Financing in half a year exceeds 20 billion! AI chip "upstart" crazily sucks money, who can shake Nvidia?

Is the investment boom in artificial intelligence over? At least in the field of AI chips, it doesn't seem to be.

According to the incomplete statistics of Science and Technology Innovation Board Daily, in the first half of 2021, domestic and foreign AI chip companies raised 27 financing cases, with a total financing amount of more than 20 billion.

At present, AI chips have been widely used in cloud computing, autopilot, intelligent security, smart phones and other fields. According to the forecast of Sadie consultant, a research institution, the size of China's AI chip market will grow to 30.57 billion yuan in 2021, an increase of 57.8% over the same period last year.

From the perspective of the market pattern, in addition to established chip giants such as Nvidia, Intel and AMD, as well as software and Internet companies such as Google, Microsoft, Facebook, Amazon, Baidu, Ali, Tencent and so on, many startups have also stepped in to develop their own AI chips, attracting a large influx of financing funds.

Nvidia's dominant Baidu and Huawei competed to join the game.

According to the deployment location, AI chips can be divided into cloud-end and end-to-side. The cloud is mainly deployed in the data center, while the end-side is dominated by a variety of terminal devices and edge devices close to the end-side.

According to the different workload, AI chips can also be divided into two categories: training-oriented and reasoning-oriented. Because of the huge amount of computation and high performance requirements, the "training" part of AI is mainly carried out in the cloud, while the "reasoning" part can be carried out in the cloud and end-to-side.

At present, Nvidia GPU chips are almost dominant in the field of cloud AI training. According to the 2020-2021 Evaluation report on the Development of artificial Intelligence Computing Power in China released by IDC, domestic GPU servers accounted for about 95% of the market share in 2020, making it the first choice for artificial intelligence acceleration programs in data centers. It is reported that Amazon AWS, Google GCE, IBM Softlayer, Aliyun, Tencent Cloud and other cloud computing manufacturers all use GPU for deep learning algorithm training in data centers.

The dominance of AI training chips has brought great growth to Nvidia. Revenue from Nvidia's data center business hit a record high of $2.048 billion in the first quarter of 2022, up 79.5% from a year earlier, according to the results. It doubled in just five quarters, compared with $968 million in the fourth quarter of fiscal 2020.

Such a potential market space has naturally attracted many giants to participate in the competition. For example, Intel, the global "big brother" in the chip field, has turned on the "buy" mode.

In 2017, Intel bought Mobileye, an Israeli self-driving company, for $15.3 billion. According to data released by Mobileye, its self-driving chip system already accounts for about 70 per cent of the production market. In 2019, Intel bought Habana Labs, another Israeli AI chip company, for $2 billion. It is understood that the Gaudi chip released by, Habana Labs directly against Standard Nvidia GPU in the field of cloud training, and successfully signed Amazon AWS, at the end of 2020 can be said to have torn a hole in the field of AI training dominated by Nvidia.

Major domestic Internet and ICT companies are also active in developing AI chips. Alibaba launched the AI reasoning chip Hanguang, Huawei launched the Penton AI chip, and Baidu's AI chip was spun off this year into a new company, Kunlun Core (Beijing) Technology Co., Ltd., and completed independent financing at a valuation of $2 billion.

In the giant scuffle, AI chip track also emerged a large number of AI start-up "upstart" companies, they are the current AI chip financing main force. During WAIC 2021 in July this year, a number of AI chip startups, such as Zhixin, Cambrian, Flint Technology, Hanbo Semiconductor, and so on, released their latest products one after another. Among them, Sakahara Technology and Hanbo Semiconductor have all announced that they will be in mass production by the end of the year.

The success or failure of domestic GPU will be known in the next three or five years.

Generally speaking, the founding teams of these new companies come from well-known international chip manufacturers such as Nvidia and AMD. For example, the founder of GPU chip design company Moore Thread is Zhang Jianzhong, former global vice president and general manager of Nvidia in China, while the founders of Hanbo Semiconductor, Flint Technology, Muxi Integrated Circuit and other companies have all worked for AMD for many years.

These startups also have the support of big Internet companies behind them. For example, long Yuan Technology's largest shareholder is Tencent, with a stake of 20.47%, while Hanbo Semiconductor has a close relationship with Kuaishou, not only its investors but also one of its customers.

At present, a number of players, such as Moore Thread, days Intelligence Core, Wall Technology, Muxi Integrated Circuit and so on, are aiming at the GPU track to compete head-on with Nvidia.

Gu Bo (a pseudonym), a senior expert in the chip industry, told Science and Technology Innovation Board Daily: "it costs a lot of money to make chips." The basic development cost of a large GPU chip is about 600 million RMB, which involves the cost of a large number of engineers. For startups, a lot of early financing is necessary, otherwise it will be difficult to get things done. On the contrary, the general financing diluted 10% of the equity, 60-7 billion valuation is a more reasonable value range. "

On the whole, domestic GPU enterprises are still in the initial stage of development, and most of the products have not yet fully entered the market. And Nvidia's mature and powerful ecological GPU, is a big challenge for them.

Gu Bo believes: "it is certainly impossible to shake the mountains easily. We can only gather smart engineers to cut in according to the most urgent needs in the country, and start from scratch and do it bit by bit." At present, there seems to be an opportunity. The world looks at the Asia-Pacific region and the Asia-Pacific region looks at China. Our demand is the most active and exuberant. And the gap of domestic chip is very big, the opportunity of this piece is all the time, see who can catch.

Zhang Sishen, managing director of China Science and Technology Chuangxing, pointed out that it is very important to be patient. "GPU is a high-threshold industry, research and development is difficult, need to be patient, enterprises need to gradually cultivate their own application ecology. We can see more clearly in the next three to five years, or we can see the result, and we look forward to eventually nurturing 1 or 2 enterprises, not relying on financing, but really relying on products to get returns from the market. "

Can AI special chip break through?

In the face of Nvidia's strong GPU ecology, many startups choose to find another way to break through from AI chips. Wu Tong, senior market manager of AI technology in Anmou China, has said that ASIC customized AI chips will become the mainstream in the market in the next five to ten years. At the same time, (DSA), a special architecture in the field of AI chips, will also become popular.

Some investors believe that startups should choose a different path if they want to compete with Nvidia. "GPU is a universal chip. However, on the reasoning side, it does not need so much versatility, and the application scenarios are more diversified. Domestic manufacturers can start with specially customized reasoning chips to reflect the advantage of differentiation. "

It is reported that although GPU has strong computing power, it also has the disadvantages of high power consumption and high cost. Compared with the training chip, the reasoning chip pays more attention to the comprehensive indicators, including energy consumption, delay, cost and so on. John Hennessy, winner of the Turing Award in 2017, pointed out that DSA, is a processor architecture optimized for application domain, which is different from general architecture, and its advantage is that it can optimize architecture for a specific class of applications to achieve better energy efficiency.

Qian Jun, founder of Hanbo Semiconductor, told Science and Technology Innovation Board Daily, "GPU is not the best architectural solution on the reasoning side, and it is easier to break through ecologically on the reasoning side than on the training side. Foreign startups, including Habana, basically adopt the DSA framework instead of GPU. At present, Hanbo's scheme can be achieved at a lower cost than reasoning GPU. "

However, some industry insiders told Science and Technology Innovation Board Daily that startups abandoned the GPU structure more out of patent considerations. "under the same framework, it is difficult to evade the technology patent of GPU. In the aspect of AI special chip, the gap at home and abroad is very small, and there is no patent barrier. Compared with GPU, the success rate is much higher. "

Science and Technology Innovation Board AI chip first Cambrian, also chose to customize the ASIC chip to enter the board, its product line includes training and reasoning chips. AI chip experts who cooperated with the Cambrian told Science and Technology Innovation Board Daily that customization and localization are the advantages of this kind of domestic chip manufacturers. "domestic manufacturers are more approachable, respond faster and iterate faster, and have a stronger ability to cooperate with customers."

Although the prospect is considerable, there is no doubt that domestic AI chips still have a long way to go. Karl Freund, former general manager of AMD supercomputing department, once told the media: "in the past five years, it has actually been an explosion of PPT core building, and startups are peddling business plans to investors." These companies usually need to go through two generations of chips before they can really start to make a profit or gain market share. "

An investor told Science and Technology Innovation Board Daily: "the AI chip looks beautiful, but it is very difficult to do." At present, there is still a lack of stable and extensive landing scenes. Investment institutions should also keep calm and avoid rushing into hot spots. "