JAKARTA - The relationship between the two biggest stars in the global artificial intelligence explosion, OpenAI and Nvidia, is beginning to show cracks. OpenAI is said to be dissatisfied with some of Nvidia's latest AI chips and has been actively seeking alternatives to replace them since last year.
This dissatisfaction has the potential to complicate relations between the creators of ChatGPT and the world's most dominant AI chipmaker, especially when the two are still engaged in jumbo investment talks.
OpenAI's strategy shift centers on an increased focus on inference chips, which are the chips used when AI models respond to user requests in real time. Nvidia still dominates chips for training large-scale AI models, but inference is now a new competitive field that is much more sensitive to speed and efficiency.
OpenAI's decision to seek alternatives in the inference chip market is a real test for Nvidia's dominance, especially amid investment negotiations between the two companies. In September 2025, Nvidia expressed its intention to invest up to 100 billion US dollars into OpenAI, which would give Nvidia a stake while ensuring OpenAI has the funds to buy cutting-edge chips.
The deal was originally expected to be completed in a matter of weeks. However, the process dragged on for months. During this period, OpenAI actually collaborated with AMD and a number of other parties for GPUs designed to rival Nvidia. One source said that OpenAI's changing product roadmap also changed computing needs and complicated negotiations with Nvidia.
Nvidia CEO Jensen Huang previously dismissed reports of tensions with OpenAI. "That's nonsense," said Huang, while confirming that Nvidia still plans to make a major investment in OpenAI.
In a separate statement, Nvidia said, "Customers continue to choose NVIDIA for inference because we deliver the best performance and most efficient total cost of ownership at scale."
An OpenAI spokesperson also confirmed that the company still relies on Nvidia for the majority of its inference fleet. "Nvidia provides the best performance per dollar for inference," said the spokesperson.
OpenAI CEO Sam Altman also played down speculation. In a post on the X platform, he said Nvidia makes "the best AI chips in the world" and OpenAI hopes to remain a "giant customer for a very long time."
But behind the calm public statement, seven sources said OpenAI was not satisfied with the speed of Nvidia chips in generating ChatGPT responses for certain types of tasks, such as software development and communication between AI systems. OpenAI is said to need new hardware that can supply about 10% of its inference computing needs.
OpenAI had explored partnerships with chip startups such as Cerebras and Groq, known for developing high-speed inference chips. However, talks with Groq stalled after Nvidia signed a $20 billion licensing deal with the company, according to a source.
Nvidia's move to license Groq technology and recruit key talent from there is seen as an effort to strengthen its technology portfolio amid the fast-moving AI industry. Nvidia said Groq's intellectual property complements its product roadmap.
OpenAI's focus on GPU alternatives since last year has been on chips with large memory embedded directly in the silicon, known as SRAM. This architecture offers speed advantages for chatbots and AI systems that must serve millions of user requests simultaneously.
Inference requires more memory than training, because chips more often retrieve data from memory than perform mathematical computations. Nvidia and AMD GPUs still rely on external memory, which adds latency and slows down responses.
Internally at OpenAI, the problem is most felt in Codex, an AI product for writing code that is now being heavily marketed. A source said that some of Codex's weaknesses are linked to the limitations of Nvidia's GPU-based hardware.
In a call with journalists on January 30, Altman emphasized that speed was a crucial factor for users of AI coding models. "Users will place a large premium on speed for programming work," he said.
Altman said one way OpenAI is meeting those demands is through its latest collaboration with Cerebras. According to him, the need for extreme speed does not always apply to casual ChatGPT users.
Meanwhile, OpenAI's competitors such as Anthropic with Claude and Google with Gemini benefit from the use of their own in-house chips, tensor processing units or TPUs, which are specifically designed for inference and are often more efficient than versatile GPUs.
Interestingly, when OpenAI began voicing its doubts about Nvidia's technology, Nvidia approached SRAM chip development companies, including Cerebras and Groq, for a possible acquisition. Cerebras declined and chose to work with OpenAI on commercial cooperation, which was announced last month.
Groq itself had discussions with OpenAI and attracted investors with a valuation of around 14 billion US dollars. However, in December, Nvidia moved quickly by licensing Groq technology in a non-exclusive cash deal, while recruiting its main chip designers.
In short, amid the global AI fever, even giants are starting to look at alternatives. Nvidia's dominance is still solid, but OpenAI's move is a signal that the AI chip power map could change faster than an app update on a phone.
The English, Chinese, Japanese, Arabic, and French versions are automatically generated by the AI. So there may still be inaccuracies in translating, please always see Indonesian as our main language. (system supported by DigitalSiber.id)
