OpenAI may not be happy with Nvidia chips

Citing 8 internal sources, Reuters said that OpenAI is looking for a new AI chip solution to serve the inference process. Nvidia’s graphics processing chip (GPU) is strong at training artificial intelligence models, but its inference ability is weaker than some competitors.

Seven out of eight sources revealed that OpenAI does not appreciate the speed of Nvidia’s hardware in helping ChatGPT process responses quickly, especially with specific issues such as the ability to develop software or AI to communicate with other software. The remaining source said OpenAI really needs new hardware, the immediate goal can meet about 10% of computing needs.

Also according to the source, since last year, Sam Altman’s company has contacted Cerebras and Groq – two famous startups in the field of AI inference hardware. However, the $14 billion deal between OpenAI and Groq was canceled because Nvidia is in the process of acquiring this startup’s assets for $20 billion.

Experts assess that the $20 billion deal between Nvidia and Groq is considered an acknowledgment that the next stage of AI is not determined solely by GPUs.

The AI ​​industry is moving from model training to direct interaction in the real world. This transformation is called inference – the process that takes place when the model answers a question, creates an image or engages in a conversation with the user. According to RBC Capital, the investment banking and capital markets division of Royal Bank of Canada (RBC), inference is becoming the dominant task in AI computing, possibly eclipsing the market for pure artificial intelligence training.

OpenAI is said to have been looking for a solution to replace GPUs through the SRAM solution – high-speed semiconductor memory, using flip-flop circuits (usually 4-6 transistors) to store data without needing to be constantly refreshed like DRAM. With the characteristics of stable data storage when there is power, fast access speed, and low latency, integrating multiple SRAMs into each chip brings speed advantages to chatbots and other AI systems when processing requests from millions of users.

Theo TechCrunchAI inference requires more memory than training, because the chip needs to spend more time retrieving data from memory than performing mathematical operations. Nvidia and AMD’s GPU technology relies on external memory, which increases processing time and slows down the speed of user interaction with chatbots.

With OpenAI, the issue becomes even more important after the company introduced and actively promoted Codex – an AI tool that helps with self-programming. However, the source said some Codex weaknesses are caused by Nvidia hardware. Currently, Codex’s competitors such as Anthropic’s Claude and Google’s Gemini all use their own AI chips, mainly Tensor processors (TPUs) produced by Google.

When OpenAI expressed caution about Nvidia’s technology, Jensen Huang’s company immediately approached a number of businesses developing chips using SRAM, such as Cerebras and Groq.

Nvidia in September 2025 also announced plans to invest $100 billion in OpenAI. If it takes place, the deal will bring large financial resources to OpenAI, as well as access to advanced chip sources to maintain its dominance in AI.

WSJ On January 30, citing sources, Nvidia’s plan was stalled, after some corporate leaders expressed skepticism. A source said that Mr. Huang in recent months has been “unsatisfied with OpenAI”, asserting that the $100 billion deal is non-binding and incomplete. He also criticized OpenAI’s “lack of discipline in its approach to business”, expressing concern with competition from Google and Anthropic.

Nvidia CEO then immediately denied: “Sam is preparing to complete the investment round and we will certainly participate with a very large amount of money, possibly the largest investment in Nvidia’s history.”

OpenAI aims to raise $100 billion and value itself at $830 billion. The company is also negotiating with many other large companies, including Amazon, with agreements that could reach $50 billion, according to Reuters.