OpenAI Weighs Google TPUs as It Works to Cut Its Nvidia Dependence

OpenAI is negotiating to buy Google's Tensor Processing Units, a move that would chip away at Nvidia's grip on the AI accelerator market. Bloomberg reported the talks on October 27. Neither OpenAI nor Google has confirmed the discussions, and no contract terms are final.

The stakes are straightforward. OpenAI needs enormous amounts of computing power to train and run its models, and today that power comes overwhelmingly from Nvidia GPUs. Buying TPUs would give OpenAI a second supplier and, in time, leverage over pricing.

What exactly is on the table?

Google's TPU is an application-specific AI chip, not a general-purpose GPU. Google began developing the processor more than a decade ago to speed up its search services. The program has since grown into a full line of AI compute chips positioned as direct replacements for Nvidia GPUs.

The TPU already has traction outside Google. Oracle said on October 27 that it plans to offer Google TPUs through its own cloud infrastructure. That matters for OpenAI, because a chip is only useful if cloud providers can actually run it at scale. Oracle's move signals that TPU capacity is becoming accessible beyond Google's own data centers.

A Google executive framed the demand side of the equation. Amin Vahdat emphasized that next-generation AI development requires a variety of computing resources. That argument aligns neatly with OpenAI's apparent shopping list.

The connection between the two companies runs deeper than a vendor negotiation. Richard Ho, who leads OpenAI's in-house chip development team, previously worked on Google's TPU program. His presence inside OpenAI means the company has direct institutional knowledge of the architecture it may now buy.

The takeaway: OpenAI is not merely evaluating a product, it is negotiating with a supplier its own chip chief knows from the inside.

How does the Broadcom chip fit in?

The Google talks are one piece of a broader diversification strategy. OpenAI has already said it plans to begin using its in-house AI chip, codenamed Jalapeno, later this year. OpenAI developed the chip jointly with Broadcom, the silicon design and packaging house that also serves other hyperscalers.

Taken together, the two moves sketch a clear sequence. Jalapeno gives OpenAI a custom accelerator of its own, while a TPU agreement would add a merchant supplier with proven cloud availability. Both reduce the share of OpenAI's compute bill that flows to Nvidia.

The takeaway: custom silicon plus a second merchant vendor is a classic two-front strategy against supplier concentration.

What changes for the AI chip market?

If OpenAI completes the purchase, it would accelerate supply-chain diversification in a market that Nvidia still anchors. The commercial mechanics are simple. When the largest AI developers commit to alternative accelerators, cloud operators follow with capacity, and chipmakers gain the volume they need to fund the next generation.

Analyst commentary accompanying the report treated the near-term effect as neutral, with mid-term sentiment also neutral and long-term sentiment positive. Those judgments, tied to confidence levels of 44, 46 and 49 percent respectively, reflect uncertainty about when diversified compute supply actually shifts demand and cost structures rather than whether the direction is favorable.

Nothing in the reported discussions guarantees a signed contract. Specific terms remain unfinalized, and both companies have stayed silent publicly. Even so, the direction of travel is visible: OpenAI has a homegrown chip arriving this year, a former Google TPU engineer running its silicon team, and a second potential supplier at the negotiating table. If a deal closes, Nvidia keeps the largest share of AI compute, but it will no longer hold it unchallenged.