1. OpenAI Google TPU purchase plan: what was reported
On October 28, a report citing a person familiar with the matter said OpenAI plans to buy artificial intelligence chips from Google, the OpenAI Google TPU arrangement that would reduce the company's reliance on Nvidia, which supplies most of the hardware used to train and run AI models. OpenAI is one of Nvidia's largest customers, so any shift in its purchasing would be closely watched.
Neither OpenAI nor Google has confirmed the plan. The price, the number of chips and the delivery dates have not been disclosed, and the claim rests on a single unnamed source. Readers should treat the details as unverified until one of the companies speaks publicly.
The timing is notable because Alphabet, Google's parent company, is due to report third-quarter earnings after the US market close on Wednesday, when investors may look for any comment on demand for its chips.
The underlying question is one of supply. Training and running modern AI systems requires enormous amounts of specialized computing, and the companies building them have competed for access to chips for several years. A developer that can draw on more than one source gains negotiating room on price and delivery, and is less exposed if a single supplier faces shortages.
2. What are Google's TPUs?
The chips in question are tensor processing units, or TPUs, which Google designed for the large-scale mathematics that AI models perform. Google first revealed them in 2016, saying it had already used them internally for more than a year.
Unlike general-purpose graphics processors, which Nvidia dominates, TPUs are specialized processors built for a narrower set of workloads. For years Google mostly used them in its own services and rented capacity through its cloud business.
In April, Google unveiled two eighth-generation TPUs: the 8t, aimed at training models, and the 8i, aimed at inference, which is the process of generating answers from a model that has already been trained. Google's performance claims compare these chips with its previous generation, called Ironwood, and not with Nvidia hardware, so the figures do not show how they stack up against a competitor.
The distinction between training and inference helps explain why chip choices matter. Training is the expensive, one-time-per-model process of teaching a system from data. Inference happens every time a user asks a question, so for a service with a very large audience it can become the larger and more continuous cost. Chips tuned for inference can therefore reduce operating expenses over time.
3. How long has OpenAI been testing Google's chips?
Reports in June 2025 said OpenAI had begun renting TPUs. OpenAI described that use as early testing and said it had no plans for large-scale deployment at the time. The new report suggests the relationship may be moving from rental to ownership.
The difference is significant. Renting means paying Google by the hour for access to machines in Google's data centers. Buying means owning the hardware, which usually implies a larger commitment, a longer planning horizon and a bigger financial outlay. Other details of any purchase are not public.
An OpenAI job listing in San Francisco adds some weight to the account. It seeks an engineer to lead work on optimizing the company's frontier models for TPUs. The work involves compilers, the software that translates a model's instructions into a form that a particular chip can run, and moving from one chip family to another is considerably harder than swapping a component.
Software compatibility is the main reason chip switching is slow. Much of the AI industry's tooling was written around Nvidia's platform, so adopting a different processor means rewriting or retuning code, testing results and retraining engineers. A company willing to take on that work is making a deliberate strategic choice, which is why the job listing draws attention.
4. A hardware chief with Google roots
Richard Ho, who leads OpenAI's hardware program, spent nearly nine years at Google, where he was among the leaders of its Cloud TPU project, before joining OpenAI. His background means the company has in-house experience with the architecture it is reportedly preparing to buy.
That history does not prove that a purchase is imminent, but it helps explain why a TPU strategy would be practical for OpenAI. Executives who know the design of a chip, and the software required to use it well, are better placed to judge whether it is worth the effort of adoption.
5. Why is OpenAI looking beyond Nvidia?
The report arrives after a series of arrangements that have already spread OpenAI's hardware bets across several suppliers. In September 2025, Nvidia signed a letter of intent to invest up to $100 billion in OpenAI, tied to the deployment of at least 10 gigawatts of Nvidia systems. A gigawatt is one billion watts, a measure of the electrical capacity such data centers would need.
About two weeks later, AMD gave OpenAI the right to buy up to 160 million AMD shares, roughly 10% of the company. That right is tied to OpenAI deploying 6 gigawatts of AMD chips.
Nvidia's eventual commitment turned out smaller than the letter of intent. The company invested $30 billion in OpenAI's funding round, well below the $100 billion figure. Nvidia chief executive Jensen Huang said in March that this may be the company's last such investment before OpenAI goes public.
In June, OpenAI introduced its first custom chip, called Jalapeño, built with Broadcom. It is designed for inference, and its first deployment is planned for the end of 2026. Seen together, these steps describe a company assembling supply from Nvidia, AMD, Broadcom and possibly Google rather than relying on one vendor.
Each arrangement carries different terms. The Nvidia letter tied money to deployment targets, and the AMD right tied share ownership to deployment of a stated number of gigawatts. Such structures link a supplier's fortunes to OpenAI's success, giving chipmakers a reason to deliver capacity and OpenAI a reason to commit to using it. Whether any comparable terms apply to a Google purchase has not been reported.
6. Where does Anthropic fit into the picture?
Anthropic, described as OpenAI's main rival, signed an agreement in April with Google and Broadcom for multiple gigawatts of next-generation TPU capacity, with the capacity expected to come online starting in 2027. If OpenAI also buys TPUs, Google would be supplying chips to the two best-known frontier AI developers at the same time.
That outcome would strengthen Google's position as a seller of AI hardware as well as a developer of models. It would also show that the companies that compete fiercely in AI products can still do business with one another in the supply chain beneath them.
Anthropic's agreement also shows the time scale of such projects. Capacity arriving from 2027 means decisions made in 2026 shape what computing is available years later, and companies must commit to hardware long before the models that will run on it exist.
7. Is Google becoming a chip vendor?
Alphabet's chief financial officer has said the company began delivering TPU systems to customers' own data centers in the second quarter, and expects most revenue from TPU system sales to arrive in 2027. That marks a change from the earlier model in which TPUs were mainly used inside Google or rented through its cloud.
Last week, Google Cloud also began offering OpenAI's GPT-6 models to select customers in a private preview. The two companies therefore already work together on distribution of models, which makes a hardware purchase look less like an isolated event and more like part of a broader commercial relationship.
For Google, selling systems that customers install in their own facilities opens a revenue stream that does not depend on cloud rental. For buyers, it offers an alternative to Nvidia systems, which have faced strong demand.
Alphabet's expectation that most of its TPU system revenue will arrive in 2027 aligns with the dates in the Anthropic agreement, though the report does not say whether any OpenAI purchase would be part of that figure. Investors will be looking for how large the opportunity is, and whether Google can manufacture and deliver the chips on schedule.
8. What to watch next
The first thing to watch is Wednesday's Alphabet earnings report, which comes after the US market close. Executives may comment on TPU demand, deliveries to outside customers, or the progress of the 2027 revenue expectation, though they may decline to discuss any individual customer.
A second marker is confirmation. If OpenAI or Google confirms the purchase, the order's size and timing would reveal how much of OpenAI's computing it is intended to cover. Without those figures, the report shows direction but not scale.
A third is the deployment of Jalapeño at the end of 2026, which will indicate how quickly OpenAI can move part of its workload onto its own silicon. The pace of that rollout, together with any TPU purchase, will show how far the company is prepared to go in diversifying away from a single supplier.
For now, the report is best read as a signal of intent rather than a finished deal. It fits a clear pattern in OpenAI's recent decisions, with every agreement so far aimed at widening the sources of computing it can draw on.
It is also important to keep the limits of the evidence in view. The report is from a single source, the companies have said nothing, and the performance claims that exist were made by Google about its own products relative to earlier versions. Until confirmed details emerge, claims that Google's chips can replace Nvidia's in OpenAI's operations should be treated with caution.