SAN FRANCISCO — OpenAI and Nvidia are facing fresh strain as OpenAI searches for faster chips to run its AI models and as talks over a proposed Nvidia investment of up to $100 billion remain unresolved, Feb. 3, 2026.
The tension reflects a broader shift in the AI boom from training giant models to serving them at scale, where milliseconds matter and infrastructure costs pile up. That shift is pushing OpenAI to diversify hardware for “inference” workloads even as Nvidia remains central to its fleet.
OpenAI Nvidia friction grows as inference becomes the new battleground
People familiar with the discussions told Reuters that OpenAI has been dissatisfied with the speed of some of Nvidia’s latest chips for certain inference-heavy tasks and has been evaluating alternatives since last year, aiming to eventually cover roughly 10% of its inference needs with non-Nvidia hardware. The company’s complaints have been especially visible around coding and agent-style workflows that demand quicker responses, according to the report. Reuters reported the details.
OpenAI has publicly tried to keep the message calibrated. After the Reuters report, CEO Sam Altman posted on X that Nvidia makes “the best AI chips in the world” and that OpenAI hopes to remain a major customer, even as it explores other options. Reuters also quoted Nvidia saying customers choose its hardware for performance and total cost of ownership at scale.
OpenAI’s most visible near-term hedge is a new commercial relationship with Cerebras, which OpenAI says is designed to add “ultra low-latency” capacity to its platform. OpenAI described the Cerebras partnership as a way to make AI respond faster on demanding workloads such as generating code, images and agent actions.
At the same time, OpenAI is increasingly operating in an ecosystem where hyperscalers are rolling out their own inference silicon. Microsoft, OpenAI’s longtime partner, recently introduced Maia 200, an inference accelerator it says will serve multiple models, including the latest OpenAI systems. Microsoft outlined Maia 200’s design and deployment as part of a broader push to improve “token generation” economics.
OpenAI Nvidia investment talks: from “$100B” headline to months of uncertainty
Against that hardware rebalancing, the companies’ capital talks have become a story of their own. Nvidia announced in September that it intended to invest as much as $100 billion in OpenAI, but negotiations have dragged on for months, Reuters reported. A separate Reuters report said internal doubts at Nvidia have contributed to the slowdown, with more recent discussions focusing on a smaller equity investment of “tens of billions of dollars.”
Nvidia CEO Jensen Huang has publicly pushed back on the idea of a rupture. In comments reported by TechCrunch, Huang called the notion of friction “nonsense,” while acknowledging the deal structure has been described as nonbinding in reporting about the talks. TechCrunch reported Huang’s response.
One complication: OpenAI’s hardware roadmap is no longer a single-vendor story. Reuters reported that OpenAI has explored or struck deals involving multiple chip suppliers and that Nvidia itself has been moving to shore up inference-adjacent technology, including a reported licensing arrangement tied to Groq. That dynamic can change the leverage and incentives on both sides in OpenAI Nvidia negotiations.
Continuity check: this isn’t OpenAI’s first move away from one-chip dependence
OpenAI’s desire to reduce bottlenecks is years in the making. In 2023, TechCrunch covered reports that OpenAI was considering developing its own AI chips to train and run future models, an early signal that reliance on any single hardware stack could become a strategic risk. TechCrunch reported that 2023 chip push.
In 2024, Reuters reported that Altman was discussing fundraising for a broader initiative aimed at expanding global chip-building capacity, underscoring how supply constraints and compute costs were already shaping strategy. Reuters reported those 2024 fundraising talks.
And in October 2025, Reuters reported that OpenAI had partnered with Broadcom to produce what it described as its first in-house AI processors, with deployment beginning in the second half of 2026. Reuters covered the OpenAI-Broadcom plan.
For now, the OpenAI Nvidia relationship remains mutually beneficial: OpenAI still depends heavily on Nvidia hardware, and Nvidia benefits from being attached to one of the most influential AI developers. The next few quarters will show whether the partnership settles into a new, more diversified equilibrium — or whether the push for faster inference chips and the stalled mega-investment reshape the balance of power.
