NewsStocksOpenAI advances AMD partnership as it works to reduce reliance on Nvidia

OpenAI advances AMD partnership as it works to reduce reliance on Nvidia

Author: Fortune Crypto·

Key Takeaways

  • •OpenAI's models currently run almost entirely on Nvidia GPUs, and the company is pursuing diversification through the AMD agreement, its own Jalapeño chip developed with Broadcom, and reportedly planned purchases of Google TPUs.
  • •TensorWave will provide 457 megawatts of capacity across multiple U.S. data center sites, representing 7.6% of the total AMD deal and requiring the company to roughly triple its staff.
  • •The 6-gigawatt AMD partnership represents capacity five times larger than OpenAI's flagship 1.2-gigawatt data center campus in Abilene, Texas, one of the world's largest.
  • •The deployment follows a staggered schedule, with the first gigawatt coming online in the second half of this year, TensorWave's capacity following in the first half of 2027, and full deployment expected by early 2028.
  • •The announcement comes amid turmoil at Nvidia, where former CEO Jensen Huang has launched a comeback campaign, but the board is working with search firms on a replacement and investigations into chip sales to China are underway.
OpenAI advances AMD partnership as it works to reduce reliance on Nvidia

OpenAI is expanding its use of AMD chips and has disclosed the first neocloud provider linked to its 6-gigawatt computing agreement with AMD, as the company works to reduce its near-total dependence on Nvidia GPUs.

OpenAI’s models currently run almost entirely on Nvidia GPUs, according to a source familiar with the company’s compute infrastructure. CEO Sam Altman said this week at a Semafor conference that it would be “hard to imagine the AI revolution without Jensen,” referring to Nvidia co-founder and CEO Jensen Huang.

Anthropic, by comparison, has a more diversified chip base. Its S-1 filing said the company relies on four main providers: Google TPUs, AMD GPUs, Amazon Trainium GPUs, and Nvidia GPUs. Anthropic has also developed a “proprietary orchestration layer” that determines which workloads run on which chips, helping reduce costs.

OpenAI is pursuing several initiatives aimed at replicating that flexibility. The company is designing its own chip architecture, called Jalapeño, and is signing agreements with suppliers other than Nvidia. OpenAI is reportedly planning to buy Google’s TPU chips. It also announced a major development in the computing agreement it signed with AMD in October, under which AMD will provide 6 gigawatts of capacity. OpenAI’s strategic partnership with AMD represents capacity five times larger than the company’s flagship data center in Abilene, Texas, which has 1.2 gigawatts and is one of the world’s largest data center campuses.

AMD is led by Lisa Su, who is Jensen Huang’s first cousin once removed. One source said Su’s least favorite question from reporters concerns the family relationship and that she has not made it part of her public image.

OpenAI has identified TensorWave as the first neocloud company associated with the AMD agreement. TensorWave will provide 457 megawatts of capacity across multiple U.S. data center sites. That represents 7.6% of the total capacity covered by the AMD deal. It will also be TensorWave’s largest deployment to date, requiring the company to roughly triple its staff to support the project, according to a company spokesperson.

The announcement comes amid turmoil at Nvidia, although TensorWave said it had been considering the timing for some time. The company was uncertain whether to proceed following Huang’s resignation but decided to announce the deal during this week’s Supercomputing 2026 conference, which it described as “one of the last major compute conferences of the year.” TensorWave wanted to release the news “ASAP since there will be a lot of great people attending, a lot of great opportunities to meet with potential hires.” Nvidia also has a booth at the conference.

The first gigawatt of capacity under the OpenAI-AMD agreement is scheduled to come online in the second half of this year. TensorWave’s capacity is expected to follow in the first half of 2027, while the full deployment is expected by early 2028. OpenAI’s proprietary Jalapeño chip, which is being developed in partnership with Broadcom, is set to scale up production in 2027. The staggered schedule means OpenAI’s diversification will develop over several years rather than through an immediate shift away from Nvidia.

OpenAI’s broader plan is to become less reliant on Nvidia over time. However, a return by Huang would still be viewed favorably by Altman as a way to preserve stability. Huang has begun a campaign to return, posting on X that he would be “honored” to return as Nvidia CEO.

For now, Huang’s return appears unlikely. Sources said Nvidia’s board is working with multiple search firms to find his replacement, and the board told CNBC that it is not considering bringing Huang back. Investigations into possible wrongdoing, specifically the sale of chips to China, are also underway. The board has created an oversight committee to ensure compliance with U.S. export controls. Huang has agreed to appear before the House in a closed-door hearing on Dec. 1.

Huang is regarded as an almost godlike figure in some AI circles, as Beatrice Nolan and Jeremy Kahn wrote in a Fortune print magazine article. His return remains uncertain.

This story was originally featured on Fortune.com.