Huawei's 950 SuperPoD and China's Domestic AI-Chip Ecosystem — The Historic Inflection Point in Liang Wenfeng's Call

In July 2026, a set of internal meeting minutes attributed to DeepSeek founder Liang Wenfeng began circulating online. The core judgment in those minutes is arguably more aggressive than most industry analysis: domestic AI-chip substitution is sitting in a historic window — and that window is closing fast.

Why NVIDIA's CUDA Moat Is Crumbling

Liang Wenfeng's first claim strikes at the heart of the industry consensus — NVIDIA's CUDA ecosystem was long regarded as an impassable barrier, but that barrier is now "crumbling rapidly." He cites three reasons:

First, AI itself can now write code. Building an ecosystem parallel to CUDA once required an army of engineers to port and adapt by hand — a near-impossible task. Now, using AI to construct that ecosystem, it becomes possible to "rewrite the whole thing" — NVIDIA's entire stack, in one pass. DeepSeek's in-house TileLang does precisely this: it writes CUDA operators in a high-level language and lets AI finish the port.

Second, the arrival of the dedicated-chip era. CUDA evolved out of gaming cards, and its underlying architecture is of a piece with gaming-card design. But AI chips are growing ever more specialized — whether at Huawei or at NVIDIA itself, the future belongs to dedicated chips no longer bound to CUDA. For a dedicated chip, CUDA's ecosystem advantage is drastically weakened.

Third, the self-fulfilling prophecy of ecosystem confidence. In Liang's view, the problem with domestic chips has never been weak hardware or a weak ecosystem — it has been insufficient production capacity. He predicts that "within the next year, one thing will be proven: the domestic-chip ecosystem is entirely fine." Once that fact is verified and confidence arrives, the remaining capacity problems will also be resolved step by step.

Huawei's 950 SuperPoD: A Drop-In Substitute in Performance and Price

Liang Wenfeng offers a remarkably specific verdict on Huawei's newly released Atlas 950 SuperPoD (the Ascend 950 super-node):

"The Huawei 950 super-node can completely substitute NVIDIA's GB200 and GB300 in both performance and price."

Built on the Lingqu interconnect protocol and a super-node architecture, the 950 SuperPoD reaches an industry-leading scale of 1,024 cards, delivering 1 EFLOPS of FP8 and 2 EFLOPS of FP4 compute, a 256 TB globally unified memory-addressing space, TB-class NPU interconnect bandwidth, and an ultra-low round-trip latency of 3μs.

Liang concedes the price is "expensive, but only expensively within limits" — a 50%, 100%, or even 200% premium is acceptable, because on tasks it is a full substitute: "every task the GB300 can do, the Huawei super-node can do, with the same latency." The only cost is that four Huawei cards stand in for one NVIDIA card, while lagging two years behind.

The key here is not the height of the performance metrics but substitutability — domestic chips have entered a stage where they can be deployed at scale, no longer an experimental alternative.

From Compute Anxiety to Power Constraints

Another of Liang Wenfeng's subtler judgments concerns the long-run bottleneck of compute. He notes that what matters most in AI comes down to three things: compute, algorithms, and data.

At the chip level there is a gap, but "technology and process iteration can close it" — what matters more is the electricity behind the compute. China holds a clear advantage in power, and the gap in electricity between the United States and China will only keep widening. In the long-run competition in AI, power may ultimately matter more than chips.

The capacity problem will persist this year, next year, the year after, and perhaps longer — but not necessarily after five years. He voices a cautious optimism: "I don't really believe that five years from now we'll still be stuck on the capacity problem."

The Open-Source Open Gambit and America's Compute Narrative

At a more macro level, Liang Wenfeng makes one sweeping judgment: whoever tries to rake in windfall profits from AI will die. He believes AI may ultimately account for 10% of human GDP, but that no one can monopolize it — "those who take less will always beat those who take more."

That judgment runs in the same vein as the open-source strategy of domestic models such as DeepSeek and Kimi K3. With 2.8 trillion parameters, a 1-million-token context window, and an open-source license, Kimi K3 struck directly at the valuation logic of America's closed-source models. Goldman Sachs, in its own report, is asking the same question: if Yang Zhilin can train a Kimi K3 that rivals the top closed-source models with limited compute, does AI really need this much compute after all?

Liang likens the affair to "a Star Wars for a new era" — the one being dragged into the quagmire, forced to burn money to clear the road, is now the United States. The question is not whether domestic chips can catch up, but whether America will keep playing this game at all.