The Ministry of Science and Technology document lays out 13 “transformative” technology projects where it wants to put government money in coming months, hoping for delivery by 2021. One is to invent new chips to run artificial neural networks, the form of software propelling the AI ambitions of Google and other tech companies.
One criterion for the project refers specifically to Nvidia: the ministry says it wants a chip that delivers performance and energy efficiency 20 times better than that of Nvidia’s M40 chip, branded as an “accelerator” for neural networks. Now two years old, the M40 is not Nvidia’s latest and greatest chip, but is still used in AI projects.
The Chinese government has targeted Nvidia before. An October call for research proposals from the National Development and Reform Commission included another request for high-powered AI chips. In August, an investment fund owned by China's State Development & Investment Corp. led a $100 million funding round in Cambricon, a Beijing AI chip startup. Cambricon announced two server chips early this month that might substitute for Nvidia chips in some AI projects if they live up to their billing.
Cambricon is part of a boom of Chinese companies and startups working on AI chips—mirroring one in the US that has seen startups and even Google look to challenge Nvidia. In October, Beijing’s Horizon Robotics, founded by veterans of search company Baidu, raised $100 million, and Deephi hauled in $40 million. Established gadget maker Huawei is collaborating with Cambricon on AI chips for phones and other devices.
Chinese officials and tech companies each have good reasons to target Nvidia, which has built a large, lucrative market supplying hardware to AI projects. The company’s stock-market value grew 10-fold in the past three years as more companies invested in AI. It has begun offering chips for robots, drones, and autonomous vehicles, and signed up partners like Toyota and Volvo.
On the government side, Chinese officials want a domestic supplier in part because of concerns about relying on foreign chips for military and other applications, says Elsa Kania, an adjunct fellow at think tank the Center for a New American Security.
The Chinese government faces many challenges in making its AI and hardware dreams come true. China produces more computer-science graduates and machine-learning research papers than the US. But the country still lags in the high-level expertise needed for advanced AI projects, says Kania.
China has struggled for years to make its chip industry more competitive with those from the US, Korea, and Japan. Attempts to foster alternatives to Intel and other US processors helped birth chips used in some of the country’s world-beating supercomputers, but not widely used alternatives for servers and PCs. Alibaba, China’s biggest cloud provider, relies on Intel and Nvidia chips, for example. “They may have aspirations, but when it comes to designing chips and building fabs to make them at scale the Chinese are still multiple generations behind,” says Paul Triolo, who tracks Chinese technology and related policy at Eurasia Group.
China’s chip initiatives have been hampered by wary US officials scrutinizing acquisitions of US semiconductor technology, who may soon turn warier. President Obama blocked a Chinese fund from acquiring a US chip firm in December 2016; President Trump quashed a similar deal in September. This month, a bipartisan group of lawmakers introduced bills to sharpen the teeth of the committee that advised on those decisions, in part motivated by China’s ambitions in chips and AI.
For now, few of China’s AI chip companies are directly attacking Nvidia’s core market selling chips for servers—despite the government’s apparent interest. Startups Horizon Robotics and Deephi, and the much larger Huawei, are instead focused on chips to bring AI functions such as understanding video to devices such as cars and cameras.
The healthy market for surveillance in China is one driver of that trend, says Chris Rowen, an investor who previously led Silicon Valley chip-design companies. Putting AI chips into cameras can make them capable of automatically spotting people, objects, or actions. Google recently touted the iage-recognition ability of its Clips camera that boasts an AI chip, although it uses it that capability to take candid family snapshots, not for surveillance.
Rowen says China’s chip startups are also eyeing the vast market potential of putting AI functions into home appliances, car parts, and other Chinese-made gadgets. “The way to make this technology proliferate is to knock down the costs,” Rowen says.
Despite the emerging trans-Pacific competition in AI, there are not clear battle lines between Chinese and American companies. State-backed Cambricon is licensing designs from Silicon Valley chip designer Arteris for the “backbone” of interconnects that move data around a chip, for example. Intel led the $100 million funding round into Horizon Robotics, which is working on AI chips for autonomous vehicles, even though it has products of its own in that market. And in September, Nvidia CEO Jensen Huang announced new deals with Chinese internet giants Alibaba, Baidu, and Tencent. “China’s AI applications are going to be running on US hardware for a while, that’s the reality at this point,” says Triolo.
