US Export Controls Just Supercharged Huawei's AI Chips

Nvidia's stranglehold on China's AI hardware market just cracked wide open, and the numbers coming out of Huawei right now are genuinely alarming to any American chip executive who thought export controls solved the problem. Huawei is now on track to ring up twelve billion dollars in AI chip sales this year—up sixty percent from last year—driven by orders from Alibaba, Tencent, and ByteDance, all racing to stockpile the domestic alternative to Nvidia after DeepSeek proved its V4 models could run entirely on Chinese silicon.

By the end of this video, you will understand exactly how a company under heavy American sanctions is accelerating faster than most analysts expected, why the DeepSeek V4 launch was the catalyst that turned Huawei's Ascend chips from a contingency plan into the primary procurement strategy for China's biggest tech companies, and what the United States government's own AI advisor is warning is coming next.

Here is the story. On May first, the Financial Times reported that Huawei expects its AI chip revenue to hit approximately twelve billion dollars in 2026, based on orders already secured, compared to seven point five billion recorded in 2025. That is a sixty percent year-over-year increase. Reuters noted it could not independently verify the details, but the reporting is consistent across multiple outlets and sources. The surge is not speculative. It is backed by confirmed orders for Huawei's Ascend 950PR processor, which entered mass production in March of this year. Huawei is also preparing to launch an upgraded version, the Ascend 950DT, in the fourth quarter of 2026.

Now, why does this matter right now, and why should you keep watching? Because what is happening in China's AI semiconductor market is no longer a regional story. It is a structural shift in global AI infrastructure. For the first time since sanctions began, Chinese cloud operators have moved from treating domestic chips as a backup plan to treating them as a primary procurement strategy. Huawei's growth is not driven by government mandates creating a forced developer base. It is driven by a market that has decided the domestic alternative is good enough—and in some cases, cost-effective enough—to deploy at scale. This matters because a parallel AI ecosystem in China, with its own hardware, software stack, and developer network, is the exact outcome Nvidia's leadership has privately warned about. A bifurcated world where the United States runs CUDA and China runs something else is not a theoretical risk anymore. It is the present reality, and Huawei is becoming the anchor of that parallel stack.

Let me give you the background. Huawei's Ascend line has been in development since around 2019, but it truly accelerated after the United States imposed increasingly severe export controls on advanced AI chips to Chinese companies. The restrictions limited access to Nvidia's most powerful processors, forcing Chinese firms to look elsewhere. Huawei responded by investing heavily in its own chip architecture, the Ascend series, built around the Compute Architecture for Neural Networks—or CANN—as a software alternative to Nvidia's CUDA. This was a multi-year bet. For years, the limitations were clear. Chinese vendors struggled with low yields on advanced process nodes, an unstable supply of high-bandwidth memory, and the overwhelming dominance of Nvidia's CUDA ecosystem, which had become so embedded in global AI development that migrating away from it carried real friction. But the pressure of sanctions, combined with Beijing's push for technological self-reliance, created the conditions for rapid iteration.

The turning point came in early 2026. Huawei debuting the Ascend 950PR at the China Partner Conference. The chip uses the same die architecture as the original Ascend 950 but is fundamentally stronger. It offers 2.87 times the computing power of Nvidia's export-compliant H20 processor. More importantly, the 950PR is the only AI chip in China to feature FP4 low-precision inference, which allows it to process AI calculations in a more compressed numerical format. This means more computations per second at lower cost. It uses HiBL 1.0 and low-cost HBM that is more cost-effective than HBM3E and HBM4E, significantly reducing the hardware investment Chinese companies need for prefill and recommendation systems. The prefill stage of inference is compute-intensive with high demand for parallel computing and lower demand for memory access bandwidth, and the 950PR was specifically designed to address that exact workload.

Then came April 24th, 2026, when DeepSeek launched its V4 model series, and this is the critical inflection point. The V4-Pro model has roughly 1.6 trillion parameters with 49 billion active. The V4-Flash has 284 billion parameters with 13 billion active. Both run entirely on Huawei Ascend chips. They ditched Nvidia. The V4-Pro is DeepSeek's biggest model ever by metrics and is optimized for complex reasoning, math, and coding. The V4-Flash is built for speed and cost efficiency. DeepSeek confirmed that the Ascend 950 SuperPod supports both models, achieving low-latency inference of twenty milliseconds for V4-Pro and ten milliseconds for V4-Flash. DeepSeek also said its Ascend chip-powered V4-Pro can outperform other open-source models in world benchmarks alongside Google Gemini Pro 3.1.

The DeepSeek launch triggered a wave of procurement activity. ByteDance, Tencent, and Alibaba were already placing orders for Huawei Ascend chips in March, but the V4 launch sent demand to new heights. Sources indicate that cloud computing and GPU rental services are also rushing to secure supply. Huawei confirmed that its entire Ascend SuperNode product line now supports the DeepSeek V4 series.

Now let's look at the data that puts this in perspective. According to an IDC report reviewed by Reuters and published on April first, 2026, around four million AI accelerator cards for servers were shipped in China in 2025. Nvidia led with about 2.2 million units, roughly 55 percent market share. AMD accounted for about 160,000 cards, or four percent. Chinese vendors collectively reached approximately 1.65 million cards, accounting for about 41 percent. Within that group, Huawei led with roughly 812,000 units, followed by Alibaba's T-Head with about 265,000 cards, and Baidu's Kunlunxin and Cambricon each at approximately 116,000 cards. That 41 percent for domestic vendors in 2025 was a milestone. It was the first time Chinese-made chips captured such a dominant share of the market.

And the trajectory is accelerating. DIGITIMES analyst Stella Weng estimated that Chinese vendors will ship 2.123 million high-end cloud AI accelerators in 2026, representing a 136 percent year-over-year increase. Huawei is expected to capture more than 50 percent of the domestic market in 2026, far ahead of competitors.

But there is a tension here, and it is real. China's total domestic AI accelerator market is growing fast, but Chinese vendors' global market share is projected to reach only about 13 percent, constrained by low yields on advanced process nodes, an unstable supply of high-bandwidth memory, and the continued dominance of Nvidia's CUDA ecosystem outside of China. Huawei's cloud computing revenue from external customers actually fell 3.5 percent in 2025 to 32.16 billion yuan, approximately 4.6 billion dollars. The ICT infrastructure segment, which includes Ascend AI chip solutions, reported revenue growth that slowed to 2.6 percent, down from 4.9 percent in 2024. Huawei is building something significant, but it is building it under constraints that are far from eliminated.

The supply chain bottleneck is particularly acute. High Bandwidth Memory remains the biggest constraint. HBM supply, along with other supply-chain limits, could slow this expansion even as demand surges. Huawei is trying to work around this with lower-cost HBM in the 950PR, but the physical limitations of advanced manufacturing nodes within China are a hard ceiling that no amount of demand can instantly overcome.

On the industry landscape, Nvidia still leads the Chinese market, but it is clearly losing ground. The shift is structural. Huawei is positioning its Atlas 950 SuperPoD and TaiShan 950 SuperPoD, debuting at MWC Barcelona, as alternatives to Nvidia's full stack. The Atlas 950 SuperPoD is expected to run on the next-generation Ascend 950DT chips and be available by the fourth quarter of 2026. Huawei has proposed its open-source Compute Architecture for Neural Networks as an alternative to CUDA and is actively inviting developers to collaborate. The question is whether open-sourcing CANN can overcome CUDA's entrenched position. Nvidia has warned that the split between CUDA and CANN ecosystems could help local companies build their own developer and customer networks that become self-reinforcing over time. Once a developer base is trained on a platform, migration costs grow exponentially.

What is next, and what should you watch for. First, the fourth quarter launch of the Ascend 950DT. This is the upgraded processor that will determine whether Huawei can maintain momentum beyond the initial surge. Second, DeepSeek's full stable release of V4 models, not just the preview versions currently available. The preview is an early look, and the final product will be the true stress test. Third, any announcements about Huawei exporting Ascend chips globally. David Sacks, the AI technology and policy advisor to the US president, predicted in an interview that Huawei may soon begin shipping its AI chips globally. He stated that China has reduced the AI semiconductor design gap with the United States from five years to merely one and a half to two years in the field of AI chipset design and packaging technology. Sacks said Huawei is one of the companies making significant strides and is very likely to become an indispensable hardware supplier in the global market. Whether that prediction materializes will depend on overcoming the HBM and manufacturing constraints I mentioned. Fourth, watch Huawei's CANN ecosystem adoption metrics. Developer migration from CUDA is the ultimate battleground, and the numbers on that front will tell you whether Huawei is building a platform or just a product. And fifth, watch Nvidia's response. If Chinese cloud operators are committing hundreds of millions of dollars to Ascend infrastructure, Nvidia's H20 and future export-compliant chips face a fundamentally different demand curve than they did even six months ago.

China is building a parallel AI stack. Huawei's twelve-billion-dollar chip revenue projection is not just a company milestone. It is evidence that American export controls, while effective at slowing specific products, are simultaneously accelerating the creation of a competing ecosystem that could be self-sustaining. The gap is closing. And once it closes, it may not open again.

Stay sharp. Jane Sterling, Sterling Intelligence.