Huawei Technologies remains a generation behind US rivals in chip performance but is working to close the gap through techniques such as cluster computing, CEO Ren Zhengfei said in comments published by the Chinese newspaper People’s Daily.
Ren said the company invests $25 billion (180 billion yuan) annually in research and development, with a growing focus on compound semiconductors.
The remarks, Ren’s first public comments on Huawei’s progress in chipmaking, come as the company continues to face US export restrictions that have blocked access to advanced semiconductor tools and high-end chips since 2019.
Huawei has developed the Ascend line of AI processors, which are used in China as alternatives to chips from the US giant Nvidia. The US Commerce Department has warned that using Ascend chips anywhere in the world could breach export control regulations.
Workarounds under scrutiny
Ren said Huawei utilizes math to overcome physical limits, non-Moore’s Law methods to boost performance, and cluster computing to offset the shortcomings of single-chip devices.
Analysts say this approach makes Huawei’s chips suitable for a wide range of AI workloads.
For instance, cluster computing strategy allows it to compete at the server level on performance, said Shrish Pant, director analyst at Gartner. While the approach may be less efficient in terms of power consumption, it is effective for many applications.
“Performance on a chip level for Huawei’s 910C is roughly equivalent to Nvidia’s H100, and it can be a good alternative to now-restricted Nvidia’s H20 chips in China, provided the rest of the pieces fall in place,” Pant said. “Since Huawei cannot access the latest and greatest tech in semiconductor manufacturing yet, they are innovating in directions like non-Moore’s law approach, and one of the examples is architectural changes like joining two reticle-size GPU dies to double performance.”
The approach also aligns with Huawei’s strengths. In cluster computing, the key challenge often lies not in building large systems but in optimizing network performance across nodes to approach peak efficiency.
“Huawei is well known for its networking capabilities and may be using proprietary software-defined networking capabilities that can accelerate the cluster,” said Hyoun Park, CEO and chief analyst of Amalgam Insights. “And there are mathematical tricks, such as the well-publicized DeepSeek use of an 8-bit floating point for training rather than the 16-bit version often used by most AI vendors.”
By simplifying model training and applying mathematical techniques that trade some accuracy for efficiency, Huawei could offset limited processing capabilities by relying more on power availability and software optimization.
Caution over sanctions
Huawei also has a vested interest in lowering expectations for its hardware on a global basis, as it is trying to avoid as many US and ally-based restrictions as possible.
“There are no real short-term victories to be had for Chinese companies to declare superior technological capabilities to the US right now,” Park said. “Global enterprises are effectively being asked to choose between the US and Chinese tech ecosystems, especially in light of current global trade issues.”
Analysts say multinationals should adopt a dual-track strategy for compliance and risk management. This includes stricter due diligence to verify the origin of AI components in data centers, cloud deals, and hardware.
“MNCs should conduct dynamic risk assessments, regularly updating compliance practices in response to BIS Entity List updates, OFAC rules, and shifting geopolitical conditions,” said Manish Rawat, semiconductor analyst at TechInsights. “It’s also essential to reassess relationships with Huawei and third-party vendors potentially using Huawei IP.”
For companies active in both the US and China, maintaining separate IT systems or cloud stacks can help avoid regulatory clashes. Best practices include dedicated export compliance teams with AI expertise, clear communication with US regulators, and steering clear of Huawei infrastructure in sensitive sectors, Rawat said.