When a group of leading Chinese artificial intelligence (AI) firms flew all the way to Silicon Valley in March 2018 to attend Nvidia’s annual GPU Technology Conference, they were told by the organizer to play down their relations with the US technology giant as growing trade tensions between the United States and China had just erupted into a public war.

Santa Clara, California-based Nvidia is one of the world’s leading makers of graphics chips, known as GPUs, typically used for PC gaming. In recent years the company has also begun to focus on AI and deep learning.

(Picture: Shutterstock)

“We were asked to pull press releases that signaled our close cooperation with Nvidia, and we just did it as requested,” a representative from one of China’s leading AI companies who attended the event told the South China Morning Post, asking to remain anonymous because of the sensitivity of the matter.

“But we all felt a chill at that time, and it is my understanding that many Chinese AI companies began to prepare a plan B when they returned to China after the event,” said the person.

China and the US are vying for supremacy in the AI field, with a number of Chinese companies coming to prominence as they expand outside the domestic market.

However, China’s emerging AI champions took a body check last week after eight of them were added to a US trade blacklist, preventing them from buying American technology, for what Washington has described as human rights violations.

“Specifically, these entities have been implicated in human rights violations and abuses in the implementation of China’s campaign of repression, mass arbitrary detention and high-technology surveillance against Uygurs, Kazakhs and other members of Muslim minority groups in the XUAR [Xinjiang Uygur autonomous region],” the Department of Commerce’s Bureau of Industry and Security said in its ruling.

The companies added to the so-called Entity List include facial recognition start-ups SenseTime Group, Megvii Technology and Yitu Technology, video surveillance specialists Hangzhou Hikvision Digital Technology and Dahua Technology, AI champion iFlyTek, Xiamen Meiya Pico Information Co and Yixin Science and Technology Co.

“The business portfolio of these eight AI firms is mainly concerned with software and solutions – so the impact [in this sense] is limited,” said Charlie Dai, a Beijing-based principal analyst at Forrester Research. “However, it is not easy to source alternatives for the hardware that the AI software relies on, such as sensors, GPUs, FPGAs and chipset design software.”

A screen shows use of facial recognition technology at the World Artificial Intelligence Conference in Shanghai on August 29. (Picture: Bloomberg)

FPGA, short for field-programmable gate array, is an integrated circuit (IC) that can be programmed in the field after manufacture.

The US decision, announced shortly before the US and China were to meet for crucial trade talks in Washington, was roundly denounced by the Chinese tech firms. In their public statements last week, many of the companies played down the likely impact, but reaction since then has been more nuanced.

On Monday, the chief executive and co-founder of AI start-up Megvii said that the ban was a “challenge” and there would be an impact on its supply of servers — as well as to its planned initial public offering in Hong Kong.

“The specific impact is that we can't directly buy products subject to US export regulations, such as x86 servers and GPUs made in the country,” said Yin Qi, the company’s co-founder and chief executive, in an internal letter to staff on Monday, a copy of which was obtained by the Post.

Nevertheless, Yin said Megvii was “well equipped for the fight” and had begun to diversify its suppliers in May, when telecommunications equipment maker Huawei Technologies was first blacklisted by the US on national security grounds.

So-called x86 servers are a family of instruction set architectures, initially developed by Intel on the 8086 microprocessor. Although several domestic brands like Inspur and Sugon provide such servers, they largely rely on patents licensed by American semiconductor giants Intel and AMD.

In the case of GPUs, domestic suppliers have so far been unable to compete with the likes of AMD and Nvidia in the commercial market.

China is home to the biggest number of AI start-ups in the world valued at US$1 billion or above, according to a February report from CB Insights, a research firm that tracks venture capital activity. Six of the 11 so-called unicorns in the annual compilation of leading AI start-ups are from China, with SenseTime taking the top spot with a valuation of US$4.5 billion.

Although Chinese AI start-ups raised US$4.9 billion in 2017, edging out their US counterparts which had US$4.4 billion in funding, many analysts said that local firms are still heavily reliant on US-made AI servers which enable the development of algorithms and deep learning training.

With the US closing more doors to China, what are these Chinese tech firms to do?

According to a Jefferies research note last week, China’s top surveillance camera provider Hikvision Digital Technology, among those recently blacklisted, will be able to find alternatives for key US components such as cameras and storage devices from domestic and Japanese vendors.

However, Hikvision remains heavily reliant on Intel and Nvidia for the CPUs and GPUs that power AI servers.

“Nvidia’s GPU outperforms other solutions in deep learning training, thanks to its CUDA toolkit and libraries,” according to the Jefferies report from analysts Rex Wu and Lydia Lin. CUDA provides high performance support as well as built-in AI functions.

CPUs and GPUs are indispensable for deep learning, which needs massive data training processed on hardware, said Kuang Kaiming, a Shanghai-based AI engineer from Diannei Biotechnology.

“They (CPU and GPU) are like roads. No matter what kind of car you drive, it needs to be on the road,” said Kuang.

This lack of competitiveness in core areas of strategic technology is one reason Chinese President Xi Jinping has called for increased self-reliance – not just for business reasons, but also to bolster military and economic security.

The issue gained extra urgency after the US government brought ZTE Corp, a Chinese state-owned telecommunications equipment company, to the brink of collapse last year for breaching US sanctions. This was followed by the US blacklisting of Huawei in May.

With the recent additions, there are about 180 entries for Chinese companies in the current US Entity List, while Russian companies number around 310 to 320.

Just like Huawei, more Chinese technologies companies are realizing that in the current environment they need a diversified list of suppliers to counter unexpected events that could threaten their very existence.

Dahua Technology, another major Chinese maker of surveillance cameras, has about 10% of its components supplied by North America, the company’s president, Li Ke, said during a teleconference last week.

People visit the Huawei 5G Experience Booth during the preview of the 2019 World Artificial Intelligence Conference in east China's Shanghai in August. (Picture: Xinhua)

“We also have a replacement strategy for those ‘hard-to-replace’ components such as CPUs, GPUs, FPGA and analogue devices,” said Li, adding that the company’s inventory currently exceeds one year to ensure it can adapt to any shocks.

“If we have limited access to chips, we will replace them, and if we can’t, we will change components,” said Hikvision’s board secretary Huang Fanghong last week. “We will redesign our products. If it’s necessary, we will create our own chips.”

Whether this is doable is open to question. Although there are Chinese alternatives to Intel and Nvidia, those are not as well-established as the US firms.

“Tech leaders in China like Alibaba, Huawei and Baidu are all making strategic investments in AI hardware, some for training and some for inferencing,” according to Forrester’s Dai. “However, the feasibility of chipset replacement depends on many other factors, such as the tape-out success rate, server compatibility, software interoperability and workload applicability.”

Some Chinese companies are making good progress with AI chips, which can power a range of applications including “internet of things” devices and autonomous driving.

Alibaba Group Holding, China’s ecommerce giant and market leader in cloud computing, last year set up a chip subsidiary called T-head (or Pingtouge in Mandarin, a nickname for the honey badger) to make its own AI inference chips.

In July, it unveiled its first processor – the Xuantie 910 – that can be used in fields including 5G mobile networks, AI and autonomous driving. Alibaba owns the South China Morning Post.

In late August, Huawei unveiled a high-end AI chip Ascend 910 for servers, claiming that it is the “world’s most powerful AI processor” targeted at AI model training, packing twice the performance of rival Nvidia’s Tesla v100.

The Shenzhen-based company also this year revealed its new Atlas 900 AI computing system that packs thousands of Ascend processors. Huawei said it has set a new world record for AI training performance.

Huawei’s AI application-specific integrated circuit (ASIC) offering, together with other domestic AI chips from companies like Horizon Robotics and DeePhi Technology, not only provide cutting-edge AI solutions, but can help lower the cost of AI chips by about 30% to 40%, according to the Jefferies report.

Huawei is the logical fallback for Chinese companies given its depth of research in ASIC design and AI custom silicon, but a bigger challenge may be acquiring access to deep data lakes, said Will Townsend, a senior analyst with US-based Moor Insights & Strategy.

“AI and machine learning is only as good as the data used to train the models. Chinese companies may be thwarted given the cybersecurity concerns raised by the West,” said Townsend.

In August, a group of researchers from Tsinghua University in Beijing created the Tianjic chip, a hybrid processor that has the architecture of a brain-inspired neuromorphic chip but can run algorithms for deep learning at the same time. These chips are more energy efficient than normal ICs because they only “fire” when required, like how neurons in a brain “spike.”

The Washington-based think tank Center for Data Innovation seconds the view that China has begun to show signs it may be able to close some of the gaps with the US, at least in AI chips.

According to research published last month, the think tank found that China had only one company among the top 15 for semiconductor sales in the world in 2019, while the European Union had two and the US six.

But take the number of firms designing AI chips in 2019 and the gap narrows – the US has 55, followed by China with 26, while the EU is in third place with 12 firms.

Even in a worst-case scenario where Chinese companies like Hikvision are banned without any GPU access, they can still “train their algorithms using China's National Super Computing Centers,” the Jefferies report said.

SenseTime started talks with Chinese supercomputing centers years ago, according to people familiar with the matter, who declined to be named as the information is private. But not all companies have prepared for this, the people added.

Relying on the cloud computing power of third parties to train AI models could raise security concerns though, as AI companies would need to share data with the third party. This may contravene privacy agreements with users, said Xia Yin, a Silicon Valley-based AI scientist.

As such, it looks like China urgently needs to speed up domestic chip production capabilities to provide emerging AI champions with hardware alternatives – or they will face an uncertain short-term future.

“The only way to overcome the challenges [of being put on the US blacklist] is to improve the supply chain in China,” said Qin Hailin, director of the Industrial Economy Institute at the China Center for Information Industry Development.