Written by |Tian Xiaomeng.
Edit |Lee Shin Ma.
Title Picture | ic photo
Masayoshi Son and AI are on the bar.
The front foot cooperates, the back foot competes. According to the "Science and Technology Innovation Board**" on February 15**, SoftBank and NVIDIA are forming an AI industry alliance to improve wireless services through artificial intelligence technology. Two days later, it was reported that Masayoshi Son, founder and CEO of SoftBank Group, wanted to raise $100 billion for AI chip companies to compete with Nvidia.
If Son's fundraising project is successful, it will be one of the largest investments in artificial intelligence since the advent of ChatGPT compared to Microsoft's previous investment in OpenAI, which is one of the reasons why SoftBank's move has attracted much attention.
It is reported that the project, codenamed "Izanagi (Japan's god of creation and life - Izanagi)", is directly led by Son, and is Son's next major attempt at a time when SoftBank has slashed investment in start-ups. One option would be $30 billion from SoftBank, with the remaining $70 billion likely coming from Middle Eastern institutions.
1. Is ARM a new turnaround?
AGI is something that every AI expert is looking for. Son once said.
In fact, Son has never wavered on the path of artificial intelligence. It can be seen from the financial report data for the third fiscal quarter of 2023 that SoftBank Vision** mainly includes SoftBank Vision ** Phase I (SVF1), SoftBank Vision ** Phase II (SVF2), and Latam Funds**. Among them, SoftBank's vision, which has attracted much attention, was the first to announce that the focus of investment was in the field of AI.
SoftBank Vision Phase II** said during the fundraising phase that it will mainly invest in AI technology, aiming to accelerate the AI revolution by investing in market-leading, technology-driven growth companies.
In terms of revenue, since its establishment, SVF1 has invested a total of US$89.6 billion, with a cumulative return of US$106.3 billion and a profit of US$16.7 billion; SVF2 has a total investment of $52.3 billion, a cumulative return of $33.3 billion, and a loss of $19 billion.
SoftBank's earnings report.
The quarter came after four consecutive quarters of losses, and SoftBank turned a profit. Yoshimitsu Goto, SoftBank's chief financial officer, attributed the earnings to "portfolios that are changing dynamically." At the same time, Goto said that SoftBank has undergone a "huge shift from Alibaba-centric to AI-centric."
SoftBank's earnings report.
It is worth mentioning that as of December 31, 2023, ARM's equity value accounted for 32% of SoftBank Group's net assets, which is already the largest single asset in SoftBank Group's portfolio.
But ARM is also full of twists and turns at SoftBank.
In 2016, SoftBank acquired Arm for about $31.4 billion** and pushed it to go private and delist. But after the failure of the WeWork IPO plan, in order to ease cash flow, SoftBank almost liquidated Alibaba's **, and similarly, Arm was also raised.
In 2020, SoftBank reached an agreement with Nvidia to plan to **ARM at a cost of $40 billion, which caused the chip industry "**, but due to strong opposition from regulators in the United States, the United Kingdom, and the European Union, in February 2022, this ** plan died."
In November 2022, Son announced that he would hand over the day-to-day work of the group to SoftBank executives such as Yoshimitsu Goto, and that he would focus on chip company ARM in the future. On September 14, 2023, ARM went public on the NASDAQ, becoming the largest IPO in the global technology sector in 2023.
On February 7 this year, ARM released strong fiscal third quarter data, which soared 42% after hours, pushing its market value to nearly $100 billion. Clearly, SoftBank, as the controlling shareholder of ARM, has received excess returns in its listing.
It is reported that the artificial intelligence chip company that Son hopes to create is to complement SoftBank's Arm.
However, whether ARM can help SoftBank stand firm in the AI field still needs to be recognized by the market.
From the perspective of ARM's business, its main business is to authorize the design and drive the construction of chips for smartphones and PCs, and a small part of its revenue comes from supercomputer chips for AI data processing, but it has not formed a monopoly position in cloud computing processors, automotive chips, and AI layout and attempts.
2. Who will die in the AI chip war?
With the sweep of ChatGPT and SORA, the competition for AI chips will also be further opened in 2024.
At the beginning of February this year, Meta Platforms confirmed that the company plans to deploy the latest self-developed custom chips in its data centers this year, and will coordinate with other GPU chips to support the development of its AI large model.
Sam Altman, the father of ChatGPT, is also trying to reshape the global semiconductor industry. Altman once posted: "OpenAI believes that the world needs more AI infrastructure than people currently plan – including fab capacity, energy, data centers, etc. Building a large-scale AI infrastructure and resilient ** chain is essential for economic competitiveness. ”
Altman is reportedly planning to seek $5 trillion-$7 trillion in funding for AI chip projects to build dozens of chip manufacturing plants over the next few years, which will be operated by multiple foundries that will produce chips not only for OpenAI, but also for other companies.
In addition, at SoftBank's earnings conference, CEO Junichi Miyagawa revealed that SoftBank's self-developed generative AI computing platform and large language model are currently making good progress.
However, while many companies are mixed, it cannot be ignored that the market value is approaching 1Nvidia, the "hegemon" of the $8 trillion AI chip market, almost monopolizes the AI chip market with a market share of 80%.
According to public information, for the product planning of the artificial intelligence market, Nvidia will supply H200 in the second quarter of this year, put GH200 GH200NVL into production, and B100, B40, GB200, and GB200NVL will also be launched this year, and the development is unstoppable. In addition, in the AI chip market, there are also technology giants such as Microsoft, Google, AMD, and Intel catching up one after another.
However, in the face of the AI chip market pattern, Sun Yongjie, a senior analyst in the communications industry, admitted to Donews that artificial intelligence is divided into two stages: training and inference, and is currently mainly focused on the training part, such as large models, and even the hottest SORA.
As far as the 'AI chip' mentioned on the market is concerned, the concept is a bit confusing. Sun Yongjie said. AI chips can be divided into general-purpose and customized. For example, Microsoft, Intel, AMD, etc. take the customized route, and the chips they launch are for some special artificial intelligence application scenarios, and the efficiency will be quite high. And NVIDIA's chips are the highest in terms of computing power and versatility.
Therefore, I think that at least at the present or for a considerable period of time in the future, Nvidia is far ahead in the training-based stage. Once it reaches the inference stage, Intel will 'get up' because the inference ability and inference efficiency of the CPU are stronger than that of the GPU, which is determined by the CPU architecture and CPU characteristics. ”
While large manufacturers are "playing AI", they have also driven the AI financing boom. According to the information, up to now, there are more than 18 chip design startups for AI large model training and inference around the world, including Cerebras, Graphcore, Bichen Technology, Moore Threads, D-Matrix, etc., with a total financing of more than $6 billion, and the overall valuation of the company has exceeded $25 billion (about 1792.).9.5 billion yuan).
In Sun Yongjie's view, the chip design market is huge, and if you go for customization, there may be a market, and there is no bubble, and you still need to squeeze out the bubble if you take the general-purpose chip route.