Key Takeaways
- DeepSeek’s newest AI model is the first to be explicitly optimized for Huawei’s domestically produced chips, marking a tangible step toward China’s goal of technological self‑sufficiency.
- While the model uses Huawei hardware for inference (the fast, response‑generation phase), training still relies on Nvidia GPUs, showing that full independence remains a work‑in‑progress.
- The announcement coincides with the upcoming Trump‑Xi summit, giving Beijing added confidence that U.S. export controls have not halted its AI progress.
- U.S. officials warn that stringent controls are pushing Chinese firms to build an alternative AI stack, potentially creating a bifurcated global market.
- Huawei plans to release a training‑capable chip later this year, but performance parity with Nvidia’s latest offerings is expected only after another year of development.
- Chinese semiconductor maker SMIC struggles with yield and power efficiency, forcing Huawei to aggregate many weaker chips to achieve needed compute power.
- Analysts argue that export controls, while intended to curb China’s advancement, are actually accelerating domestic innovation and ecosystem integration between chipmakers and AI model developers.
DeepSeek’s Milestone Announcement
When the Chinese start‑up DeepSeek released its latest artificial intelligence model last month, it signaled a meaningful shift in the nation’s quest for homegrown technology. For the first time, DeepSeek declared that its new model had been optimized to run on chips made by Huawei, China’s leading telecom and equipment giant. This optimization represents a concrete milestone in Beijing’s long‑standing effort to develop advanced semiconductors domestically and reduce reliance on Western innovation, particularly the Nvidia GPUs that dominate the global AI hardware market.
Timing and Political Context
The timing of DeepSeek’s announcement is noteworthy. It came just before the scheduled summit between former President Donald Trump and Chinese President Xi Jinping, a meeting that has repeatedly touched on technology trade issues. By highlighting progress with Huawei chips, DeepSeek gave Beijing fresh confidence entering the talks, suggesting that U.S. export controls on Nvidia’s most powerful AI processors have not derailed China’s AI development trajectory. The announcement thus serves both a technical and a diplomatic purpose, reinforcing China’s narrative of resilience amid external pressure.
U.S. Export Controls and China’s Adaptive Strategy
U.S. export controls have long targeted the sale of cutting‑edge Nvidia chips to Chinese entities, aiming to limit Beijing’s ability to train state‑of‑the‑art AI models. However, rather than freezing China’s AI ambitions, these restrictions have compelled firms like DeepSeek and Moonshot AI to redesign their systems around the constraints. As Wei Sun, a principal AI analyst at Counterpoint Research in Beijing, put it, “U.S. export controls are not freezing China’s AI development; they are forcing China to build an alternative stack.” Companies are now exploring how their models can run on a broader range of processors beyond Nvidia’s offerings, turning a limitation into an incentive for domestic innovation.
Technical Detail: Inference on Huawei Chips
DeepSeek’s latest model leverages Huawei hardware specifically for the inference stage—the process whereby a trained AI model responds quickly and accurately to user queries. Inference typically demands less computational power than training, making it a more feasible target for early‑generation domestic chips. By optimizing for inference on Huawei’s processors, DeepSeek demonstrates that Chinese chips can already support practical AI applications, even if they are not yet capable of handling the most intensive training workloads.
Training Still Depends on Nvidia
Despite the inference breakthrough, DeepSeek continues to rely on Nvidia chips for training its model, according to two semiconductor industry sources who spoke anonymously. Training involves feeding massive datasets through the model to adjust its internal parameters, a process that requires far greater compute capacity and memory bandwidth than inference. The sources noted that DeepSeek still accesses Nvidia GPUs, likely through remote use of chips housed in data centers outside China, a loophole that allows Chinese firms to obtain high‑performance computing without directly importing the hardware.
Huawei’s Chip Roadmap
Huawei has acknowledged the current gap and outlined a clear development timeline. The company plans to release a chip capable of supporting AI training later this year. However, Huawei executives cautioned that it will take an additional year after that release before its training chips can match the performance of Nvidia’s current offerings. This staggered approach reflects the reality that closing the performance gap will require both hardware advances and software optimizations that take time to mature.
Implications for a Bifurcated AI Market
The growing split between Chinese and American AI infrastructure echoes a warning from Jensen Huang, Nvidia’s chief executive, who has argued that strict export controls will ultimately push Chinese firms to build domestic alternatives, leading to a bifurcated market: Chinese AI systems running on Chinese chips while the West continues to rely on American hardware. Huang contends that such a division could diminish U.S. influence over China’s AI industry, even as Nvidia stands to benefit from unfettered access to the Chinese market if restrictions were eased.
Challenges in Chinese Chip Manufacturing
Realizing Huawei’s ambitions hinges on the capacity of China’s semiconductor foundries, most notably Semiconductor Manufacturing International Corporation (SMIC). SMIC, which fabricates many of Huawei’s chips, has struggled to produce them at scale. The chips it manufactures tend to exhibit higher defect rates and greater power consumption compared with those made by established foreign rivals like TSMC or Samsung. To compensate, Huawei has adopted a strategy of stitching together large numbers of these less‑efficient chips to achieve the compute power of more advanced processors—a tactic that depends heavily on SMIC’s ability to ramp up volume production.
Export Controls as a Catalyst for Innovation
Analysts from TechInsights and former Biden administration officials note that while export controls have constrained China’s ability to produce the large volumes of advanced semiconductors needed for cutting‑edge AI, they have also spurred Chinese tech firms to innovate in new ways. Companies are redefining what constitutes success in the AI race. Rather than solely pursuing the raw compute power afforded by billions of dollars’ worth of top‑tier chips, firms like Huawei are betting that future competitiveness will stem from building an integrated ecosystem—combining chips, AI models, and applications—that is “good enough” for most real‑world uses. This ecosystem approach allows tighter co‑design between hardware and software, potentially yielding performance gains that offset raw hardware deficiencies.
Close Collaboration Between DeepSeek and Huawei
DeepSeek’s technical papers outline specific modifications that chipmakers could implement to improve performance with its models. When announcing its latest model, Huawei highlighted a “close collaboration of chip and model technologies from both parties.” Jacob Feldgoise, an analyst at the Center for Security and Emerging Technology at Georgetown University, described DeepSeek’s outreach as a call to the void: “DeepSeek is calling out into the void to Huawei and other companies, ‘Please make these changes so we can get better performance out of your chips.’” Such feedback loops are critical for domestic chipmakers aiming to tailor their products to the precise needs of leading AI models.
Outlook Ahead of the Trump‑Xi Summit
As the Trump‑Xi summit approaches, analysts expect that frustration over U.S. export controls will be a topic of discussion. Jiang Tianjiao, an associate professor at Fudan University in Shanghai, observed that “chip export controls have consistently been an issue China opposes,” but added that as China’s chip‑making capabilities improve, officials may be less inclined to interfere with efforts to reduce dependence on American technologies. The fate of Nvidia’s H200 chip—a high‑performance GPU that Trump cleared for sale to China two months after his last meeting with Xi—remains uncertain. Commerce Secretary Howard Lutnick told a Senate Appropriations Committee that no H200s had actually reached China, and Nvidia’s regulatory filings show zero revenue from H200 sales there to date. This ambiguity underscores the ongoing negotiation between Washington’s desire to oversee chip use in China and Beijing’s push for self‑reliance.
Conclusion
DeepSeek’s optimization for Huawei chips marks a tangible, though partial, step toward China’s ambition of establishing an autonomous AI stack. While inference tasks can now be handled domestically, training still leans on Nvidia’s superior hardware, highlighting the remaining gap in performance and yield. The evolving landscape suggests that U.S. export controls, intended to curb China’s advancement, are instead accelerating domestic innovation and encouraging a shift toward integrated, ecosystem‑based AI solutions. Whether this leads to a sustained bifurcation of the global AI market or eventually converges as Chinese chips catch up will depend on continued progress in semiconductor manufacturing, strategic collaboration between firms like DeepSeek and Huawei, and the diplomatic dynamics playing out at forums such as the upcoming Trump‑Xi summit. The coming months will be crucial in determining how far China can close the technological divide and what the resulting balance of power will look like in the AI era.