Nvidia Corp. is reportedly in talks to backstop a $250 billion loan for OpenAI Group PBC.

The Wall Street Journal on Sunday cited sources as saying that the financing would cover the cost of a new data center campus in Ohio. According to CNBC, the site is expected to provide 10 gigabytes of computing capacity. That’s equal to the power consumption of about eight million homes.

The campus is reportedly being developed by SB Energy Inc., a SoftBank Group PBC unit that raised $500 million in funding from OpenAI last year. It builds not only data centers but also energy infrastructure such as battery storage facilities. In 2024, SB Energy constructed a 900-megawatt solar park near Austin for a Google LLC data center.

According to CNBC, the $250 billion loan that Nvidia is in talks to backstop would only cover OpenAI’s lease and certain construction costs. The sum reportedly doesn’t include the graphics processing unit that the Ohio site will host. It’s believed the chips could more than double the project’s price tag to over $500 billion.

The size of the prospective deal hints that the campus will host multiple data centers. Meta Platforms Inc.’s Hyperion campus in Louisiana, which is set to provide half the computing capacity of SB Energy’s Ohio site, is expected to host up to 11 buildings. 

It’s unclear what chips OpenAI plans to install in the facilities. However, the Journal reported that Nvidia is weighing a deal that would see it finance $350 billion worth of graphics card purchases for OpenAI. If that agreement materializes, it’s highly likely the Ohio campus will host Nvidia silicon.

Years to build

A 10-gigawatt campus would take several years to build. As a result, it’s likely that OpenAI will not use Nvidia’s current-generation silicon but rather the more capable GPUs it plans to launch in the near to mid-term.

The chipmaker is gearing up to release a new graphics card called the Rubin Ultra in the second half of 2027. Nvidia will ship it as part of a 72-accelerator rack called the Rubin Ultra NVL576. According to the company, the system will provide 15 exaflops of performance when processing data in the NVFP4 format. Nvidia’s current-generation systems can manage up to 3.5 exaflops. 

The company plans to follow up the Rubin Ultra with a chip series called Feynman in 2028. It will be the first Nvidia graphics card lineup to feature die stacking, which means that its chiplets will be placed atop one another in a three-dimensional configuration. Additionally, the company plans to equip Feynman with a custom variant of HBM, the high-speed memory that GPUs use to store data.

Feynman racks are expected to host as many as 1152 graphics cards. That’s a more than tenfold increase over Nvidia’s Vera Rubin appliances. The density boost will be made possible by the introduction of switches with co-packaged optics technology. CPO hardware makes it possible to link the GPUs in a rack without chips called transceivers that are usually needed for the task. That lowers hardware costs and power use.

If it materializes, Nvidia’s debt backstopping deal with OpenAI would be a major win for SoftBank. The latter company reportedly plans to take its SB Energy unit public later this year at a valuation exceeding $50 billion. A $250 billion contract from OpenAI could help boost investor interest in the offering. 

Image: Nvidia