Google invests in Anthropic

Background

Anthropic is an AI company that builds large language models. Training and running models at that scale takes an enormous amount of computing power, and almost no model developer owns all the chips and data centres it needs; the usual arrangement is to rent capacity from a cloud provider. Google sits on both sides of that relationship with Anthropic, as an investor and as a supplier of computing, and the agreements between the two companies are easier to read with both roles in mind.^[1][2]

The October 2023 investment

In October 2023, Google invested $500 million in Anthropic and committed a further $1.5 billion over time.^[1] Reuters reported in late October 2023 that Google had agreed to invest up to $2 billion in Anthropic, describing the company in its headline as an OpenAI rival.^[3] The Wall Street Journal reported the agreement before Reuters did.^[3]

The numbers fit together. The $500 million paid at the start plus the $1.5 billion to follow add up to the $2 billion ceiling that the Reuters report described.^[1][3] The shape of the deal — part of the money up front, the rest over time — is also a common way for a large investor to spread a commitment rather than commit the whole sum at once.

A second commitment

In March 2025, Google agreed to invest another $1 billion in Anthropic.^[1] Taken with the earlier amounts, the investment commitments described here come to roughly $3 billion.

Buying compute, October 2025

On 23 October 2025, Anthropic announced that it would expand its use of Google Cloud to as many as 1 million TPUs.^[2] A TPU, or tensor processing unit, is a chip designed for machine-learning calculations rather than for general-purpose computing; renting them by the million is a way of buying training and serving capacity without building the data centres oneself. The deal was worth tens of billions of dollars and was expected to bring more than 1 gigawatt of capacity online in 2026.^[2] At the time, Anthropic served more than 300,000 business customers.^[2]

Power, rather than chip count, has become a common way to describe AI data-centre build-outs, because the electricity and cooling available at a site set a hard limit on how many chips can be run there. A gigawatt is a very large amount of power, comparable to the output of a large generating station, so the figure describes a build-out on the scale of a substantial industrial facility rather than a single server hall. The chip count and the power figure describe the same commitment from two angles: how much hardware, and how much electricity to run it.

Why it matters

The sequence is short and easy to misread. The October 2023 announcement was an investment with a staged structure;^[1] the March 2025 agreement added to it;^[1] the October 2025 announcement was different in kind — a commercial agreement to buy cloud capacity, not another share of the company.^[2] That is why its price is expressed in tens of billions of dollars while the investment commitments are expressed in single billions.

For Anthropic, the arrangement ties its ability to train and serve models to one supplier's build-out schedule. For Google, the same arrangement makes the investor a vendor: money put into the company returns partly as cloud revenue when the company rents chips back. Arrangements of this kind are one reason cloud contracts have become a standard part of how AI labs are financed, alongside direct investment.

The reporting history is a small detail with some value. The Wall Street Journal had the agreement first, and Reuters followed in late October 2023 with the figure of up to $2 billion.^[3] Where a figure like that comes from — briefings, filings or sources close to the deal — is rarely stated, which is why the announced numbers and the reported numbers should be kept apart rather than merged into a single total.

Finally, the customer figure in the October 2025 announcement puts the size of the business in view: more than 300,000 business customers at the time the compute expansion was announced.^[2] Commitments of this scale are made against that kind of customer base, and the capacity is meant to come online in 2026.^[2]