OpenAI's Google Chip Plan Tests Its Nvidia Dependence

OpenAI plans to buy Google's AI chips, a move that could give one of Nvidia's major customers another source of computing power. OpenAI intends to buy Google's tensor processing units; CNBC subsequently said it confirmed the intention. Neither company has announced a purchase, and the price, number of chips and delivery schedule remain undisclosed.

Those missing details determine how much changes for anyone using OpenAI's products. A small order might help engineers learn to run models on different hardware. A large deployment could give OpenAI more capacity to train models or generate answers. The plan leaves both possibilities open. It also arrives after OpenAI tried renting Google's chips and while the company works on a chip of its own. Each route could serve a different job, and moving between them takes work.

A Relationship that Began with a Trial

Google's tensor processing units, or TPUs, are processors built for the calculations AI models perform. Google revealed the chips publicly in 2016, after already using them in its own data centers. Customers could access TPUs through Google's cloud, paying for computing time while Google controlled the underlying hardware.

OpenAI's interest predates this week's purchase plan. In June 2025, it had begun renting Google's chips. It had no active plans to use them in its products, OpenAI subsequently said. The activity was early testing without plans to deploy the chips at scale, Decrypt said. That was OpenAI's position during the rental trial; the reported purchase would raise a different question about deployment.

Renting is a relatively contained way to find out whether a model can run well on unfamiliar hardware. Buying chips would pose an operational question: where to put the systems and which workloads to move onto them. A possible purchase would go beyond the rental trial, with its practical effect depending on the order's size and how OpenAI uses it.

Google has also changed what it sells. It began delivering TPUs for customers' own data centers in the second quarter of 2026, after years of developing them for its own use. This week, Alphabet said Oracle would offer access to TPUs through its cloud. Google already has its own cloud and has begun direct hardware sales; Oracle is a planned additional channel. A sale to OpenAI would fit that broader push, even as the two companies compete in AI products and search.

The incentives run in both directions. Google could gain a prominent customer for its chip business. OpenAI could gain another way to secure computing capacity. Their rivalry does not make the transaction implausible; it makes the terms, including how much control OpenAI would have over its hardware, consequential.

The Chip Needs a Software Route In

An AI model cannot be shifted to a different processor by changing a purchase order. Software must translate its operations into instructions the processor can execute, and engineers must optimize how the model uses the chip. OpenAI has advertised for an engineer to help optimize its frontier models for TPUs. The role points to preparation, rather than a completed migration.

Consider the two broad jobs an AI company buys compute to perform. Training builds a model through repeated calculations over data. Inference runs a trained model to produce an answer for a user. Google introduced separate eighth-generation TPUs in April, the 8t for training and the 8i for inference. OpenAI has also unveiled Jalapeño, a chip designed with Broadcom for inference, with initial deployment planned for the end of 2026. A chip designed to serve answers and a possible order of Google hardware need not address the same immediate workload.

That makes OpenAI's approach look like several routes being developed at once. Renting Google capacity can help test a workload. A purchase could provide hardware OpenAI deploys itself. Jalapeño is an attempt to build capacity around OpenAI's own requirements. Each route carries engineering work and a different degree of control. The mix will matter more than the number of chipmakers on a procurement list.

Nvidia also has a relationship with OpenAI beyond selling processors. It has invested tens of billions of dollars in the company, The Information reported. That connection does not prevent OpenAI from seeking other chips, but it complicates any tidy story in which a Google order ends the existing relationship. The practical measure of diversification would be workloads running at scale on alternative systems, alongside the capacity and timing needed to keep them there.

Anthropic Offers a Comparison, with Limits

Another Google customer shows what a more developed TPU arrangement can look like. Anthropic agreed in October 2025 to use up to 1 million TPUs, CNBC said. That agreement puts a public upper limit on the chip count, while OpenAI's planned purchase has no disclosed quantity. An agreement to use up to that number does not mean Anthropic has deployed all of those chips.

Anthropic's arrangement also shows why chip procurement is only part of the bill. Broadcom described a $63 billion financing vehicle, backed by Blackstone and three banks, that would buy TPUs and networking equipment and lease them to Anthropic for five years. In an IPO filing, Anthropic said it uses chips from Google, Amazon, Nvidia and AMD to select hardware for different workloads, CNBC said. Its stated approach is to match a task to a chip, rather than require one architecture to do everything.

OpenAI may have different workloads and commercial terms, so Anthropic's commitments cannot predict the size or performance of an OpenAI deployment. They do show the planning behind access to another chip supplier: hardware, networking, financing and software have to meet in the same place. Google stands to sell into that demand even when its customers compete with its own AI models.

For OpenAI's users, a second supplier could eventually mean more capacity or more flexibility when demand rises. The route from a planned purchase to that outcome runs through delivered chips and models that operate effectively on them. The engineering will determine what a Google order can actually do.