For two years the largest buyer of AI computing has had one supplier for the chips that matter. That arrangement is starting to loosen.
OpenAI plans to buy Google’s tensor processing units, according to people familiar with the matter, a step that would give the company a second source of custom silicon alongside Nvidia. The value of the order and the terms attached to it were not disclosed.
The distinction between renting and buying is the substance of the report. OpenAI has leased TPUs through Google’s cloud since June 2025, an arrangement that gave it access to capacity without owning hardware or committing to an architecture. A purchase changes the calculation: the chips become an asset with a depreciation schedule, and the model developer gains control over where they run and for how long.
Google began selling TPUs directly to selected customers this year, and has agreed to place them in Oracle’s cloud starting in 2027. Until recently the chip existed to serve Google’s own products and its cloud tenants. It is now a product line with external buyers, including one that competes with Google’s own models.
That competition is the part analysts flagged first. Selling accelerators to a rival model developer is a business decision that a hardware company makes easily and a research organization finds harder. The revenue is attractive and the strategic cost is visible: every TPU that trains an OpenAI model is capacity not used by Gemini.
Alphabet’s shares rose on the report. Nvidia’s were roughly flat.
The Nvidia relationship is not small enough to replace. OpenAI is one of the chipmaker’s largest customers, and in August Nvidia committed more than $100 billion in credit support for OpenAI’s data center in Ohio, a facility that will run entirely on Nvidia architecture. Nvidia’s chief executive has said OpenAI’s existing and planned deployments will amount to roughly $600 billion of Nvidia computing by 2030.
Read those two facts together and the shape of the deal becomes clearer. Nvidia has financed a customer’s buildout on the understanding that the buildout uses Nvidia equipment. A purchase of TPUs elsewhere does not cancel that commitment, but it does establish a second lane, and the existence of the lane is what changes the negotiation at renewal.
There is a technical reason the second lane is hard to open. A decade of software investment has made Nvidia’s programming model the default for researchers, and most training code in the industry assumes it. Moving a workload to TPUs means rewriting kernels, retraining engineers and accepting that some libraries do not exist. Google has built its own toolchain for exactly this reason, and the effort required to use it is the moat Nvidia collected by shipping software for free.
OpenAI has an incentive to pay that cost. A single supplier sets price, allocation and delivery schedule, and a customer whose roadmap depends on a product it cannot buy elsewhere has little standing in that conversation. Two sources do not make a buyer powerful, but they make the threat credible.
The timing is not accidental. Demand for accelerators has outrun supply for three years, and every large buyer of AI computing has spent that period signing multiyear reservations rather than purchasing hardware outright. Cloud rentals solved the immediate problem. They also concentrated dependency, since the company that rents capacity also sells a competing model.
Google’s position is the more complicated one. Its cloud business competes with the same companies that buy its chips, its research division competes with their models, and its hardware team now has external customers whose orders help justify fabrication and packaging commitments. Companies have managed that set of conflicts before in semiconductors, usually by treating the chip business as a separate profit center with its own quotas and delivery obligations.
For the industry, the broader pattern is that buyers of AI computing are beginning to multi-source. Meta has taken equity positions with three chipmakers and placed orders across them. Oracle is putting four accelerator families in the same data centers. The assumption that one architecture would serve every workload is being replaced by a portfolio approach in which training, fine-tuning and inference run on different hardware chosen for cost.
Inference is where that shift happens first, because serving a model is a simpler engineering problem than training one and because the volumes are larger. A company that can run inference on cheaper silicon without touching its training stack captures most of the savings with a fraction of the migration risk.
Nvidia’s answer has been to extend beyond the chip. Its credit support for customers, its stakes in cloud providers and its guarantees around data-center leases have turned a component supplier into something closer to a financier of the buildout. That strategy keeps demand attached to its architecture, and it also puts the company’s balance sheet behind its customers’ construction schedules.
The question the TPU report raises is how much of that structure is durable. If the largest model developers can run meaningful workloads on a second architecture, then Nvidia’s pricing power depends less on the absence of alternatives than on the cost of switching, which falls a little each time a competitor proves a workload at scale.
None of this describes a rupture. OpenAI will keep buying Nvidia systems, and the Ohio facility will be built as planned. What changed this week is that a company which once had one option has begun to behave as though it has two, and in a market defined by allocation, the willingness to walk is the only power a buyer has.
