OpenAI is unhappy with some NVIDIA chips and is looking for alternatives

OpenAI, one of the most prominent companies in the artificial intelligence world, began searching last year for alternatives to some of NVIDIA’s newest chips. Eight sources familiar with the situation told Reuters that this move could seriously complicate relations between two of the main players in the AI boom.

Shift in focus: from training to inference

Previously, OpenAI — like other industry leaders — relied on NVIDIA primarily for training large models, a field where the company’s GPUs dominate. Now, however, the priority is shifting toward inference — the stage where a trained model answers user queries in ChatGPT, generates code, or interacts with other software.

Sources say OpenAI is dissatisfied with the performance of NVIDIA hardware specifically in inference workloads, especially for software development and scenarios where AI systems communicate with other programs. The company is looking for a new hardware platform that could eventually cover up to 10% of its inference computing needs.

Negotiations with NVIDIA have dragged on

In September 2025, NVIDIA announced plans to invest up to $100 billion in OpenAI. The deal was expected to give OpenAI funding to purchase new chips, while NVIDIA would receive a stake in the company. Closing was expected within weeks, but negotiations have now dragged on for months.

One reason for the delay is OpenAI’s shift in priorities. The company needs chips with large amounts of on-die SRAM (static random-access memory), which significantly speeds up inference, where data must constantly be fetched from memory. At NVIDIA and AMD, this memory is typically external, adding latency. OpenAI began discussing potential cooperation with startups Cerebras and Groq, which design chips with large amounts of on-chip SRAM.

NVIDIA moves quickly

While negotiations with OpenAI were ongoing, NVIDIA itself approached Cerebras and Groq. A deal with Cerebras did not materialize, as the startup preferred a commercial partnership with OpenAI. NVIDIA, however, signed a $20 billion licensing agreement with Groq, effectively bringing key chip designers over to its side.

NVIDIA says Groq’s technologies fit well into its own product roadmap and remains confident in its leadership.

“Customers continue to choose NVIDIA for inference because we deliver the best performance and total cost of ownership at scale,” the company said.

In response, OpenAI stated that it still uses NVIDIA hardware for the vast majority of its inference workloads and considers NVIDIA chips the best in terms of price-to-performance.

After the article was published, OpenAI CEO Sam Altman wrote on X:

“NVIDIA makes the best AI chips in the world, and we hope to remain a huge customer of theirs for a very long time.”

Competition is intensifying

While NVIDIA continues to dominate model training, inference is becoming the new battleground. Google has long relied on its own TPUs, which give it an advantage in inference speed. Anthropic and other companies are also exploring alternatives. If OpenAI does reduce its dependence on NVIDIA for inference, it would be a serious test of NVIDIA’s dominance in AI hardware.

In brief

OpenAI is unhappy with inference performance on some of NVIDIA’s newer chips and has been seeking alternatives since last year. This has slowed negotiations over a multibillion-dollar investment. The company discussed partnerships with Cerebras and Groq, but NVIDIA secured Groq through a licensing deal. OpenAI continues to use NVIDIA for most tasks but is actively testing other platforms. Inference is becoming a new front in AI competition — and while NVIDIA remains the leader, pressure is mounting.