DeepSeek V4 Runs on Huawei Chips: China AI Autonomy Signal, SMIC +10%

DeepSeek launched a preview of a new V4 model on April 26, 2026 specifically adapted for Huawei's AI chips, with Huawei confirming its hardware was used in portions of the V4 training process. The model topped Hugging Face's trending rankings faster than any model previously recorded. The Pro version outperforms other open-source models on world-knowledge benchmarks, trailing only Google's closed-source Gemini Pro 3.1. Chinese chipmaker stocks rallied on the news: SMIC rose 10%, Huahong Semiconductor rose 15%. Nvidia shares moved upward separately on Intel's strong revenue forecast.

The significance is not the benchmark score. It is the architecture decision: DeepSeek has demonstrated that a frontier-class open-source AI model can be trained on Huawei chips rather than Nvidia hardware. The dependency that US export controls were designed to enforce — Chinese AI labs need Nvidia GPUs to train competitive models — is now publicly broken.

What Changed Between DeepSeek V4 Pro and This Release

DeepSeek V4 Pro, released April 24, was trained predominantly on Nvidia hardware (either stockpiled pre-controls H100s or newer chips acquired through intermediaries). The April 26 Huawei-adapted model is different in kind, not just degree.

The Huawei adaptation means two things simultaneously. First, the model was optimised to run inference efficiently on Huawei's Ascend AI chip series (specifically the Ascend 910B and 910C). Second — and more significantly — Huawei confirmed that some of the V4 training runs used Huawei hardware. Training a frontier-class model on Huawei chips requires that those chips are competitive with Nvidia on the training workload profile. DeepSeek's willingness to publish this validation is a direct demonstration of Huawei chip training capability.

The Hugging Face ranking matters for a different reason. DeepSeek V4 becoming the fastest model to hit the number one trending slot means the developer community immediately adopted it at scale for testing, experimentation, and production evaluation. That adoption velocity will generate real-world performance data on Huawei hardware within days, providing independent validation (or refutation) of the training capability claim.

The Export Controls Logic: What Just Happened

US semiconductor export controls since 2022 have operated on a core assumption: without access to Nvidia's most capable training GPUs (A100, H100, H200, B200), Chinese AI labs cannot train frontier-class models competitively with US labs. Restrict the chips, constrain the models.

DeepSeek has now demonstrated this assumption is wrong — three times in eighteen months:

DeepSeek R1 (late 2024): Competitive with GPT-4-class models trained at a fraction of the stated compute cost, using reported stockpiles of older Nvidia H800 chips.

DeepSeek V4 Pro (April 24, 2026): 1.6T parameters, beating Claude on coding benchmarks, open-source.

DeepSeek V4 Huawei (April 26, 2026): Frontier-class model trained on Huawei hardware with no Nvidia dependency.

Each iteration has closed a different gap in the "export controls will slow China" thesis. V4 Huawei closes the hardware supply gap: even without any Nvidia chips at all, DeepSeek can train frontier models. The policy foundation for chip export controls as an AI containment strategy now requires a new argument.

Huawei Ascend: What the Training Validation Means

Huawei's Ascend 910B has been China's primary domestically produced AI accelerator. Its performance claims — roughly comparable to an Nvidia A100 on training workloads — have been disputed by Western chip analysts who note that Huawei's stated benchmarks are not independently verified and that the chip is manufactured on SMIC's 7nm DUVi process rather than TSMC's more advanced nodes.

DeepSeek using Ascend chips for V4 training provides the most credible independent performance validation Huawei has received. If DeepSeek — which has demonstrated aggressive compute efficiency and willingness to publish honest benchmark comparisons — chose to use Huawei chips for training runs, those chips are performing at a level that makes frontier model training viable.

The SMIC 7nm DUVi process connection matters here. We covered the AEI "Lithography Loophole" report showing SMIC producing near-frontier chips via multi-patterning on older DUVi machines. The Ascend 910B is one of the primary products coming off those SMIC lines. DeepSeek V4 Huawei is, in effect, a live demonstration that SMIC's DUVi-fabricated AI chips are good enough to train frontier AI models.

SMIC +10%, Huahong +15%: The Market Signal

Chinese semiconductor stocks rallied sharply on the DeepSeek V4 Huawei announcement. SMIC rising 10% and Huahong Semiconductor rising 15% in a single session reflects market repricing of the domestic Chinese AI chip demand opportunity.

The market logic: if DeepSeek — the most benchmarked and credible Chinese AI lab — is actively using SMIC-fabricated Huawei chips for training, then the addressable market for domestic Chinese AI chips just became concrete. Previously, Chinese cloud providers (Alibaba, Baidu, ByteDance) and AI labs were using domestically produced chips for inference while relying on Nvidia stockpiles for training. The training gap closing means domestic chip demand expands from inference-only to the full training-and-inference cycle.

Huahong Semiconductor is the second-largest Chinese chip foundry after SMIC, producing mature-node chips that supply the broader Chinese electronics industry. Its 15% rally on this news reflects derivative demand: more SMIC training chip production requires more supporting semiconductor components from suppliers like Huahong.

The Hugging Face Signal: Developer Adoption Velocity

DeepSeek V4 becoming the fastest model to top Hugging Face's trending rankings is a developer adoption metric that matters independently of benchmark scores. Hugging Face trending reflects active download and experimentation by the developer community — the people who build production AI applications.

The Huawei-adapted model topping Hugging Face immediately means developers are actively evaluating whether they can deploy V4 on Huawei hardware. For developers in China building AI applications, this is directly relevant: AWS China and Azure China both have constraints on Nvidia hardware availability; Huawei Cloud with Ascend chips is the primary alternative. A frontier model optimised for Huawei hardware changes the economics of Chinese cloud AI deployment.

For developers outside China: the open-source weights mean V4 (including the Huawei-optimised version) is freely available. Developers with access to Huawei hardware through their enterprises or through Chinese cloud providers can now deploy frontier-class models without Nvidia dependency.

What This Means for US Export Control Policy

The DeepSeek V4 Huawei announcement lands the same week as the AEI report on China's DUV lithography loophole and the House bill targeting capability-based chip controls. The policy momentum is clear: the current export control framework has not prevented China from reaching frontier AI model capability, and the hardware gap is now demonstrably closed.

The next legislative move — likely in the export control reauthorisation cycle — will need to answer a harder question: if China can train frontier AI models on Huawei chips fabricated on SMIC DUVi lines, what specific export controls would meaningfully constrain Chinese AI capability without simply restricting US chip company revenues? The answer is not obvious, and DeepSeek has just made the policy problem harder.

Key Takeaways

• DeepSeek launched Huawei-adapted V4 model April 26, 2026: trained partly on Huawei Ascend chips; fastest model ever to top Hugging Face trending; Pro version trails only Google Gemini Pro 3.1 on world-knowledge benchmarks
• Nvidia dependency broken: this is the third successive DeepSeek release demonstrating Chinese frontier AI without Nvidia reliance; V4 Huawei closes the training hardware gap specifically
• SMIC +10%, Huahong +15%: market repricing domestic Chinese AI chip demand from inference-only to full training cycle; Ascend 910B validated as frontier training-capable hardware
• Export control implications: core assumption of chip controls (Nvidia restriction = model capability constraint) is now publicly refuted; policy needs new argument
• Connects to DUV loophole: Ascend chips are SMIC DUVi-fabricated; DeepSeek V4 Huawei is a live demonstration that SMIC's near-frontier chips are good enough to train frontier AI models
• Developer implication: frontier open-source model available without Nvidia hardware dependency; Chinese cloud AI deployment economics change; Huawei Cloud Ascend deployments now viable for frontier inference and training

For the DUV lithography loophole that enables Huawei chip production, read China's DUV Lithography Loophole: SMIC and Huawei Printing Near-Frontier AI Chips. For the DeepSeek V4 Pro benchmark context, read DeepSeek V4 Pro: 1.6T Parameters, Beats Claude on Coding. For the competing model landscape, read OpenAI GPT-5.5 Released: Agentic Coding Upgrade.