OpenAI seeks Nvidia alternatives amid chip performance concerns, despite public denials

Behind the carefully worded statements of friendship between OpenAI and Nvidia lies a more complicated reality. Both companies have spent recent weeks insisting publicly that their relationship remains strong—Sam Altman praising Nvidia's chips as the world's best, Jensen Huang dismissing reports of tension as nonsense. Yet according to eight sources who spoke to Reuters, OpenAI has been quietly shopping for alternatives to Nvidia's hardware since last year, driven by frustration with how slowly the chips process certain kinds of artificial intelligence tasks.

The tension centers on a fundamental shift in how AI companies now spend their money and engineering effort. Nvidia built its dominance on chips designed for training—the computationally brutal work of teaching large language models like ChatGPT to recognize patterns in massive datasets. That business has been enormously profitable. But as AI models have matured, the real-world bottleneck has moved elsewhere. When millions of users ask ChatGPT a question, the system doesn't need to train anything new. It needs to run inference: taking the already-trained model and using it to generate an answer. This is a different problem, requiring different hardware.

Seven of the eight sources told Reuters that OpenAI is unhappy with the speed at which Nvidia's current chips can handle inference, particularly for specialized tasks like software development and AI systems communicating with other software. The company is looking for hardware that could eventually supply about ten percent of its inference computing needs—a meaningful slice of its total infrastructure. The alternative chips OpenAI is pursuing are built around a different architecture: they embed large amounts of SRAM, a type of high-speed memory, directly onto the same piece of silicon as the rest of the processor. Cramming more SRAM onto a single chip creates speed advantages for systems that need to process requests from millions of concurrent users. It's a different engineering philosophy than what Nvidia optimized for.

OpenAI has already signed deals with AMD and other chipmakers to build and test these alternatives. Those negotiations, sources say, did not sit well with Nvidia. The timing makes the friction especially pointed. In September, Nvidia announced plans to invest as much as one hundred billion dollars in OpenAI, a deal the company's CEO called the largest in computing history. The investment would give Nvidia a stake in OpenAI and provide OpenAI with capital to purchase advanced chips. But the deal has stalled. Negotiations that were expected to close within weeks have dragged on for months, and recent reports suggest Nvidia is now considering cutting its planned investment in half.

The public posture from both sides remains one of harmony. Altman wrote on social media that Nvidia makes the best AI chips available and that OpenAI hopes to remain a massive customer for years to come. An OpenAI spokesperson stated that the company relies on Nvidia for the vast majority of its inference operations and that Nvidia delivers superior performance per dollar. Huang, for his part, called the idea of tension between the companies nonsense and reaffirmed Nvidia's commitment to the investment. Yet the gap between these statements and the reporting from multiple sources close to the negotiations suggests that what both companies are saying for public consumption and what they are actually doing behind closed doors have begun to diverge. The question now is whether that divergence will widen as OpenAI's search for inference alternatives matures and as the stalled investment deal continues to languish.