The fkra blog
AI industry · 5 min

Anthropic sells through its cloud suppliers

Its cloud partners help it reach customers and give those partners several ways to influence the business.

Listen to this article

5:54
0:00

AI may have been used to generate this audio.

Press Space or K to play or pause, and J or L to go back or forward 15 seconds.

Sources checked 3 October 2026.

A customer can buy Claude through a cloud provider rather than directly from Anthropic. For the customer, that may mean one familiar procurement process and an existing bill. For Anthropic, the sale adds another company to the relationship.

Reuters, citing a confidential prospectus, put 2025 cloud-channel sales at $2.16 billion and distribution fees at $351 million. [1]

Amazon and Google combine investor, supplier and competitor roles. [1] Those roles deserve examination before reaching for a claim about circular money. A marketplace sale can represent a separate customer paying for a useful service. The channel relationship alone does not establish that the revenue is artificial.

Anthropic sales channels
Reuters’s 2025 breakdown; the remainder is calculated. [1]

Dividing fees by channel sales gives about 16.3%. That arithmetic does not measure gross margin. The remainder still has to cover the resources used to serve requests and the other costs of running the business. A customer’s payment can be real while the vendor’s cost of acquiring and serving that customer remains high.

Revenue, marketplace sales and the fees paid back
Reuters on the 2025 filing [1]: revenue of nearly US$4.6 billion, about US$2.16 billion of it sold through the Amazon and Google cloud marketplaces, and roughly US$351 million paid back to them in distribution fees. US$ millions.

It also would not show what a direct sale would have cost. Selling without the channel could require more sales staff, a separate procurement review or customer support that the marketplace relationship partly simplifies. To assess the economics, the useful comparison is the contribution from comparable customers acquired through each route, including what it costs to keep them. Public channel percentages alone cannot supply that comparison.

Why the arrangement is attractive

Imagine a large company whose procurement team already buys through a cloud marketplace. Adding another approved service there could be easier than negotiating a new vendor relationship. That is a plausible distribution advantage, although the size of it would need customer-level evidence.

The computing relationship has a different purpose. In its November 2024 AWS announcement, Anthropic described work with Amazon on Trainium and identified AWS as its primary cloud and training partner. [2] A lab building around that infrastructure can gain access to capacity and engineering support. It also has to account for the work involved in moving away from it.

Partner role

What Anthropic receives

What to examine

Investor

Capital

Ownership and governance terms

Computing supplier

Training and serving capacity

Commitments, capacity and switching work

Sales channel

A route to customer procurement

Fees and control of the customer relationship

Competitor

No automatic benefit from this role

How competing products affect negotiations

The terms matter more than the labels. A long-term capacity agreement could improve planning while reducing flexibility. A marketplace fee could be worthwhile if the channel reaches customers the company would struggle to acquire directly. There is no useful verdict without looking at the benefits alongside the obligations.

The four roles in the table should remain separate in a financial account. An investment purchases an ownership claim. A cloud contract purchases services. A marketplace collects a customer’s bill and charges for distribution. Transactions between the same companies can be connected without being economically identical. To argue that a specific arrangement disguises demand, we would need evidence about that arrangement and its accounting.

There is still a concentration question without such an allegation. Imagine negotiations over future computing capacity while the same counterparty controls an important route to customers. Each contract can be reasonable on its own, yet the combined relationship gives both parties reasons to consider the others. Understanding bargaining power requires looking across the agreements.

One counterparty, four roles
Illustrative: the roles the article describes for Amazon and Google, not data. Each role has its own contract and its own accounting. [1]

It is tempting to solve that with a rule to use several vendors. A second supplier helps only to the extent that work can move to it. A company can pay two providers and still depend on one for a critical capability. Redundancy has to be demonstrated in the application, including the operational work of switching.

What a buyer should ask

A company choosing an AI provider has a smaller version of the same problem. It may already depend on one vendor for identity, cloud hosting and procurement. Buying its model access through that vendor can reduce administration. It can also make a later change more involved than switching an API address.

For a business-critical application, I would map the parts of the service a replacement must provide. Model behaviour is one part. Usage reporting, permissions, billing and regional availability may also affect the move. The map should show what can be changed independently and what requires a coordinated migration.

A practical migration rehearsal could take one bounded customer workflow and run it through an alternative route. Keep the acceptance cases, inspect differences in output, and record the work required to preserve permissions and usage records. The result might justify staying with the current supplier. It would at least turn an abstract claim of portability into a tested scope, with the untested parts named.

That exercise does not require pretending that switching is free. Some dependencies are worth accepting. It does require naming them before a contract, an integration and a set of operating procedures make them harder to revisit.

Anthropic confirmed a confidential draft S-1 submission in June for a proposed IPO, subject to review and market conditions. That announcement does not establish a completed listing. [3] The reporting around the prospectus gives a reason to read the eventual public documents for concentration, commitments and channel terms, rather than rely on the valuation headline.

A partnership can work well for both sides and still leave one side with fewer alternatives. The relevant details are in channel fees, capacity commitments and migration requirements. Those are the terms a buyer or investor needs to examine before deciding how comfortable to be with the concentration.

Sources

0
65 views
AI economicsAnthropicCloudVendor dependenceProcurement
Mosab Alrasheed
Get the next entry

One email when we publish. Research, product decisions, and what teams report back.

Now playingAnthropic sells through its cloud suppliers