI systems to be built and deployed without relying solely on Nvidia, which is why V4 could ultimately have an even bigger impact than R1—accelerating adoption domestically and contributing to faster global AI development overall.”
The White House circulated a memo on distillation techniques—methods for transferring capabilities from large models to smaller ones—shortly before V4’s release. The timing suggests policymakers are reassessing whether current controls address the mechanisms by which Chinese labs achieve competitive performance. If algorithmic innovation compensates for hardware limitations, the calculus underlying export policy shifts.
MIT Technology Review notes that Chinese government coordination pushed Huawei integration. The strategic intent is clear: demonstrate self-reliance in the AI stack from silicon through serving infrastructure. Whether that translates to sustained competitiveness depends on how Huawei’s roadmap compares to Nvidia’s and whether China can replicate advances in GPU architecture, memory bandwidth, and interconnect technology.
What to Watch
Independent benchmark results will clarify whether V4’s self-reported performance holds across diverse tasks. Watch for adoption rates among Chinese enterprises, which would signal confidence in Huawei-based infrastructure. Track whether US labs respond by lowering prices or accelerating context window expansion. Monitor Congressional testimony and policy papers for shifts in how Washington frames export control effectiveness. The $200 billion AI infrastructure market is at stake, and V4 suggests the competition is less one-sided than recent narratives assumed.