Stargate, the $100 billion supercomputer

In artificial intelligence, not only the software counts.

Oliver Thansan
Oliver Thansan
30 March 2024 Saturday 10:22
5 Reads
Stargate, the $100 billion supercomputer

In artificial intelligence, not only the software counts. The amount of data that must be processed is so high that only great computing power guarantees a good result. Furthermore, large data centers are needed, because queries to large language models, with hundreds of thousands of parameters, are done over the Internet, so hardware is essential and all technology companies know that they are going to have to invest a lot in this aspect. According to the online newspaper The Information, Microsoft and OpenAI are planning a data center with a supercomputer called Stargate that could cost close to 100 billion dollars, about 92.6 billion euros at the current exchange rate.

The new facility projected by Microsoft and OpenAI should be operational in 2028. Microsoft would be responsible for this multi-million dollar investment. Stargate would be built in the United States and would be the largest supercomputer in a series of several that both companies have proposed creating in the next six years as part of their plans for AI.

The information suggests that the person mentioning the provisional cost of around $100 billion spoke directly with OpenAI CEO Sam Altman, who maintains an excellent relationship with Microsoft CEO Satya Nadella.

The Stargate supercomputer would be the culmination of a five-phase plan agreed upon by both companies, which is currently in its third phase. The information indicates that Microsoft is already working on the fourth phase, with the construction of a supercomputer smaller than the previous one and which is scheduled to come into operation from 2026.

None of the companies have commented on this information. A Microsoft spokesperson, without directly mentioning the Stargate supercomputer, told Reuters in an email: “We are always planning for the next generation of infrastructure innovations needed to continue pushing the frontier of AI capability.”

The speed with which artificial intelligence is being adopted requires investments in large data centers that have skyrocketed demand for specialized chips for this technology. The information about Stargate indicated that the project will have several chip supplier companies. Microsoft itself announced last November the creation of a duo of chips to work with artificial intelligence.

Two weeks ago, the CEO of Nvidia, Jensen Huang, presented the new Blackwell computing platform and skyrocketed its stock market price, becoming the third American company – after Apple and Microsoft – to reach a value of more than two billion. Dollars.

The Blackwell platform includes among its components the B200 chip, which has a power of 208 billion transistors and far exceeds the performance of Nvidia's AI accelerators, which are leaders in this sector. According to Huang, artificial intelligence now needs “bigger GPUs.” Nvidia chips are not CPUs, central processors, but GPUs, graphic processing units, more capable in calculations. The latest large AI language models can be trained with 2,000 Blackwell GPUs and consume 4 megawatts of power in 90 days. The same task with the previous Nvidia chips would require 8,000 GPUs with a consumption of 90 megawatts in the same period.