Elon Musk is not Donald Trump’s only friend. More discreet, Masayoshi Son, the boss of the tech investment giant SoftBank, has long shown signs of confidence in the Republican’s policies. Already, in 2016, he celebrated the start of Donald Trump’s first term by pledging to invest $50 billion in the United States. The Japanese has set the bar even higher for this second term. In December, invited to the sumptuous Mar-a-Lago residence, he announced 100 billion dollars of investments in the United States. Donald Trump teased Masayoshi Son on this occasion by asking him to double down: “He can afford it, you know!”, he declared to the audience. “Masa” kicks in but compliments the chosen one for his negotiating skills. A well-oiled show.
SoftBank had one last trick in store. During the inauguration ceremony, he unveiled, with the database specialist Oracle and the AI leader OpenAI, a titanic data center construction project with a budget of $500 billion over five years and called Stargate. A real “Marshall plan for AI,” observes Kevin Polizzi, president of Unitel Group, specializing in the cloud. “And ideal for OpenAI, which will be able to precisely adjust the architecture of the centers to its needs.”
SoftBank is playing big here. In recent years, several of his bets have not had the expected success. Covid-19 has disrupted its Uber foal. The world has shunned the pizza-making robots of his start-up Zume. And the WeWork adventure in which SoftBank had invested $9 billion was a resounding failure. Failures which weighed on SoftBank’s accounts and damaged its image as King Midas with infallible flair.
Behind Stargate, the support of Donald Trump
His new project Stargate also presents challenges. The first ? Electricity supply. Already, the US grid is showing signs of weakness in the face of AI’s voracity – data centers are expected to consume 9% of the country’s electricity by 2030, double their current level. “The Americans need to increase their production quickly. They know that the development of nuclear power will take time, which is why they are accelerating in gas and oil,” observes Dali Kilani, co-founder and CTO of FlexAI .
Stargate project stakeholders will also need to find unoccupied areas near these sources of energy and internet interconnection. “There is hardly any suitable private land left, which is why the State will mobilize federal land,” notes this AI infrastructure specialist.
Finding the right chips to equip these future data centers will also be a major concern. SoftBank, which took a firm course towards AI well before the announcement of the Stargate project, has put a lot of emphasis on the semiconductor design specialist Arm, which it owns 90%. Rightly so: this group’s expertise in chips for smartphones and connected objects will be valuable in bringing these small devices into the era of generative AI. Arm does not, however, develop powerful processors like those from Ndivida that the entire sector is chasing. And TSMC, which is responsible for manufacturing these, is “at maximum production capacity, it will not be able to go any faster,” confides Dali Kilani.
Elon Musk makes fun of Stargate
Last challenge to face: cost management. Such a project requires a lot of money. Surprisingly – given his support for Donald Trump – Elon Musk was one of the first to point this out, mocking the project praised by the president which he supports but in which his old adversary, Sam Altman, participates alongside the investor Japanese. “SoftBank doesn’t have the money. I have it on good authority that it has not yet secured more than $10 billion,” the billionaire tweeted on January 22. “This is false, as you undoubtedly know. Would you like to come and visit our first site already under construction?”, quickly retorted the boss of OpenAI. This titanic project will nevertheless be complex to scale.
“How many of these billions of dollars will really be put to good use is difficult to say. Those who quickly jump on the AI train do not always have an optimal business model and are likely to buy equipment that will be outdated in some time given the speed at which this is evolving, But countries that do nothing run a greater risk: they create a structural delay,” warns Kevin Polizzi of Unitel Group. Even if it presents certain risks, SoftBank’s bet is smart. Because in recent months, a discovery has upset all projections in the world of AI.
A discovery shakes up the world of AI
With recent families of AI capable of “reasoning” such as OpenAI’s o1 model, “we realize that computing power matters even more than we thought a year ago”, explains Karim Beguir , the founder of the start-up InstaDeep. It is not only strategic during the training phase of a model. “It can also be useful to reallocate computing power when you ask an AI something (Editor’s note, what we call “inference”) in order to make it think about a specific question. The model will, thanks to this, generate new data, explore a whole field of possibilities around our request and reason about it”, explains the expert. Instead of having an almost instantaneous response like with early versions of ChatGPT, AIs like o1 take a few minutes to think, but their responses are much better.
“We are considering future releases where [l’IA] would think for hours, days or even weeks. Inference costs would be much higher. But what price would we be willing to pay for a new cancer treatment? For revolutionary batteries? For a proof of the Riemann hypothesis (Editor’s note, one of the unsolved mathematical problems considered to be the most important of our time). AI can be much more than a chatbot”, emphasized on X Noam Brownone of OpenAI’s leading researchers. In this context, positioning yourself in AI infrastructures is wise. SoftBank understands this well. The stock market too: after the announcement of the Stargate project, the company’s price jumped by 10%.
Even though Europe has a network of cutting-edge EuroHPC supercomputers (three of which are in the top ten worldwide), it pales in comparison this week to the 500 billion dollars put on the table across the Atlantic. “The Trump administration understood that AI and its ability to increase productivity were going to be a real issue in economic warfare. After restricting the sale of Nvidia processors to 120 countries, including some Europeans, the United States is structuring itself to take full advantage of AI when the EU only plans 1.5 billion for AI factories”, laments Stéphane Roder, president of AIBuilder, in a press release. The game is not lost, however. Mario Draghi’s wise recommendations to make the EU more competitive seemed to be gathering dust since September. Ursula von der Leyen, however, promised on January 21 that the roadmap that the Commission would present next week would rely heavily on it to restore Europe. Knock on wood.
.