DeepSeek overturns the impression that domestic chips are not user-friendly: Huawei's 950 supernode fully replaces Nvidia's GB300 and others.

Recently, DeepSeek's leader Liang Wenfeng's internal remarks sparked a major uproar online. The composure and confidence displayed in his speech, along with his deep analysis of the entire industry, caught many longtime observers off guard.

Unlike the practical concerns outsiders typically focus on—such as commercialization and profitability—Liang's current strategic direction appears aimed at pairing mature, iterated domestic chips with fully self-developed domestic AI large models to directly challenge the global AI technology narrative that the U.S. has maintained for years.

Based on the leaked internal remarks, Liang's outlook on the development prospects of domestic AI chips is far more optimistic than most industry observers' assessments.

He believes Nvidia's traditional CUDA moat is collapsing faster than expected, citing three core reasons: AI can now autonomously generate code to build complete ecosystems, newly introduced advanced programming languages have significantly lowered the difficulty of rewriting operators, and the AI computing card market has fully surpassed gaming cards in scale, leaving Nvidia with no reasonable justification to forcibly bind the ecosystems of these two product lines together.

These technological shifts, combined with the reality of U.S. chip export controls, have effectively created an unprecedented historical window of opportunity for the domestic AI chip industry.

He made a clear prediction that within the next year, the entire market will completely overturn the entrenched perception that "domestic chips are simply unusable." DeepSeek has already entered deep collaboration with Huawei, securing 16,000 Ascend 950 cards. While the overall equivalent computing power only matches about 4,000 Nvidia B-series chips—insufficient to support training the next-generation flagship model—it is more than enough to help Huawei validate and run its entire computing ecosystem.

He also candidly acknowledged that Huawei's core issue has never been technical capability, but rather production ramp-up speed. Four Huawei chips roughly match the comprehensive performance of one Nvidia card, with paper specifications trailing by about two years. However, as long as pricing and ecosystem can achieve parity, this level of performance gap is entirely acceptable across the industry.

In this industry leader's view, the ecosystem concerns outsiders previously raised about domestic chips are far from a deadlock. In the past, people generally believed domestic chip ecosystems were immature, unusable, and difficult to work with. But he is confident that within at most a year, the entire team will overturn the industry's entrenched impression through tangible, real-world results.

He stated bluntly that the hardware performance and supporting ecosystem of domestic AI chips no longer have fundamental issues—the only bottleneck dragging things down is production capacity failing to keep up.

There is no technical barrier to adapting domestic large models to domestic AI chips, and Nvidia cannot stop this trend. In a fully open, normal commercial environment where everyone could freely purchase Nvidia's high-end chips, advancing domestic substitution would indeed be more difficult. But now that Nvidia's high-end chip exports to China are completely restricted, all domestic AI practitioners are forced to pivot to self-developed computing power—effectively creating the most direct fertile ground for nurturing the domestic ecosystem.

Liang admitted he is very optimistic about the development of the domestic computing power track. Based on current trends, Nvidia's export control policy toward China is essentially digging its own commercial grave. Huawei's super-node architecture—the computing cluster built on Ascend 950—can fully achieve parity with Nvidia's flagship GB200 and GB300 computing products in terms of comprehensive performance and cost-effectiveness.