In October 2022, the US Department of Commerce dropped a bombshell on the AI industry. New export controls banned the sale of Nvidia's A100 and H100 GPUs (Graphics Processing Units) to China — the chips that were powering virtually every major AI lab in the country. A year later, the rules tightened further, adding the H800 and even the downgraded H20 to the restricted list. The message was unambiguous: the United States intended to choke off China's access to the most advanced AI hardware.
Four years later, something unexpected has happened. China's AI chip industry has not collapsed. It has accelerated. A new generation of Chinese AI chips — from Huawei's Ascend to Biren Technology's BR100 to Cambricon's Siyuan series — is now powering AI training and inference at scale. These chips are not as good as Nvidia's best. But they are good enough — and they are improving faster than almost anyone predicted.
The Sanctions: A Timeline
To understand where China's AI chip industry is today, you have to understand what it is up against. The US export controls, administered by the Bureau of Industry and Security (BIS), have evolved through several waves:
October 2022: The first wave banned the A100 and H100, Nvidia's flagship data center GPUs. Nvidia responded by creating the A800 and H800 — compliance versions with reduced interconnect bandwidth — and sales to China continued.
October 2023: The second wave closed the A800/H800 loophole by adding a performance density threshold. Nvidia created the H20, a further-reduced version, but even that was eventually restricted. Advanced semiconductor manufacturing equipment — particularly EUV (extreme ultraviolet) lithography machines from Dutch firm ASML — was also banned from export to China.
December 2024: The third wave added advanced HBM (high-bandwidth memory) to the restricted list, cutting off China's access to the high-speed memory chips that are essential for training the largest AI models.
The cumulative effect was to deny Chinese AI labs the three things they needed most: the best GPUs, the best chipmaking tools, and the best memory. China's AI chip industry had to build everything with one hand tied behind its back.
Huawei Ascend: The Incumbent Goes All-In
Huawei was the natural leader of China's AI chip push. The company had been developing its Ascend series of AI processors since 2018, long before the sanctions hit. When the A100 ban arrived, Huawei already had a product in the market. The sanctions gave it something more valuable than technology: a captive market of desperate Chinese AI companies that had no other choice.
The Ascend 910B, released in 2023, is the workhorse of China's AI chip ecosystem. Built on a 7nm process by SMIC (Semiconductor Manufacturing International Corporation), it delivers roughly 256 teraflops of FP16 (16-bit floating point) performance — comparable to an Nvidia A100 in raw compute, though with significant caveats in software maturity and memory bandwidth. It is the chip that powers Huawei's Cloud AI services and is used by most major Chinese AI labs, including Baidu, iFlytek, and the state-backed Beijing Academy of AI.
The Ascend 910C, which began sampling in early 2026, pushes performance to roughly 400 teraflops, closing the gap with Nvidia's H100. More importantly, Huawei has built a full software stack — the CANN (Compute Architecture for Neural Networks) platform — that provides a TensorFlow and PyTorch-compatible development environment. This is the hardest part of competing with Nvidia. Nvidia's CUDA (Compute Unified Device Architecture) platform, with its 20-year head start and millions of developers, is the real moat — not the hardware. Huawei's CANN is not yet a CUDA replacement, but it is the closest thing China has.
The scale of Huawei's investment is staggering. The company's HiSilicon chip design unit has grown from roughly 7,000 employees before the sanctions to over 20,000 in 2026. Huawei Cloud now operates tens of thousands of Ascend clusters across China, and the company claims that Ascend-powered training can achieve 80-90% of the efficiency of Nvidia GPU clusters for most workloads. Independent benchmarks are mixed, but the consensus is that the gap is narrowing rapidly — from "unusable" in 2023 to "workable" in 2025 to "competitive" for many applications today.
Biren Technology: The Dark Horse
If Huawei is the incumbent, Biren Technology is the dark horse. Founded in 2019 by former AMD and Qualcomm engineers, Biren raised over $700 million before the sanctions hit. Its first product, the BR100, was supposed to be manufactured on TSMC's 7nm process and directly challenge the A100. The sanctions killed that plan — TSMC was forced to stop working with Biren almost overnight.
What happened next is a case study in Chinese chip industry resilience. Biren pivoted to SMIC's 7nm process, redesigned the chip around the available manufacturing capability, and launched the BR104 — a smaller but functional AI accelerator — in 2024. The BR104 delivers roughly 200 teraflops of INT8 (8-bit integer) performance, making it competitive for AI inference workloads. The full BR100, now manufactured on SMIC's enhanced 7nm process, began shipping in limited volumes in early 2026. Early benchmarks suggest it can match or exceed the A100 on certain AI training tasks, though it falls short on memory-intensive workloads due to HBM restrictions.
Biren's survival is significant for a reason beyond the technology. Before the sanctions, China's AI chip startups were considered heavily dependent on TSMC (Taiwan Semiconductor Manufacturing Company) and Western EDA (Electronic Design Automation) tools. Biren proved that a Chinese AI chip startup could lose access to the world's best foundry and the world's best design tools — and still ship a competitive product. It was not pretty. It was not fast. But it worked.
Cambricon and the Inference Frontier
Not all AI chips are for training. The larger market — and arguably the more strategically important one — is AI inference: running already-trained models to generate text, recognize images, or make predictions. And in inference, Chinese chipmakers have a natural advantage: the chips do not need to be as powerful, the software requirements are simpler, and the market is massive.
Cambricon Technologies, a spinout from the Chinese Academy of Sciences, is China's leader in AI inference chips. Its Siyuan 590, built on a 7nm process and designed specifically for large language model inference, can serve a 70-billion-parameter model at roughly 40 tokens per second — competitive with Nvidia's L40S inference GPU at a fraction of the cost. Cambricon's chips are used in over 100 Chinese data centers and power everything from Alibaba's recommendation engines to ByteDance's content moderation systems.
The inference market is where China's AI chip industry may ultimately break Nvidia's dominance. Training a large model is a one-time capital expense. Running inference on that model — serving millions of users — is an ongoing operational expense. The company that can deliver the lowest cost per token will capture the inference market, and Chinese chipmakers are pricing aggressively. A Cambricon inference card costs roughly $3,000, compared to $8,000-$12,000 for an Nvidia inference GPU. For Chinese cloud providers running billions of inference queries per day, that difference adds up to hundreds of millions of dollars.
The Software Moat: CUDA and Its Challengers
The hardest part of building an AI chip is not the hardware. It is the software. Nvidia's CUDA platform, launched in 2006, is the de facto standard for GPU programming. It has over 4 million developers, hundreds of optimized libraries, and two decades of accumulated optimization for every major AI framework. Any chip that wants to compete with Nvidia must either be CUDA-compatible — which requires a license Nvidia has no incentive to grant — or build an entirely new software ecosystem from scratch.
China's AI chipmakers are doing the latter, and the effort is enormous. Huawei's CANN, Biren's BIRENSUPER, and Cambricon's Cambricon Neuware are all attempts to build CUDA alternatives. They provide compatibility layers that allow PyTorch and TensorFlow code written for CUDA to run on Chinese chips with minimal modification. The compatibility is not perfect — performance drops on complex workloads, and some advanced CUDA features have no equivalent — but it is improving with every release.
The Chinese government is accelerating this effort through standardization. In 2025, the MIIT released a national AI chip programming interface standard that all Chinese chipmakers are required to support. The goal is to create a unified Chinese AI software ecosystem — similar to what CUDA did for Nvidia — that makes it easy for developers to target any Chinese AI chip without learning a new platform for each one. Whether this standard gains traction or becomes another government mandate that the market ignores remains to be seen.
The Road Ahead
China's AI chip industry still faces enormous challenges. The lack of EUV lithography means Chinese chips will remain one or two process nodes behind TSMC's leading edge for the foreseeable future. SMIC's 7nm process is impressive given the constraints, but TSMC is shipping 3nm chips at volume and developing 2nm. The HBM ban is an even bigger bottleneck — without access to Samsung and SK Hynix's latest memory, Chinese AI chips struggle on the most memory-intensive AI training workloads.
But the trajectory is clear. In 2022, China's AI chip industry was almost entirely dependent on Nvidia. In 2026, it has a diversified ecosystem of domestic chipmakers covering training and inference, with a software stack that is maturing rapidly. The gap with Nvidia is still measured in years — but it is no longer measured in decades. And the Chinese market, with its insatiable demand for AI compute and its government-mandated preference for domestic chips, is large enough to sustain a domestic AI chip industry even if it never catches up to Nvidia on raw performance.
The lesson of the sanctions is not that export controls are ineffective. It is that they are effective in the short term and counterproductive in the long term. By cutting off China's access to Nvidia's best chips, the US created the one thing that no amount of R&D (Research and Development) funding could have produced: a market. Chinese AI companies that would never have considered a domestic chip are now building their entire infrastructure on Huawei Ascend and Cambricon. Chinese software developers who would never have learned CANN are now writing code for it every day. The sanctions did not stop China's AI progress. They redirected it — and they may have accelerated it in ways that will only become clear with time.