The Financial Times broke the number on Friday and it landed harder than anyone in Washington wanted to admit. Huawei expects roughly $12 billion in AI chip revenue in 2026, up from $7.5 billion this year, a 60 percent jump in twelve months. By year end, the company expects to control 60 percent of China's domestic AI chip market. The other 40 percent goes to a mix of SMIC, Cambricon, and Biren. Nvidia's share, by Jensen Huang's own admission this week, is zero.

Read that one more time. The most valuable semiconductor company in human history sells nothing into the second largest economy in the world. Not because Chinese buyers do not want H100s. Because the United States Commerce Department made the sale illegal.

Huawei has been preparing for this exact outcome for five years.

The Ascend 950PR Is the Real Story

Huawei's flagship Ascend 950PR entered mass production in March. The spec sheet is what matters. Two PFLOPS of FP4 throughput, 128GB of local HBM, native MXFP4 support, and the ability to run DeepSeek V4 on a single chip without partitioning. The 950DT successor is on schedule for Q4 2026 mass production with 144GB of HBM and an upgraded interconnect. None of that requires Nvidia, TSMC's most advanced process node, or any technology covered by the BIS export rules.

The DeepSeek V4 angle is the part Western analysts keep missing. DeepSeek's team optimized V4's kernels and quantization specifically for the Ascend instruction set. Not Nvidia CUDA. Not even as a fallback. The model that triggered the January 2025 panic in San Francisco now runs natively on Huawei silicon and runs faster on Huawei silicon than on the H800s the export controls were trying to ration. The hardware is good. The software is co-designed. The customers are local. The supply chain is closed.

That is not a chip company catching up. That is a chip company that already won.

How the Export Controls Produced the Opposite of Their Goal

The original BIS rules in October 2022 were narrow. They restricted the highest end Nvidia chips and the most advanced lithography tools. The theory was that Chinese AI labs could not train frontier models without American hardware and could not buy American hardware. By 2025 that theory was obviously broken. By 2026 it is dead.

Three things happened. First, Chinese labs proved you do not need an H100 cluster to train a frontier model. DeepSeek did it on H800s. Manus did it on a mix of H800s and Ascend 910s. Qwen 3 did it on Ascend 920s. The performance gap, once two generations, is now closing inside one. Second, the controls created the strongest possible incentive for Beijing to subsidize a domestic alternative. They did. Roughly $48 billion in state-backed silicon funding flowed into Huawei, SMIC, and a network of design houses between 2023 and 2026. Third, the controls cut Nvidia out of any ability to influence the Chinese roadmap. Once the customers were forced to switch, switching back became a strategic risk for them. Why rebuild a software stack on Nvidia when the rules might tighten again next year?

Jensen Huang has been saying this on every earnings call since Q2. The June 2025 print was the first time he said it without hedging. The October print made it explicit. The Q4 print will quantify the lost revenue. By the FT's math, Huawei's $12 billion in 2026 is a $12 billion line item that used to flow to Nvidia.

It is also a $12 billion line item that creates the manufacturing scale Huawei needs to push the next generation. The 950DT in 2027 will not be a copy of an H100. It will be a chip designed inside the constraints of Chinese fab capacity, optimized for the workloads Chinese hyperscalers actually run, and priced to undercut whatever Nvidia is allowed to sell into the rest of Asia.

What Comes Next

The export controls were a four year bet that the United States could prevent China from achieving AI compute parity. Four years in, the bet has lost. China has parity. Not at the bleeding edge. Not for every workload. But for the workloads that matter for production deployment, training mid-frontier models, and serving inference at scale, the Ascend 950 series is enough. Huawei does not need to beat Nvidia. It needs to be enough. It is.

The next phase of this story is exports. Not to China. From China. Huawei has begun pitching Ascend systems to Saudi Arabia, the UAE, Indonesia, and Malaysia. Those countries used to buy Nvidia. Now they have a second supplier with comparable performance, an unrestricted license, and a state actor willing to subsidize delivery. The diffusion rule, the BIS framework that limits AI chip shipments to most of the developing world, was built to slow this exact scenario. It is going to fail too, for the same reason the China controls failed. You cannot ban your way to a customer base.

The cleanest read on the FT story is that the chip war is over and Huawei won the inside lane. The American policy goal of preventing a Chinese AI compute stack failed completely. There is now a fully domestic Chinese stack, and as of this quarter, that stack is profitable. Whoever in the Commerce Department is still drafting tighter rules should read the Huawei revenue line one more time, then ask whether more rules will produce more lost market share or less. The answer should be obvious.

Nvidia just got proof. Zero is not a typo.