Skip to content

OpenAI Signs Six-Year Deal for 457 MW of AMD Compute From TensorWave

News
Liquid-cooled AI server rack with copper supply pipes and blue coolant hoses running to the cabinet.
The OpenAI TensorWave deal sets aside 457 megawatts of AI compute for OpenAI. The six-year deal will use AMD chips at several U.S. data center sites. The first sites are due to come online in the first half of 2027. TensorWave aims to have the full build ready by early 2028.
The cloud firm announced the deal on November 16. The company will set up AMD Instinct MI455X chips in Helios racks. OpenAI will have sole use of the capacity covered by the deal. The firms kept the price and site names private.

How the OpenAI TensorWave deal fits AMD’s plans

OpenAI and AMD first set out their six-gigawatt plan in October 2025. The plan spans more than one line of chips. The firms set the start of the first gigawatt for the second half of 2026. They planned to use the MI450 series. TensorWave will use chips from that line and follow its own time frame.
The broader deal also gave OpenAI a warrant for up to 160 million AMD shares. The right to buy shares comes in steps. It grows as OpenAI hits goals. One goal is to put more chips to work. The new deal shows where part of that growth is due to come from.
AMD Ventures has invested in TensorWave. TensorWave says it is the first specialist AI cloud named as a host under the deal. It also says the job is its largest yet. The cloud firm raised more cash in June to build out its sites. It says it has secured more than two gigawatts of data center space under long-term deals. The work for OpenAI includes setting up the racks and keeping them cool. It must also link the racks so the chips can share data.

What 457 megawatts covers

TensorWave calls the figure IT capacity. It refers to the power set aside for the gear that does the work. Fans, pumps and other parts of a site draw power as well. The firm did not say how much extra power the sites would need. Actual power use will rise and fall with the load on each site.
The Helios rack design puts 72 AI chips in a rack with EPYC chips and Pensando network gear. Some jobs are too big for one rack. The sites will use fast Ethernet links so racks can share work and move data. The host gave no rack or chip counts by site. OpenAI has yet to say how much of the space it will use to train models. The sites could also run models for users.

When the sites are due to come online

The deal gives AMD another way to grow its data center business. AMD also sells host chips for systems that use AI chips from other firms. Here, it will supply both types of chip. TensorWave will run them for OpenAI.
The near-term test is whether the first sites can come online on time. The firm has yet to say how much space will be ready at that point. OpenAI is due to gain access to the full 457 megawatts by early 2028 if the build stays on track.
loading

Loading