Chinese chip designers including Tencent Holdings are aggressively pitching their AI chips as an alternative to Nvidia's, hoping that U.S. export restrictions will encourage customers to switch, four people familiar with such discussions said.
California-based Nvidia controls as much as 90% of China's $7 billion market for chips used to process massive amounts of data to develop artificial intelligence (AI) software.
However, US strategic technology controls that intensified in October have encouraged even smaller names, such as state-backed Hygon Information Technology and startup Iluvatar CoreX, to take on the US giant.
Huawei Technologies is generally considered to be making the most progress, with the Ascend 910B being compared to Nvidia's A100 in terms of computing power, but not in overall performance.
But Tencent and smaller AI players are accelerating chip product launches and staging more marketing pushes, betting that while the U.S. rules affect only the most advanced chips, they could still keep customers away from Nvidia, the people said.
Tencent, China's largest social media and gaming company, which also sells cloud services to third-party customers, has promoted services using the AI inference chip Zixiao developed with deep learning startup Enflame, touting performance comparable to that of some Nvidia chips, two of the people said.
One of the people said Tencent is pitching its Zixiao v1 variant as a cheaper alternative to Nvidia's A10, which is used for AI image and speech recognition applications. It's also promoting an upcoming v2Pro variant optimized for AI training that could replace Nvidia's now-locked L40S, the person said.
Tencent uses Zixiao chips internally and does not sell them directly to external customers. It rents computing power to customers through its cloud services, which also offer Nvidia or AMD chips as an option.
Tencent said it has no plans to develop Zixiao beyond the current version.
“Tencent designed Zixiao to complement our cloud products and solutions in compliance with laws and regulations. It is only available to customers through Tencent Cloud's enterprise services,” a company spokesperson said.
Others pitch direct sales. Tencent-backed Enflame, which has an AI training accelerator chip called Yunsui, and Iluvatar CoreX, which makes the Tiangai graphics processing unit (GPU), have also promoted upcoming upgrades to their offerings as a replacement for Nvidia's advanced A100 chip, two of people said.
Hygon brings its recently released GPU, Shensuan No. 2, on the market, which was designed from the start to be compatible with Nvidia's chip computing platform CUDA, meaning Nvidia users can switch chips with minimal design changes, a third person said.
Last month, startup Intellifusion announced the Deepedge10 chip to compete with the upcoming H20 chip that Nvidia designed to comply with the latest export restrictions.
A fourth person said Intellifusion pushed its announcement to take advantage of Nvidia's situation by licensing the chip to customers even though it has yet to go into mass production.
Enflame, Iluvatar CoreX, Hygon Information and Intellifusion did not immediately respond to requests for comment. Nvidia declined to comment.
PRODUCTION CAPACITY
Technology companies in the world's second-largest economy have begun to focus on Nvidia alternatives. Tencent itself has said that the US government has started looking for domestic sources of AI training chips. And internet search giant Baidu placed a sizeable order for Huawei chips, sources said last month.
“Our competitors in China, that's a lot of startups… There are about 50 startups focusing on AI. Huawei is a formidable competitor. Other US competitors, Intel and AMD are all very tough competitors. We have no shortage to competition,” Nvidia CEO Jensen Huang told reporters in Singapore last week.
Should Chinese chip designers win orders, they could still compete for manufacturing capacity given the restrictions imposed on U.S. foundries like TSMC to work with Chinese companies, said Lucy Chen, vice president of Isaiah Research.
“The majority of China's advanced manufacturing processes and advanced packaging capabilities are likely to be prioritized for use by Huawei. These emerging companies must devise a strategy to overcome the constraints of US restrictions and production cuts,” she said.
Still, the restrictions have created a market opportunity as tech giants pursue a strategy to stock more types of AI chips instead of just Nvidia's, making sustainability of their AI strategy a priority rather than performance, says Nori Chiou, investment director at White. Oak capital.
“The original goal (of the United States) was to slow down China's AI capabilities, but in fact the related action has increased China's self-development ability,” he said.
“Many Chinese cloud giants are working to build their AI ecosystems without American chips because of these limitations.”
(Reporting by Yelin Mo in Beijing, Fanny Potkin in Singapore and Brenda Goh in Shanghai; Editing by Tom Hogue and Christopher Cushing)