Key Takeaways
- DeepSeek CEO Liang Wenfeng states Huawei’s Ascend AI processors trail Nvidia by roughly two years, with performance parity on the horizon.
- Huawei’s Atlas 950 SuperPoD matches Nvidia’s GB300 in performance per price but demands four times the chip count, underscoring resource constraints.
- Market research from Bernstein projects Huawei will capture about 50% of China’s AI chip market in 2026, while Nvidia’s share drops to around 8%.
- Huawei’s Ascend chips still lag in training frontier models, but show strong inference performance relative to Nvidia’s H20 according to company claims.
DeepSeek CEO Puts Huawei’s AI Chip Gap at Two Years, but Market Forces Are Accelerating the Catch-Up
Liang Wenfeng, chief executive of Chinese AI startup DeepSeek, stated during an investor conference call that Huawei’s Ascend AI processors are approximately two years behind Nvidia’s latest offerings. He also noted that the gap is closing quickly as Huawei’s Atlas 950 SuperPoD can match the performance of Nvidia’s GB300 NVL at a competitive price, though it requires four times the number of chips. The remarks come amid a rapid realignment of China’s AI chip market.
Table of Contents
Inside DeepSeek’s Technical Assessment of Huawei vs Nvidia
During the same call, Wenfeng outlined specific product comparisons. DeepSeek found that Huawei’s Atlas 950 SuperPoD delivers equivalent performance to Nvidia’s GB300 NVL in terms of both capability and price, according to the executive. However, the critical detail is that for every Nvidia GB300 processor, DeepSeek would need four Ascend chips to achieve comparable results.
Wenfeng stated that training a model similar in scale to existing frontier AI systems would require roughly 50,000 Nvidia GB300 units or 200,000 Huawei Atlas 950 processors. This four-to-one ratio represents a significant hardware requirement that DeepSeek must factor into its infrastructure planning.
The DeepSeek CEO also highlighted that Nvidia is effectively ‘digging its own grave‘ by limiting access to its most advanced chips in China, according to Huawei Central. As Huawei’s solutions become more cost-effective, demand for Ascend processors is surging, accelerating Huawei’s market penetration.
Why Market Data Points to a Dramatic Power Shift
Independent market research from Bernstein, as reported by AP NEWS, provides a quantitative backdrop to Wenfeng’s qualitative assessment. Bernstein analysts estimate that in 2025, Nvidia and Huawei each held approximately 40% of China’s AI chip market. For 2026, however, the projections shift dramatically: Nvidia’s share is expected to fall to around 8%, while Huawei’s rises to about 50%.
In terms of technical capability, analysts view Huawei’s most advanced commercial AI chips, the Ascend 950 series, as roughly comparable to Nvidia’s H200 — a product that Nvidia has never sold in China and from which it has generated no revenue due to export controls. Meanwhile, DeepSeek has already adapted its latest V4 model for Huawei’s Ascend architecture, signaling a deepening reliance on domestic hardware.
According to Tom’s Hardware, Huawei claims that its Ascend 950PR accelerator delivers approximately 2.87 times the inference performance of Nvidia’s H20 at a quarter of the cost — a claim that has not yet been independently benchmarked at production scale. Industry analysts caution that while Huawei’s inference performance is competitive, its chips still lag significantly in training cutting-edge AI models, a crucial area where Nvidia’s architecture retains a clear lead.
What This Means for AI Infrastructure Procurement
The convergence of DeepSeek’s internal testing and independent market data paints a clear picture: Huawei’s Ascend chip ecosystem is maturing rapidly, and for many AI inference workloads, it is already a viable alternative to Nvidia. But the four-to-one chip ratio and the lag in training performance mean that the transition will not be seamless for compute-intensive frontier model training.
For AI startups and enterprises in China, the calculus is shifting. While Nvidia remains the gold standard for training, the combination of export restrictions and rising domestic competition is forcing a reevaluation of supply chains. Andres SEO Expert closely monitors these developments to help technology leaders make informed decisions about infrastructure and AI strategy.
If you are navigating these changes and need to optimize your AI infrastructure or enhance your digital presence, get in touch with Andres. And if you are looking to build or scale AI-driven automation and programmatic workflows, explore how our AI automation services can streamline your operations. For more insights into our approach, visit Andres SEO Expert to learn how we blend technical analysis with strategic execution.
Frequently Asked Questions
What did DeepSeek’s CEO say about Huawei’s AI chip gap compared to Nvidia?
Liang Wenfeng stated that Huawei’s Ascend AI processors are approximately two years behind Nvidia’s latest offerings, but noted the gap is closing quickly as Huawei’s Atlas 950 SuperPoD matches Nvidia’s GB300 NVL in performance and price, though requiring four times more chips.
How many Huawei Ascend chips are needed to match one Nvidia GB300 processor?
According to DeepSeek’s testing, four Ascend chips are required to achieve comparable performance to a single Nvidia GB300 processor. For frontier-scale models, DeepSeek estimates needing 50,000 Nvidia GB300 units or 200,000 Huawei Atlas 950 processors.
What are the projected market shares for Nvidia and Huawei in China’s AI chip market in 2026?
Bernstein analysts project that by 2026, Nvidia’s share will fall to around 8% (from ~40% in 2025), while Huawei’s will rise to about 50% (from ~40% in 2025), reflecting a dramatic shift driven by export controls and domestic adoption.
How does Huawei’s Ascend 950PR compare to Nvidia’s H20 in inference performance?
Huawei claims the Ascend 950PR delivers approximately 2.87 times the inference performance of Nvidia’s H20 at a quarter of the cost. However, these claims have not been independently benchmarked at production scale, and analysts caution that training performance still lags behind Nvidia.
What are the key limitations of Huawei’s chips for training AI models?
While Huawei’s inference performance is competitive, its chips still lag significantly in training cutting-edge AI models. Nvidia’s architecture retains a clear lead for compute-intensive frontier model training, making the transition not seamless for such tasks.
Why are Chinese AI companies shifting towards Huawei’s chips despite performance gaps?
Export restrictions limit access to Nvidia’s most advanced chips in China, while Huawei’s solutions are becoming more cost-effective and competitive for inference workloads. Market forces, including surging demand for Ascend processors, are accelerating Huawei’s market penetration.
How has DeepSeek adapted its models to Huawei’s Ascend architecture?
DeepSeek has already adapted its latest V4 model for Huawei’s Ascend architecture, signaling a deepening reliance on domestic hardware. This adaptation allows DeepSeek to leverage Huawei’s chips for inference while planning infrastructure around the higher chip count required.
