OpenAI is quietly reshaping the balance of power in the chip industry, and it is doing so for a simple reason: some of the Nvidia hardware it relies on is no longer meeting its needs. Behind the scenes, the company is testing rival accelerators, signing multibillion‑dollar supply deals, and racing to bring its own silicon online. The outcome will determine not just OpenAI’s costs and capabilities, but who controls the next era of artificial intelligence infrastructure.
People familiar with the company’s plans say the dissatisfaction centers on how certain Nvidia chips handle the most demanding inference workloads, especially for coding tools and other latency‑sensitive products. That frustration is now colliding with a broader strategic push by Sam Altman to secure long‑term compute at unprecedented scale, turning OpenAI from a captive customer into an increasingly assertive hardware power in its own right.
Why some Nvidia chips are no longer enough
From the outside, Nvidia still looks unassailable, yet inside OpenAI the picture is more complicated. I have learned that executives and engineers have been unhappy with the way specific Nvidia accelerators perform on inference, the stage where models like ChatGPT and GitHub Copilot respond to user requests in real time. According to Sources, that dissatisfaction is focused on chips used in production services, where every millisecond of delay and every watt of power translates into user experience and margin pressure.
The concerns are not abstract. In coding assistants, where developers expect near‑instant completions, even small inefficiencies can make a premium GPU feel like the wrong tool. One detailed Scoop reports that OpenAI has not been satisfied with Nvidia chips for certain applications like coding and has been actively looking at alternatives, a shift that potentially complicates its long‑standing Nvidia partnership. Another account from Feb underscores that the company is exploring options specifically for inference, not just training, suggesting a targeted response to where Nvidia’s current lineup is weakest for OpenAI’s use cases.
Cerebras and other challengers move into the frame
Once a customer of essentially one supplier, OpenAI is now assembling a roster of chip partners that looks more like a diversified portfolio than a single bet. The most striking example is a reported Cerebras Poses arrangement worth about 10 billion dollars, which gives Cerebras a chance to prove that its wafer‑scale processors can stand in as an Alternative to Nvidia With a large, real‑world deployment. For OpenAI, that kind of deal is not just about raw performance, it is leverage, a way to show Nvidia that it is willing to move meaningful workloads elsewhere if its requirements are not met.
At the same time, OpenAI is broadening its relationships with other incumbents. A recent agreement added Cerebras to a lineup that already included Nvidia, AMD, and Broadcom, with OpenAI publicly confirming that it is designing its own accelerators and that Broadcom will help manufacture some of these custom AI chips. The chip deal highlights Broadcom CEO Hock Tan as a central figure in that collaboration and notes that Later, OpenAI and Broadcom unveiled a partnership that had been in the works for months. In effect, OpenAI is turning its own demand into a proving ground for every chipmaker trying to rival Nvidia, while keeping enough optionality to shift workloads as the technology and pricing evolve.
The custom chip bet with Broadcom and TSMC
OpenAI’s most consequential move, however, is the decision to build its own silicon. The company has been working with Broadcom on a custom accelerator, a project that reflects how hungry OpenAI is for compute and how unwilling it is to leave its fate entirely in Nvidia’s hands. That collaboration, first described as a way to secure more predictable supply and tailored performance, positions Broadcom as both a manufacturing partner and a design collaborator, even as Nvidia is investing 100 billion dollars in its own infrastructure and customers like OpenAI contemplate spending hundreds of billions of dollars on AI capacity over time.
The manufacturing side of that plan is starting to come into focus. According to reporting that cites Commercial Times, OpenAI is expected to deploy a custom AI chip on TSMC’s N3 process by the end of 2026, with a second‑generation design already planned for the A16 node. The Jan analysis notes that Please treat this as an early signal of how aggressively OpenAI is moving to secure TSMC capacity and that Commercial Times has framed TSMC as a critical partner in securing sufficient manufacturing capacity. A separate report explains that OpenAI aims to start mass‑producing its custom AI Chip in 2026, with the design developed in partnership with Broadcom and often compared to Google’s Cloud TPU project. That Sep report underscores how central the Chip program is to OpenAI’s long‑term economics and how closely it mirrors the path taken by other hyperscalers that outgrew off‑the‑shelf GPUs.
Strategic shift and the Sam Altman factor
Behind the technical details sits a broader strategic shift. OpenAI is no longer content to be a price taker in a market where Nvidia’s GPUs set the pace and the bill. Instead, it is advancing a Strategic Shift Towards that explicitly aims to reduce reliance on Nvidia and to bring more of the hardware stack under its own control. That plan includes accelerating custom chip development, deepening ties with Taiwan Semiconductor Manufacturing Company (TSMC) by 2026, and using a mix of in‑house and partner silicon to match specific workloads to the most efficient hardware. In practice, it means OpenAI can reserve Nvidia’s most advanced GPUs for frontier model training while steering inference and specialized tasks to cheaper or more tailored accelerators.
Sam Altman’s ambitions are even larger than that. According to one detailed account, Altman ( Sam Altman ) wants to raise trillions of dollars for an AI chip initiative that would give OpenAI and its allies far greater control over the supply chain. The Altman plan is framed explicitly as a way to secure improved prices and to control the development of the AI ecosystem, not just to eke out incremental performance gains. If even a fraction of that capital materializes, it would reshape the economics of chipmaking, from how fabs are financed to which architectures get priority on the most advanced process nodes.
What this means for Nvidia and the wider AI race
For Nvidia, OpenAI’s moves are both a warning and a validation. On one hand, the fact that a marquee customer is unhappy with some chips and is actively testing alternatives shows the limits of any single vendor’s dominance. On the other, Nvidia’s own roadmap is hardly standing still. A widely watched video analysis notes that even as Nvidia introduced Vera and Rubin, described as its most powerful AI platforms yet, OpenAI and Tesla began moving away to build their own silicon. That Jan commentary captures the tension: Nvidia is racing ahead with Vera and Rubin, while some of its largest customers hedge by designing custom hardware that could eventually displace at least part of their GPU fleets.
The stakes extend beyond one supplier relationship. Analysts are already asking whether Nvidia can remain the most valuable chip company if major buyers like OpenAI, Alphabet, and Tesla increasingly rely on their own accelerators. One Dec forecast argues that Nvidia’s dominance in computing is coming into question and that Tech giant Alphabet might start selling a product that competes directly with Nvidia’s GPUs in some applications, a move that would further erode Nvidia’s lock on AI infrastructure. In that context, OpenAI’s dissatisfaction with certain Nvidia chips is not an isolated complaint, it is part of a broader realignment in which the biggest AI players seek to own more of the stack, from model weights to wafers.
