Masayoshi Son’s Quest to Become the Next Nvidia
The billionaire SoftBank founder promises to invest $100 billion in the US as part of a sweeping redemption plan.
By Min Jeong Lee and Ian King
December 22, 2024 at 5:00 PM EST
When Masayoshi Son made a surprise appearance at Mar-a-Lago alongside Donald Trump last week, the founder of Japan’s SoftBank Group Corp. vowed to invest a stunning $100 billion in the US over the next four years. The president-elect joked that perhaps SoftBank would like to double that target to $200 billion. Not to be outdone, Son replied with a wide smile, “I will really try.”
That may not be an exaggeration. The Japanese entrepreneur is quietly plotting one of his trademark bet-the-company moves that’s likely to cost as much as his initial pledge — and perhaps a good bit more. Over the past few months, according to people directly involved, Son has developed a singular obsession: how to build the next Nvidia Corp. with his own chip and rake in some of the tens of billions of dollars being spent on artificial intelligence hardware.
It's a strategy that could lead to wide-ranging investment in chip production, energy capacity and related technologies. Son's aim is to have the first batch of shipment-ready AI chips by 2026, with a prototype available as early as the summer of next year, according to people familiar with the details.
The 67-year-old sees himself as a trailblazer in AI, and yet he’s watched Nvidia emerge as the most powerful player in the field. The US chip company holds a near-monopoly on the semiconductors used to train AI models like OpenAI’s ChatGPT, lifting its market valuation to more than $3 trillion. Now, with a sprawling portfolio of startups and a 90% stake in Arm Holdings Plc — whose chip designs won a near-monopoly for smartphones because of their energy efficiency — Son is ready to become a serious contender.
Conscious of the damage to his legacy from a series of disastrous investments, Son is determined to develop SoftBank’s own chips to help make AI available everywhere, said the people, asking not to be identified discussing confidential strategies. Eager to address AI’s chronic thirst for electricity, he anticipates Arm will stand at the center of these efforts with Taiwan Semiconductor Manufacturing Co. as a likely manufacturing partner. He’s also seeking knowhow from SoftBank’s recently acquired Bristol-based Graphcore Ltd., which designs larger processing units to speed up AI software, they said.
Plans remain in flux, including financing for SoftBank, which already has trillions of yen in debt. One of the many wild cards is a possible role for Intel Corp. Earlier this year, Arm approached Intel about the possibility of acquiring part of the Santa Clara-based chipmaker and was rebuffed. Since then, Intel CEO Pat Gelsinger, who had insisted that the company should remain whole, has been ousted by the board. SoftBank has already reached out to potential clients, however, with at least one company — a cloud AI provider — expressing interest in using Son’s chip in data centers. Ultimately, the grand plan is to create an ecosystem that reaches homes and offices and people on the road, taking AI far beyond the scope of software training in data centers, where Nvidia is focused.
Son has only hinted at these ambitions publicly, and representatives of SoftBank and Arm declined to comment. SoftBank Chief Financial Officer Yoshimitsu Goto declined to go into details during an interview at SoftBank headquarters, but he made clear that his boss no longer has financial constraints on his ambitions.
“We are ready to go in with as much money as Son needs. We can always find a way to do any project, even while standing by our financial policy,” Goto said. “It wouldn’t be wise to put a limit on something from the start.”
SoftBank is in need of a makeover. The last time Son made an investment pledge to Trump in 2016, Son’s Vision Fund pumped well over $100 billion into hundreds of startups worldwide, feeding one of the biggest venture capital bubbles that resulted in the implosions of the likes of co-working company WeWork, construction company Katerra, smart window maker View, direct-to-consumer goods company Brandless and robot-made-pizza delivery company Zume. While the two Vision Funds have recovered from the worst of their losses, they’ve hurt Son’s rags-to-riches legacy that spans the launch of iPhones in Japan, a pioneering rollout of broadband in the country and an early bet in Alibaba Group Holding Ltd. that yielded thousands-fold returns.
Enter Arm — now the single most valuable part of SoftBank’s portfolio. Buoyed by Nvidia’s use of its architecture in AI accelerators, the chip designer’s value reached a record $195 billion value this year, more than six times the $32 billion price SoftBank paid in 2016. And Son is working with Arm CEO Rene Haas to expand into AI accelerators, Nvidia’s playground.
Here, Nvidia’s graphics processing units are far ahead of the competition in terms of capabilities and ease of use. Even Advanced Micro Devices Inc., the strongest contender so far, has had little success closing the revenue and market share gap with Nvidia. But the sheer volume of computing power and money AI needs is creating openings for competing chip designs, spurring tech powerhouses like Amazon.com Inc. and Alphabet Inc., along with dozens of startups, in a race to fulfill demand that’s impossible for Nvidia to satisfy by itself.
“There is scope for new entrants precisely because the market’s going to grow so much and change so much,” said Chris Miller, author of Chip War: The Fight for the World’s Most Critical Technology. “It’s not surprising that SoftBank, given its ambitions and its scale, is also looking at this market.”
Son is betting that he has an edge. Arm’s energy-efficient architecture — critical for helping to pack more capabilities into mobile devices — now dictates the designs of central processing units that work alongside Nvidia’s chips in training AI. It’s a largely backstage role. But what if Arm could take center stage with its own accelerator?
At stake is a multi-trillion-dollar question of which company AI will anoint its next semiconductor champion, following Nvidia, TSMC and Broadcom Inc.
“Having Arm within SoftBank gives us a huge upper hand in the AI race. It means Son and SoftBank are in a position to take the lead,” Goto said. “My job is to see Son’s strategy through while keeping the promises we made to our stakeholders. I have no doubt that we can pull this off.”
When SoftBank bought Arm in 2016, Son called the Cambridge-based company his crystal ball into the future of tech. But once Son won $60 billion in capital commitments from the Saudi and Abu Dhabi wealth funds, insights from Arm became less critical: His Vision Fund could now channel unheard-of sums into startups to boost their growth, and he could replicate with brute scale his mega-successes with Yahoo and Alibaba across dozens of companies, SoftBank executives said at the time.
By 2022, it was clear that deploying so much capital in a small window of time had led to flawed picks and insufficient due diligence, with the minimum $100 million ticket size magnifying miscalculations. The result of the billionaire’s exuberance and optimism was a $20 billion loss in the Vision Fund segment for the year to March 2022.
Son stopped attending the Vision Fund’s high-stakes global pitch meetings, a key step in earmarking hundreds of millions of dollars. SoftBank hit the brakes on investments and fired hundreds of Vision Fund staff. In the fall of 2022, Son took full responsibility for the bloodbath and ceded day-to-day operations to his lieutenants. He also disappeared from earnings calls.
Son’s stated purpose was to focus on bringing Arm’s stock back to the market. But he also needed time for introspection. “I couldn’t stop crying for days,” he told shareholders in June last year. He realized he didn’t want to end up as just a business operator. “I want to be an architect of the future,” he said. “I may only be able to play a small part, and I may not have the ability to contribute much. But this is what I want to do, even if I die while drawing up the plans.”
Behind the scenes, Son was working with Haas to transform Arm. For most of its history, the company focused on licensing basic building blocks, charging pennies on the dollar per device that used its technology. But under Haas, and with Son’s support, Arm began to move up the value chain in red-hot arenas, becoming more of a chipmaker trying to charge more for complete chip designs.
At the same time, Son was thinking hard about what role SoftBank could play in AI. He often turned to a secret ten-member iChat group of trusted lieutenants to float ideas. Here, an excited Son might send more than a dozen messages in a day, some with iPad drawings or flow charts, others with free-form riffs on the future of technology or, in one case, a photo of a four-wheeled, red robot on which he’d drawn fingers.
Then in June of this year, Son had a eureka moment at 4 o’clock in the morning. He texted the chat group in his excitement. During an annual shareholders’ meeting that same day, he compared his elation to solving a complex linear equation after working on it night and day for a year. “I rushed to make a whole bunch of calls to the US to make sure my thought process was correct,” he said without elaborating. “Watch me. We are going to do this.”
Later, he urged his iChat brain trust to revise any five-year business plan every week. They would need to be flexible and ready for anything.
For Son, the new chip project would bring him back full circle to a photo of a microprocessor he held in his hands when he was a student. The glimpse the photo gave him of a technology-fueled future brought him to tears, he often said. The difference is that now, Son has access to $25 billion in cash and the ability to raise billions more from banks and other investors to bring about that future himself. In one scenario, code-named Izanagi after the Japanese god of creation and life and containing the initials for artificial general intelligence, SoftBank would seek as much as $100 billion to bankroll the chip project, Bloomberg News reported earlier this year. He might also tap hyperscalers in a project financing scheme and leverage tens of billions of dollars into hundreds of billions of dollars, one person familiar with Son’s thinking said.
Son’s ambitions to build his own chip empire got an added boost from conversations with OpenAI CEO Sam Altman. The two had kept in touch after their first meeting in 2019, when Son offered OpenAI $1 billion in investment, which never materialized. Son, who turns to ChatGPT constantly to ask questions including “How do I make a million dollars?” was one of the first people in Japan to understand that a paradigm shift had occurred. Often, the two talked about the shortage of semiconductors needed to power large language models and other AI applications.
Each Nvidia chip costs tens of thousands of dollars. In response to the high prices and long wait times for the graphics processing units, Son and Altman tossed around enormous figures that might be needed to build a rival AI chip. Son suggested that they would need $3 trillion to do so successfully, one person familiar with the matter said. Altman then countered with a $7 trillion figure, a sum that was later reported by the Wall Street Journal.
For technical expertise, however, Son turned to Arm. Since its founding in 1990 by a small group of engineers, Arm has grown to become the owner of the world's most widely used microprocessor technology. In that instance being a latecomer to the field had been an advantage, helping Arm leapfrog years of development to come up with a clean and efficient design. The idea is that Arm’s engineers could repeat what they did 35 years ago, guided by Son’s vision of what an AI-tailored chip needs.
Son spoke with Haas regularly, with talks sometimes lasting hours. Discussions often returned to the Izanagi venture, a push that would complement Arm’s IP assets and build an AI chip powerhouse within SoftBank’s AI ecosystem. The talks ran alongside conversations around other projects including a model similar to Blackrock Inc.’s push to create a $30 billion fund with Microsoft Corp. to build AI data warehouses and energy infrastructure, one person said.
Haas’ ambitions to transform Arm into a different, bigger chip company aligned well with Son’s. Still, the two CEOs weren’t always in agreement. Haas, who formerly worked on Nvidia’s computing products and helmed Arm’s IP product division, is acutely aware of the challenges of chip design and the sector’s manufacturing hurdles. After getting repeated push-backs from Haas, a frustrated Son at one point demanded to speak with other Arm executives, a request that Haas refused. The two continue to talk daily, people familiar with the two CEOs’ relationship said, and Haas, who is on board with Son's ambitions, has become his go-to for the practical implementation of the grand plans.
SoftBank is betting that AI chip designs – a field that has historically been unpredictable and competitive – will change dramatically over the next decade or two. One open question is whether it’s more or less efficient for a single GPU to conduct both training and inference, as in Nvidia’s case. If it were more energy-efficient and faster for the two functions to occur separately, this might provide an opening for a new player. Son believes Arm is well-positioned to support inference in connected devices from smartphones to robot control systems at the edge of the network, according to a person familiar with the matter.
“Because AI is such a nascent industry, we’re in the early stages,” Miller said. Today, most AI accelerators are used for training, but in the future, a much larger share of computing power will likely go to putting that training to work to make predictions or infer things about the data, he said. “Nvidia is obviously the first leader in the market for AI accelerators, but that market certainly will grow and could change a lot over the coming decade.”
Others are also angling for the same projected demand, however. AMD, which argues its chips are better than Nvidia’s in performing some workloads, plans an update of its accelerators next year that it says will bring about a big jump in inference speeds. Amazon, which invented cloud computing operations with custom-built hardware, is also using its scale to muscle its way into creating an Nvidia fighter. The e-commerce company’s engineers are working to deliver its newest Trainium AI accelerator to data centers by the end of the year, and its sprawling infrastructure may help jumpstart momentum.
Old hands in the chip industry have largely watched SoftBank’s efforts with bemusement. Son’s skills as an operator served him well in software sales, magazine publishing, mobile network rollouts and supporting young entrepreneurs, but he’s never been tested in chipmaking’s stark and capital-intensive realm of nanometers and clean rooms. Even Nvidia almost went under multiple times before gaining the clout to get TSMC to develop manufacturing processes for its latest chips.
In the end, SoftBank’s chances of success largely rest on Arm’s intellectual property. Other weapons in Son’s arsenal include SoftBank’s client base — although only a fraction of Amazon’s — in a Japan that’s eager to pay top dollar to catch up in AI. The tech group also operates renewable energy projects in the US that help power Google and other hyperscalers’ data centers. Son also has a team of investment experts at a smaller and leaner Vision Fund hunting for companies and technologies that he needs, including the Graphcore deal.
“Our vision, which I think is grounded in reality, actually aligns with some of the grand vision that Son has promoted as well,” Nigel Toon, the CEO of Graphcore, said at the time of SoftBank’s acquisition in July. “We are part of the delivery behind a very grand vision.”
Many details in SoftBank’s chip plans remain under discussion including how many companies and investors will be involved. While TSMC is the preferred manufacturing partner, SoftBank may seek other partners to secure capacity and technological support. Son is also paving the way for a strategic partnership with OpenAI. The Vision Fund segment, which swung to a profit in the September quarter, invested $500 million in the startup and launched a $1.5 billion tender to buy more stock from OpenAI employees through this month. SoftBank is committing additional capital to the second Vision Fund next year, boosting the fund's size from $60.8 billion at end-September, according to people familiar with the matter.
In his texts, Son reminds his confidantes to think on a scale of centuries. As immense as the chip project is, it’s only one of many ideas he’s pursuing and could end up comprising 50% of SoftBank’s new direction or just 0.5%, one person familiar with Son’s thinking said. SoftBank’s recent announcements have spanned a supercomputer using Nvidia’s upcoming Blackwell chips, faster AI processing in cars and a number of Vision Fund deals around robotics companies. Robots would be pivotal to delivering real-life benefits of AI to people, argues Son, who at one point brought a mock robot design to Jony Ive’s home to propose a possible venture.
During SoftBank’s shareholder meeting in June, one attendee asked about Bloomberg’s report on the Izanagi chip and how much progress had been made. Son smiled. “Let’s not talk about specifics,” he said.
“We don’t want to show our hand too soon. This is a world of pros, and our success hinges on the results we achieve. We can’t comment on our methodologies ahead of time,” he said. “But I can tell you I’m committed to delivering results.”
— With assistance from Mark Bergen, Shirin Ghaffary, and David Stringer
bloomberg.com |
