China’s beautiful achievements in AI have one obtrusive weak spot: entry to compute—the uncooked processing energy that fuels AI and depends on giant volumes of superior semiconductors. The U.S. presently has a tenfold benefit over China in complete compute capability, a niche that will solely widen over time. U.S. tech companies are pouring billions of {dollars} into new knowledge facilities and might reap the advantages of the newest chip developments from Nvidia and AMD or their very own self-developed AI chips.
In the meantime, the efficiency and quantity of international AI chips that Chinese language companies can get hold of have gone down over time because of more and more stringent U.S. export controls. Chinese language tech leaders equivalent to Tencent, Baidu, and DeepSeek have known as out compute constraints as a key bottleneck to sooner AI improvement.
Huawei’s domestically produced AI chips, often called the Ascend collection, may seem to be the apparent resolution to China’s compute challenges. However there’s a catch: Chinese language tech companies don’t wish to use Huawei’s chips, which lag behind their international counterparts, for coaching their AI fashions. In 2024, Chinese language firms purchased round 1 million Nvidia H20 chips in contrast with an estimated cargo of 450,000 Huawei Ascend 910B chips.
Solely a handful of state-backed firms in China have used Huawei chips to coach their fashions, together with iFlytek, SenseTime, and China Cellular. Chinese language firms are dragging their ft on switching to home AI chips regardless of stress from Chinese language central authorities businesses to take action.
Chinese language AI builders overwhelmingly choose utilizing Nvidia chips—even severely performance-degraded ones—and go to nice lengths to entry them. Lots of China’s high AI fashions at present are nonetheless skilled on Nvidia’s {hardware}, together with DeepSeek’s V3 mannequin and Moonshot’s Kimi K2 mannequin. In anticipation of the U.S. ban on Nvidia’s H20 chips, ByteDance, Alibaba, and Tencent rushed to spend $16 billion to stockpile roughly 1.3 million to 1.6 million H20 items.
On the finish of 2024, ByteDance had deliberate to spend $7 billion to entry Nvidia chips on servers outdoors of China. Chinese language tech firms have been scouring black markets throughout Asia in addition to e-commerce websites to accumulate banned Nvidia chips for as a lot as double their regular valuedouble the value. Chinese language patrons have even resorted to purchasing Nvidia’s RTX gaming chips as substitutes, although they aren’t designed for AI workloads, and smuggling onerous drives full of knowledge in another country to coach fashions on servers outdoors of China.
Why are China’s AI builders so reluctant to modify from Nvidia to Huawei, whilst their entry to Nvidia chips turns into more and more constrained?
First, Nvidia’s degraded chips on the market to China nonetheless outperform Huawei’s chips in some necessary dimensions. Huawei’s Ascend 910B chips use older HBM2E reminiscence expertise, providing solely two-thirds of the reminiscence capability and 40 % of the bandwidth of Nvidia’s H20 chips.
Huawei’s newer Ascend 910C chips, that are ramping up manufacturing this yr, provide 80 % of the H20’s bandwidth however nonetheless use the older HBM2E reminiscence normal that’s two generations behind essentially the most superior AI chips. This hole in reminiscence efficiency is especially necessary given the rise of reasoning fashions and inference, the place reminiscence bandwidth performs an important position.
A second key cause why Chinese language tech firms can’t simply give up Nvidia is similar cause American tech firms can’t, both: CUDA. Nvidia’s parallel computing platform, launched in 2006, has collected customers and is tightly built-in with PyTorch, the dominant AI framework, making a mature software program ecosystem that locks builders into Nvidia’s AI programs.
For Chinese language tech companies, switching away from Nvidia means rewriting code, abandoning this industry-leading infrastructure, and dropping entry to the purposes in CUDA libraries constructed up over years by world builders. Huawei’s alternate options—its CANN platform and MindSpore framework, launched in 2018 and 2019—are newer, much less mature, and tormented by technical points together with bugs, crashes, and overheating.
With a much smaller AI {hardware} person base than Nvidia’s programs, Huawei lacks the high-volume, real-world suggestions from main prospects wanted to quickly refine its chips and software program. Because of this, Huawei’s AI options are unable to reap the benefits of the sort of iterative optimization that made China a worldwide chief in different industries.
Whereas entry to Nvidia chips is changing into more and more troublesome, the availability of Huawei’s chips stays each constrained and unsure. U.S.-led export controls on semiconductor manufacturing tools to China have restricted the nation’s chipmaking capabilities.
Specifically, Huawei and SMIC have been struggling to ramp up manufacturing of superior chips on the 7-nanometer course of degree or beneath. A scarcity of entry to excessive ultraviolet lithography (EUV) machines from ASML and U.S. instruments for key duties equivalent to etching and deposition have made it troublesome for SMIC to fabricate superior chips exactly and reliably, protecting its manufacturing yield far beneath {industry} chief TSMC.
Whereas SMIC is making regular progress and Huawei is on observe to promote over a million Ascend dies this yr, Huawei additionally illegally procured greater than 2 million of TSMC’s logic dies, a core chip element, for its Ascend 910B and 910C chips in 2024. Satirically, as a result of Huawei is already closely sanctioned, it faces little punishment for skirting export controls on this method.
Chinese language firms are additionally cautious of the extra business and geopolitical dangers concerned with Huawei, which has been a frequent goal of the U.S. authorities for years. For instance, the U.S. Division of Commerce warned in Could that utilizing Huawei chips “anyplace on the planet” would violate U.S. export management guidelines earlier than later adjusting its announcement.
Huawei will not be solely a chip provider to Chinese language tech firms but additionally a robust competitor. Huawei is China’s second-largest cloud service supplier and has developed its personal open-source Pangu household of AI fashions. Different Chinese language tech firms are jostling with Huawei to offer cloud companies not simply inside China however globally as nicely—making Huawei’s chips an unpopular choice for companies competing in the identical house.
Nevertheless, this might all change if the U.S. makes the flawed selections.
Whereas Huawei’s chips have decrease bandwidth reminiscence efficiency than Nvidia’s H20 chips, Huawei’s 910B and 910C chips already provide larger complete processing efficiency (TPP) and higher power effectivity (TPP/watt) than the H20. Removed from being a “highly effective chip” as some have claimed, the H20 truly has worse computational efficiency and power effectivity than Nvidia’s older A100 chips, launched again in 2020. Furthermore, Huawei has been capable of enhance the computational efficiency of its Ascend chips, even after switching manufacturing from Taiwanese chip producer TSMC to China’s personal SMIC following U.S. export controls.
Maybe extra importantly, Huawei has been making vital progress on the degree of AI computing programs. Huawei lately unveiled its CloudMatrix 384 system, made up of 384 of Huawei’s newest 910C chips and a novel all-optical networking strategy. In keeping with SemiAnalysis, Huawei’s new CloudMatrix system outperforms Nvidia’s state-of-the-art GB200 NVL72 system on key dimensions, equivalent to compute energy (how briskly the chip can course of giant volumes of knowledge), reminiscence bandwidth, and built-in networking.
Whereas Huawei’s new system is much more expensive and energy-intensive than Nvidia’s counterpart, which can restrict buyer adoption, it marks a placing development in system-level efficiency, which can be much more necessary than particular person chip efficiency for scaling up giant AI compute clusters.
In a latest technical paper, Huawei has already proved that its new CloudMatrix system might be efficiently used to coach superior AI fashions. The pricing and power points will seemingly be manageable for Huawei because it continues to take a position closely in R&D and obtain vital state assist.
As Huawei’s AI programs proceed to enhance, U.S. export management insurance policies have to be fastidiously calibrated to keep away from pushing China’s AI {industry} too far. If China’s home AI chips proceed to enhance, whereas U.S. chips out there in China are additional downgraded by export controls, there might be a crossover level the place the efficiency of Chinese language chips clearly exceeds that of American chips out there in China.
The essential tipping level could possibly be if China’s largest tech firms, equivalent to Alibaba, Tencent, and ByteDance, throw their formidable sources towards working with Chinese language AI chipmakers. This is able to kick off a optimistic suggestions loop for China’s AI chipmakers, significantly Huawei, increase the software program libraries and instruments for creating a whole Chinese language AI hardware-software ecosystem. As soon as this course of is underway, it will additionally mark a degree of no return for American AI chipmakers like Nvidia within the China market.
There are already some indicators of this potential shift. DeepSeek and ByteDance are experimenting with utilizing Huawei’s AI chips to run their AI fashions. Ant Group, a by-product of Alibaba, is even testing using Huawei’s chips for mannequin coaching. Huawei’s Ascend developer neighborhood has grown almost tenfold prior to now 4 years, although it nonetheless stays far smaller than Nvidia’s.
Different Chinese language AI chipmakers moreover Huawei are making progress as nicely, together with Cambricon, Biren, Moore Threads, Enflame, and Hygon. Cambricon noticed its first-quarter income surge greater thanforty‑fold final yr and obtained a big order for its AI chips from ByteDance, Cambricon’s income is forecasted to develop 3.7 instances to five.5 billion yuan this yr in accordance with Goldman Sachs. Nvidia’s CEO, Jensen Huang, has stated that Nvidia’s market share in China has declined from 95 % to 50 %—a declare supported by different credible evaluation.
The U.S. wants a extra subtle strategy to export controls. The reversal of the H20 chip ban by the Trump administration was a step in the correct path. On the similar time, the White Home’s new AI motion plan accurately acknowledges that profitable the AI race with China relies on making the U.S. tech stack, together with its AI chips, the dominant platform for world AI improvement.
Semiconductor export controls will not be so simple as tightening the valve on a faucet. China’s AI chip dilemma isn’t just a {hardware} downside however an ecosystem one. Huawei now has entry to most of the key sources it must develop superior AI chips, together with financing and expertise. Nevertheless it’s lacking a big and devoted buyer base that’s dedicated to co-refining the software program and {hardware} Huawei presents.
A wise strategy to export controls would give attention to setting a efficiency threshold for AI chips that may be bought to China primarily based on a window between U.S. and Chinese language {hardware} capabilities. The efficiency threshold must be excessive sufficient to outperform China’s home {hardware} choices to make sure Chinese language builders stay on U.S. platforms. On the similar time, it must be low sufficient to take care of a big efficiency hole with {hardware} programs out there to American builders.
Ideally, this efficiency threshold would come with a buffer, equivalent to a 50 % efficiency benefit over Chinese language {hardware} programs on key metrics, in anticipation of enhancements in Chinese language {hardware} choices. A daily yearly replace, with advert hoc adjustments for surprising developments, would seemingly be enough to regulate for advances made in Chinese language AI chips whereas offering sufficient coverage stability for {industry} individuals.
The overarching coverage objective is obvious: Make sure the U.S. continues to steer the world in AI. By constraining China’s entry to cutting-edge chips with out pushing Chinese language AI builders to make the leap to China’s personal home chips, the U.S. can use export controls to assist make this a actuality.
The views expressed by the authors don’t characterize these of their affiliated establishments.
