DeepSeek CEO: Nvidia Is Quietly Digging Its Own Grave

At an investor meeting, Liang Wenfeng did not mince words. The CEO of DeepSeek painted a raw, almost cinematic picture: China is sprinting to close a compute gap that once looked unbridgeable, and the old leader may be surprised by how fast the field moves.

DeepSeek currently runs the equivalent of about 20,000 Nvidia H100 chips, most of which arrived in the past month or two. That capacity sounds huge. It is. But it still only lets them activate roughly 10 billion parameters at a time. Compare that with the top-tier models today that can keep around 800 billion parameters live simultaneously. The difference is not small. It is structural.

How Huawei fits into the equation

Liang singled out Huawei's Atlas 950 nodes, powered by Ascend 950 processors, as a real alternative. His claim was provocative: although a single Ascend 950 is less efficient than an Nvidia H100, in practice a cluster of these Huawei nodes can match the end-to-end performance of Nvidia's GB200 and GB300 racks. The math, he said, looks roughly like four Ascends for every one Nvidia chip, but total throughput and latency can align close enough that the substitution is feasible.

Nvidia is digging its own grave, he warned. Price differences matter less, Liang argued, because even if Huawei hardware were 50, 100, or 200 percent more expensive, the capability to deploy at scale and the political advantages of a domestic supply chain could outweigh pure cost comparisons.

There is a catch. Training state-of-the-art large models is brutally hungry. To train a model on the scale of the current AI giants, Liang estimates you would need about 50,000 GB300-class Nvidia processors or roughly 200,000 Ascend 950 chips. Those are training numbers only; research and experimentation add their own compute demands on top. This helps explain why the compute resource gap between China and the United States remains the largest single obstacle.

So what is DeepSeek doing about it? For now, the firm is expanding its infrastructure aggressively. But Liang also stressed a different battleground: data. Quality labeling, he said, is the slow choke point. You can throw chips at a problem, but if your labels and curated datasets lag, progress stalls. Companies such as OpenAI and Anthropic have a head start here because they moved earlier and poured more resources into high-quality annotation.

The takeaway is not just about chips or national pride. It is about the constellation of hardware, software, talent, and time. China is building the compute capacity. It is also investing in the time-consuming, less glamorous work of dataset hygiene and labeling. If both tracks advance, the map of AI competitiveness could look very different within a few years.

Liang's message was blunt but strategic: this is not a single battle over a single chip. It is a long campaign where alternatives that once looked second-rate can become decisive once scaled and integrated into a full stack. That shift may be the moment Nvidia will look back on and call pivotal.