On May 5, 2026, it became public: Anthropic has committed to paying roughly $200 billion to Google Cloud over five years — for compute on Google TPUs, co-designed with Broadcom. That is $40 billion per year, beginning in 2027.
For context: Anthropic’s annualized revenue stood at $30 billion as of April 2026. The annual cloud bill to Google already exceeds today’s annual revenue. The numbers only work in 2027 if growth continues — and if it doesn’t, the structure becomes toxic.
What This Deal Actually Means#
Anthropic is not Google’s partner. Anthropic is Google’s largest single customer.
The $200 billion commitment represents more than 40 percent of Google’s entire cloud revenue backlog. Google has simultaneously pledged up to $40 billion in investment in Anthropic — $10 billion immediately, $30 billion tied to performance targets. Amazon holds a further $25 billion commitment.
One more number: Alphabet reported record profits of $62.6 billion in Q1 2026 — according to Fortune’s analysis, nearly half of that came from the revaluation of its Anthropic stake, not from operational business.
An organization this deeply embedded in an investor’s balance sheet cannot structurally pursue open conflict with that investor.
The Governance Facade#
Anthropic created the Long-Term Benefit Trust (LTBT) to preempt exactly this objection. The Trust is supposed to prevent commercial interests from overwhelming the safety mission. It sounds like a safeguard.
It isn’t — at least not based on what can be reconstructed from the non-public contract text. The Harvard Law School Corporate Governance Blog analyzes: the trustees, according to this analysis, hold no independent enforcement power. Shareholders can amend the Trust by supermajority — without trustee consent. Who currently sits on the Trust is also not fully documented: Jason Matheny departed in December 2023, Paul Christiano in April 2024.
A governance instrument that cannot be enforced and whose composition remains opaque is not a safeguard. It is scenery.
The Pentagon Argument as Distraction#
In February 2026, Dario Amodei rejected a demand from the U.S. Department of Defense to remove contractual exclusions covering mass domestic surveillance and fully autonomous weapons systems.
That demonstrates willpower. Not structural independence.
The relevant question is not whether Anthropic can resist a single contract. The relevant question is which research questions get asked at all when the primary funder is simultaneously the largest investor. Which results get published. Whether internal criticism of Google products ever arises — and if not, why not.
Anyone defending Anthropic’s independence would need to show: Anthropic has published something that damages Google, or has taken action that visibly contradicts Google’s interests. Such evidence is absent. That does not prove that the dependency is already shaping the agenda — but it gives no reason to assume it isn’t.
Structural dependency requires no explicit commands. It operates through omissions.
When Contract Partners Are More Candid Than Press Releases#
Anthropic is the only frontier lab offering its models across all three major cloud platforms — AWS, Google Cloud, Azure. Maximum distribution reach through maximum platform dependency: that is not coincidence, it is a business model — and it ties Anthropic to the same players, on both sides of its balance sheet, that it holds as investors.
Broadcom stated the consequence of this arrangement plainly in an SEC filing:
“Anthropic’s utilization of the enhanced AI compute capacity is dependent on Anthropic’s continued commercial success.”
Commercial success. Not safety progress.
Sources: CNBC/The Information, 05.05.2026 · VentureBeat — Revenue $30B · humai.blog — Revenue vs. Compute · androidheadlines — 40% Cloud Backlog · TechCrunch — Google & Amazon Investment · Fortune — Alphabet Q1 2026 · Harvard Law LTBT Analysis · LessWrong — LTBT Weakness · TechCrunch — Amodei/DoD · Broadcom via letsdatascience — SEC Filing
