DeepSeek founder Liang Wenfeng said at a recent investor meeting that China’s AI gap with the United States comes down primarily to computing resources rather than talent shortages or technical limits. According to meeting notes obtained by Yicai, Liang said, “Our biggest gap with the U.S. lies in resources. All the differences we see in talent, model capability, and the application layer stem from the gap in compute resources.”
DeepSeek says it remains 12 to 18 months behind top U.S. AI firms
Liang said DeepSeek currently has compute equivalent to about 20,000 NVIDIA H-series GPUs. He also said the most advanced models globally have already reached 800 billion active parameters, while China’s top models are still only in the tens of billions, leaving a gap of roughly 10x.
Even if DeepSeek were to spend the full amount of its latest fundraising round of more than 50 billion yuan, or about $7.4 billion, on compute procurement, Liang said it still would not be enough to train a model at the same scale. Overall, he estimated that DeepSeek trails leading U.S. AI companies by about 12 to 18 months. At the same time, he said the company achieved comparable results with about one-twentieth of the compute used by rivals, and plans to use that as a benchmark to shrink the gap to 3 to 6 months under current chip constraints.
DeepSeek and Huawei are building a path outside CUDA dependence
The other side of the compute gap is software. The meeting notes show DeepSeek is working closely with Huawei to optimize its models for the Huawei Ascend computing stack. It is also using its in-house high-level programming language, TileLang, together with a proprietary compiler in an effort to reduce dependence on NVIDIA’s CUDA ecosystem.
According to analysis compiled by Citrini analyst Jukan, DeepSeek V3 was still trained on NVIDIA GPUs, but the technical route now under discussion has already cut back software-level dependence on CUDA in a meaningful way. Jukan said inference efficiency improved significantly, while performance loss was limited to 1% to 2%.
If TileLang and the in-house compiler are successfully ported to Huawei chips, Jukan said the ecosystem problem that has long weighed on Chinese chips could see an initial fix in about a year. If that happens, the main bottleneck left would be manufacturing capacity.
Hardware performance still trails, with a two-year lag in Jukan’s estimate
On the hardware side, Liang said Huawei’s AI Supernode 950 SuperPod is now competitive with NVIDIA’s GB200 and GB300 systems in both performance and price. He added that DeepSeek is expected to secure about 16,000 Huawei AI chips.
Jukan offered a more specific conversion ratio. At present, he said, it still takes about four Huawei cards to match the performance of one NVIDIA card, implying an overall technology lag of around two years. By that measure, the chip gap between China and the U.S. is roughly a 4x hardware performance gap plus a two-year time lag.
China’s open-source AI model boom is still reinforcing NVIDIA in the near term
Jukan also focused on NVIDIA’s relationship with China’s AI market. In his view, the broad circulation of open-source AI models in China is, for now, reinforcing NVIDIA’s dominance rather than weakening it. These models are trained on NVIDIA GPUs and released with NVIDIA hardware as the optimization target, which naturally makes NVIDIA GPUs the preferred platform for inference as well.
He said that dynamic protects neocloud providers that built inference businesses on NVIDIA GPUs, while also giving NVIDIA some leverage against hyperscalers such as Google and Amazon as they move faster on in-house ASIC development. “No one is running DeepSeek V4 or Kimi K3 inference workloads on TPU or Trainium, at least not yet,” Jukan said.
If China reaches compute independence, NVIDIA’s moat could weaken
Even so, Jukan said he is bearish on NVIDIA over the medium to long term. The concern, he argued, is not simply that NVIDIA could lose the China market. The larger risk is that once Chinese models are no longer trained and optimized on NVIDIA GPUs, hardware choices in the inference market may begin to open up as well.
That would raise the relative competitiveness of Google TPU and Amazon Trainium, while eroding the strategic position of neocloud companies that rely on NVIDIA GPUs. In that scenario, NVIDIA’s soft lock on the broader AI compute market through CUDA would start to weaken at its foundation.
Jukan said the DeepSeek investor meeting revealed more than one Chinese AI company’s anxiety over compute. It also pointed to a technical path that is starting to take shape. Liang’s description of the AGI roadmap, spanning chain-of-thought reasoning, AI agents, continual learning and self-evolution, was framed around a direction that does not put NVIDIA at the center.
It may still be too early to say CUDA’s moat is approaching its endpoint. But multiple companies are already trying to work around it, and that much is getting harder to dispute.