February 4, 2026 OpenAI has been looking beyond Nvidia for parts of its artificial intelligence infrastructure, according to people familiar with the matter. The ChatGPT maker has grown dissatisfied with the performance of some of Nvidia’s latest chips for certain tasks, particularly those tied to AI inference, the stage where trained models generate responses for users.
While Nvidia remains the industry leader in chips used to train large AI models, inference has emerged as a new competitive battleground, one where speed and memory access are increasingly critical.
Seven sources said OpenAI has been evaluating alternative hardware since last year, seeking chips that can deliver faster responses for workloads such as software development and AI-to-software communication. The effort is notable because it comes as OpenAI and Nvidia remain in talks over a potential investment that could see Nvidia commit up to US$100 billion to the startup.
Nvidia announced last September that it intended to invest heavily in OpenAI as part of a partnership that would also help fund the startup’s massive chip purchases. The deal was widely expected to close quickly but negotiations have stretched on for months. During that time, OpenAI’s product roadmap has evolved, changing the type of computing resources it needs and slowing progress on the agreement, one person familiar with the talks said.
Nvidia chief executive Jensen Huang dismissed reports of friction over the weekend, calling suggestions of tension “nonsense” and reiterating Nvidia’s intention to make a major investment in OpenAI. Nvidia said in a statement that customers continue to choose its technology for inference because of performance and cost advantages at scale.
OpenAI, for its part, said Nvidia still powers the vast majority of its inference fleet and delivers the best performance per dollar. After the story surfaced, OpenAI chief executive Sam Altman wrote on X that Nvidia makes “the best AI chips in the world” and said OpenAI hopes to remain a major customer for years to come.
Behind the scenes, however, OpenAI has been exploring hardware designed specifically for inference. Sources said the company ultimately wants alternative chips to supply roughly 10 per cent of its inference capacity. Discussions have included startups such as Cerebras and Groq, which focus on designs that embed large amounts of memory directly onto the chip, a feature that can reduce delays when models fetch data.
Inference workloads tend to be more memory-intensive than training, meaning systems often spend more time waiting on memory access than performing calculations. Nvidia and AMD GPUs rely heavily on external memory, which can slow response times for chatbots handling millions of queries.
Nvidia has not stood still. As OpenAI signalled its interest in alternative designs, Nvidia approached companies working on memory-heavy chips about possible acquisitions, people familiar with the matter said. Cerebras declined and instead struck a commercial partnership with OpenAI announced last month. Groq, meanwhile, held talks with OpenAI and attracted investor interest valuing it at about US$14 billion, but by December Nvidia moved to license Groq’s technology in a non-exclusive, all-cash deal and hired away key chip designers.
