Buying another family of processors does not mean abandoning the previous one. The proposed acquisition of Google chips attributed to OpenAI illustrates this nuance: behind the competition between brands lies an industrial question—the amount of computing capacity available to train and run artificial intelligence models.
On 28 October 2026, Clubic relayed a report by The Information that OpenAI was considering buying Google TPUs. This is a press report about a plan. It does not allow a signed contract, a delivery or the replacement of Nvidia GPUs to be presented as established fact.
A plan to buy, not yet a documented installation
A Reuters dispatch dated 27 October cites the same journalistic source. It says that the financial terms and other details had not been made public; at the time, OpenAI had declined to comment and Google had not immediately responded to the agency.
The fact that several articles have appeared does not, therefore, amount to several independent confirmations of the purchase. The articles refer back to an initial report. This distinction matters to readers: the possibility of negotiations tells us neither how many chips are involved, nor which generation, nor when they might become available for use.
The issue can nevertheless be explained without turning a conditional into a certainty. To do so, three things must be distinguished: the nature of the processors, how they are accessed and how the software that uses them is organised.
What a TPU is
TPU stands for “Tensor Processing Unit”. In its technical documentation, Google describes these components as specialised circuits designed to accelerate machine-learning calculations. They are designed in particular for large matrix operations: calculations involving arrays of numbers that are widely used in neural networks.
The provider also specifies that code intended for these processors runs through the XLA compiler. The hardware is therefore not an interchangeable component whose name alone is enough to predict the result. The programme, its structure and the operations requested all contribute to the performance achieved.
Google also distinguishes situations suited to TPUs from those where GPUs or conventional processors may be more appropriate. Its documentation cites limitations relating, among other things, to certain custom operations or high-precision computing requirements. Specialisation is an advantage for suitable tasks; it does not establish universal superiority.
Renting computing power and buying chips are two different decisions
A company can use an accelerator without owning it, by accessing cloud infrastructure. Buying hardware involves a different arrangement: it is necessary to determine where it will be installed, who will operate it and how it will be integrated with the rest of the system.
The reported plan involving OpenAI does not provide the information needed to describe this arrangement. It would therefore be premature to draw a new map of its data centres or to announce a precise saving per query.
Consider a purely illustrative economic comparison. A team can rent capacity when it needs it, or invest in infrastructure that it expects to use over the long term. In the second case, the price of the component does not tell the whole story: a resource left unused for part of the time can affect the calculation. In the first, flexibility of access says nothing, by itself, about the cost over several years. These questions help put the story into context; they are not known figures for the OpenAI–Google plan.
Diversification did not begin with this report
One public development provides firmer context. The joint announcement by AMD and OpenAI on 6 October 2025 announced a multi-year agreement covering six gigawatts of infrastructure equipped with AMD GPUs, spread across several generations.
The announced schedule envisaged an initial deployment of one gigawatt using the Instinct MI450 series, starting in the second half of 2026. This document confirms the announcement and its scope; it is not proof that all the installations have now been completed. Nor are the six gigawatts a number of chips or a direct measure of how many responses ChatGPT could produce.
This precedent puts the TPU possibility in the context of an already documented move towards using several technologies. It does not confirm a purchase from Google. Above all, it avoids portraying each additional supplier as an exclusive switch from one camp to another.
The questions that will help assess what happens next
If an official announcement is made, four pieces of information will be particularly useful. Which generation of components is involved? Is the programme intended for training, running the models or both? When will the capacity actually be available? What results will be measured using the software actually in use?
It will also be necessary to distinguish between capacity ordered, hardware delivered and an operational service. All these stages may be announced, but they do not describe the same level of progress.
For end users, changing computing provider does not directly guarantee either a lower subscription price or a better response from the model. Those are separate outcomes that must be verified independently. The OpenAI–Google story is therefore interesting as a sign of possible diversification, provided we follow the evidence of implementation rather than already treating it as the verdict in a chip war.
Illustrative photograph: server equipment; it does not depict an OpenAI TPU installation.
