Chinese AI startup DeepSeek has begun optimizing its newest large language model, V4, to run on Huawei’s Ascend processors, a move that directly challenges Nvidia’s grip on the AI training hardware market. The decision, confirmed in April 2026 reporting from both the Associated Press and Bloomberg, comes as tightening U.S. export controls continue to restrict Chinese companies’ access to Nvidia’s most advanced GPUs.
DeepSeek shook the AI industry in early 2025 when its R1 reasoning model matched top Western competitors at a fraction of the reported training cost. Its V3 model, released shortly before, demonstrated that a relatively small Chinese lab could compete with the likes of OpenAI and Google on benchmark performance. Now, with V4, the company appears to be making a bet that it can maintain that competitiveness while shifting away from the Nvidia hardware that has powered virtually every frontier AI model to date.
A preview is live, but the full launch is on hold
The AP reported that DeepSeek rolled out a preview of V4 with confirmed support for Huawei’s Ascend chips, framing the release as a concrete step toward reducing reliance on American semiconductor companies. Bloomberg, drawing on a CCTV account, reported a different angle: DeepSeek had postponed the full V4 launch specifically to deepen its integration with Huawei’s Ascend hardware.
The two accounts are not necessarily contradictory. DeepSeek may have released a limited demonstration while holding back the production-grade version for further optimization. But neither outlet explicitly confirms that reading, and the company has not clarified the exact rollout timeline.
What both reports agree on is that DeepSeek’s engineering effort goes beyond basic compatibility. According to Bloomberg, the company is performing hardware-specific tuning, reworking parts of the model’s computational pipeline to extract better performance from Huawei silicon rather than simply porting code written for Nvidia’s CUDA environment. That level of investment suggests DeepSeek views Ascend not as a stopgap but as a primary platform.
The performance question no one can answer yet
The biggest gap in the current reporting is performance data. Neither the AP nor Bloomberg published benchmark comparisons between V4 running on Huawei Ascend and V4 running on Nvidia hardware. Without those numbers, the practical significance of DeepSeek’s optimization work remains an open question.
A model that runs on domestic Chinese chips but performs significantly worse than its Nvidia-trained equivalent would be a political gesture. A model that closes the gap, or matches it, would represent something far more consequential: proof that China’s domestic semiconductor ecosystem can support frontier AI development.
Key technical details are also missing. Neither source specifies whether V4 was trained entirely on Ascend processors, fine-tuned on them after initial training on Nvidia GPUs, or simply deployed for inference on Huawei hardware. The distinction matters enormously. Full training on Ascend would signal deep confidence in domestic chips. An inference-only deployment would suggest Nvidia hardware remains essential for the most computationally demanding phase of model development.
Huawei’s Ascend chips and the domestic alternative
Huawei’s Ascend 910B and 910C processors are the most advanced AI accelerators manufactured in China. They were designed as direct competitors to Nvidia’s A100 and H100 GPUs, though independent benchmarks have generally shown them trailing Nvidia’s chips in raw throughput and the maturity of their software tooling. Huawei’s CANN (Compute Architecture for Neural Networks) framework serves as the Ascend equivalent of Nvidia’s CUDA, but it has a far smaller developer community and fewer optimized libraries.
DeepSeek’s decision to invest serious engineering resources in Ascend optimization could help change that calculus. If a high-profile model like V4 runs well on Huawei hardware, it validates the platform for other Chinese AI labs and application developers. Baidu, Alibaba, and ByteDance have all explored Ascend compatibility for their own models, but DeepSeek’s effort appears to be the most publicly visible commitment to date.
Huawei itself has not issued public statements about the scope of its collaboration with DeepSeek. Whether the partnership involves dedicated engineering support, custom firmware, or simply access to Huawei’s existing SDK is not established in available reporting.
U.S. export controls as an accelerant
Washington’s escalating chip restrictions have shaped this story at every turn. The U.S. Commerce Department first imposed sweeping controls on advanced AI chip exports to China in October 2022, then tightened them in October 2023 to close loopholes that allowed modified chips to reach Chinese buyers. Those rules cut off legal access to Nvidia’s A100, H100, and subsequent high-end GPUs for Chinese companies.
The intended effect was to slow China’s AI progress. The unintended effect, visible in DeepSeek’s V4 strategy, has been to accelerate investment in domestic alternatives. By forcing Chinese firms to plan around restricted access to Nvidia chips, U.S. policy has raised the strategic value of every viable domestic accelerator and given companies like DeepSeek a powerful incentive to accept short-term performance trade-offs in exchange for long-term supply-chain independence.
Bloomberg’s reporting reinforces this interpretation. According to its account, DeepSeek chose to delay a flagship product launch to prioritize Huawei chip integration. That is not the behavior of a company grudgingly adapting to sanctions. It is the behavior of a company that has decided supply-chain sovereignty is worth real competitive cost.
What this means for Nvidia
Nvidia’s immediate financial exposure is limited. The company has already lost most of its direct China revenue for high-end AI chips due to export controls. But the structural risk is harder to dismiss. If Chinese AI labs build mature, optimized software ecosystems around Huawei and other domestic accelerators, they reduce their dependence on Nvidia’s pricing, product cycles, and CUDA platform over time.
Today, the global AI research community overwhelmingly assumes Nvidia’s CUDA as the default environment. Training scripts, open-source libraries, and optimization techniques are built around it. DeepSeek’s Ascend-specific work is a step toward fracturing that assumption, at least within China. If the effort succeeds, future Chinese models could be designed from the ground up for domestic hardware, making it difficult for Nvidia to recapture the market even if export controls were eventually relaxed.
Nvidia has not issued a formal response to the V4 development, according to available reporting. The company’s silence likely reflects both the early stage of DeepSeek’s release and a broader reluctance to draw attention to the growing viability of competing hardware platforms.
A test case still waiting for its verdict
The verified facts point in a clear direction: DeepSeek is making a serious, resource-intensive push to build its next flagship model around Huawei’s Ascend chips. The company has previewed V4 with Ascend support and, according to Bloomberg, delayed the full launch to get the optimization right. That commitment is real and worth watching.
But commitment is not the same as capability. Until independent benchmarks, developer feedback, or DeepSeek’s own technical disclosures reveal how V4 actually performs on Huawei hardware, the most important question remains unanswered: Can Chinese-made chips support frontier AI models that compete with the best Nvidia-trained systems?
If the answer turns out to be yes, DeepSeek’s V4 will mark a turning point in the global AI hardware race. If not, it will still stand as evidence that China’s AI industry is no longer waiting for permission to build its own path forward.
More from Morning Overview
*This article was researched with the help of AI, with human editors creating the final content.
