Oracle reportedly buying 400,000 Nvidia chips for first Stargate data center

Oracle Corp. is buying $40 billion worth of Nvidia Corp. chips to build a data center for OpenAI, the Financial Times reported today.

The deal is believed to involve about 400,000 GB200 Grace Blackwell processors. Each chip combines two of Nvidia’s latest Blackwell B200 graphics cards with a 72-core central processing unit.

Oracle will use the chips to power a sprawling data center currently under construction in Abilene, Texas. The facility will reportedly consume 1.2 gigawatts of power once it’s fully operational. That corresponds to the electricity usage of about one million households.

The facility is part of Stargate, OpenAI’s initiative to build a network of artificial intelligence data centers in the U.S. The ChatGPT developer has partnered with Oracle, Softbank Group Corp. and Abu Dhabi-based fund MGX on the project. Stargate is expected to cost up to $500 billion over the next four years.

The Financial Times cited sources as saying that OpenAI and SoftBank will each invest $18 billion in the project for a majority stake. Oracle and MGX, in turn, have reportedly committed $7 billion each.

The Texas data center where Oracle plans to install its 400,000 Nvidia chips is set to come online in mid-2026. It’s owned by Crusoe Energy Systems LLC, a Denver-based data center startup, and investment firm Blue Owl Capital. The two companies have reportedly raised about $15 billion from institutional backers, mostly in the form of debt, to finance the project.

Oracle has leased the facility for 15 years. It will make the data center’s computing capacity available to OpenAI for use in AI projects. The companies teamed up after the ChatGPT developer ended an exclusive cloud hosting agreement with Microsoft Corp., one of its largest investors, last year.

The GB200 Grace Blackwell chip that will power the data center can run large language models using up to 25 times less electricity than its predecessor. It also includes other enhancements. According to Nvidia, the chip uses built-in AI models to anticipate technical issues. Meanwhile, a module known as the Decompression Engine speeds up the task of retrieving information from databases.

The chip’s two GPUs are made by Taiwan Semiconductor Manufacturing Co. using a modified version of its N4P process. N4P is a heavily-upgraded iteration of the company’s five-nanometer node.

During the chip manufacturing workflow, transistors are etched into processors using beams of laser light. Those laser beams are not projected directly onto the processors but first go through so-called photomasks. A photomask is a panel that filters some of the light to customize the manner in which transistors are formed.

Compared with TSMC’s earlier processes, N4P requires fewer chip manufacturing steps that involve a photomask. That increases GPU production speeds, which should make it easier for Nvidia to keep up with demand for its chips.

Oracle’s Texas data center is set to be joined by more  Stargate facilities in the U.S. down the line. The company has also partnered with OpenAI to build a Stargate campus in the United Arab Emirates. That data center, which will reportedly use an enhanced version of Nvidia’s GB200 Grace Blackwell chip, is set to come online next year. 

Image: Nvidia