China's AI Chip Ecosystem at a Pivotal Moment, Viability Set for Proof Within a Year

During a recent investor exchange, DeepSeek founder Liang Wenfeng outlined a transformative outlook for domestically developed AI chips. He revealed the company's current compute capacity is roughly equivalent to 20,000 H-series chips, with most of this infrastructure deployed within the last couple of months.

Resource Gap Persists, Talent Pool is Comparable

"The gap with the United States is primarily in resources, translating to a lag of about 12 to 18 months, or perhaps 6 to 12 months," Liang analyzed. "However, there's almost no gap in talent, as we're essentially drawing from the same pool of technical experts."

New Technologies Are Eroding the CUDA Moat

Liang believes the long-standing ecosystem advantage held by NVIDIA's CUDA is undergoing a fundamental shift, driven by two key forces:

  • AI Reshapes Development: "With AI now capable of writing code, we can use AI itself to build the ecosystem," he stated, suggesting a redefinition of development barriers and efficiency.
  • Breakthroughs in Programming Languages: He highlighted emerging tools like TileLang, an open-source, high-performance AI operator programming language developed by a Peking University team. "Using such high-level languages to rewrite CUDA operators allows for rapidly recreating the entire ecosystem."

Market dynamics are also fostering decoupling. "The compute card market was once smaller than the gaming card market, but that has reversed. When compute cards become the larger segment, the necessity for them to remain bundled diminishes."

The Coming Validation of Ecosystem Viability

Liang characterizes the current phase as a "historic opportunity." "We believe that within the next year, a critical assertion will be validated: the ecosystem for domestic chips is completely viable." He acknowledged the past perception of domestic chips being "unusable or difficult to use," but expects this view to be overturned within 12 months.

Practical Deployment and Trade-offs

Discussing specific hardware, Liang used a domestic computing card as an example. "Our procurement aim is to help foster its ecosystem." He noted that its super-node variant can match the performance and price point of comparable NVIDIA products.

"The price is somewhat higher, but not prohibitively so. The main trade-off is requiring four domestic cards to match the compute of one NVIDIA card, coupled with a release timeline lag of about two years," he added, comparing currently available domestic hardware to NVIDIA's offerings from two years prior.

From a total cost of ownership perspective, Liang offered a practical comparison: "NVIDIA cards are typically depreciated over five years, while domestic cards might be over three years at most. The performance is good this year, acceptable next year, but energy consumption could become a more significant factor thereafter."