Anthropic's compute obligations are now a story about corporate finance, not just technology

In the same week, a number surfaced that is worth sitting with. Krishna Rao, Anthropic’s CFO, told Invest Like The Best that the company has already locked in over $100 billion in compute commitments, split across deals with Google and Amazon, each covering five gigawatts of capacity. Rao described five gigawatts of TPUs starting in 2027 and five gigawatts of Amazon’s Trainium chip. That figure alone would be notable. The additional detail, that another $50 billion is set to flow in from deals inked just last month, transforms the story from a large technology bet into something more structurally unusual.

Harry Stebbings, on 20VC, placed the Google commitment alone at $200 billion over five years. The two figures do not sit neatly together, and the sourcing matters: Rao is speaking as a first-party participant with direct knowledge of the deals; Stebbings is characterizing the arrangement from the outside. Both framings agree on the direction and scale, but a careful reader will hold the specific numbers separately rather than treating them as corroborating data points. What both conversations confirm is that the commitments are extraordinary in absolute terms, and that the Google relationship in particular has taken on a structural weight that goes beyond a standard cloud contract.

Stebbings added a claim that, if accurate, reframes the relationship from Anthropic’s perspective entirely. He suggested that Anthropic’s revenue commitment to Google now accounts for roughly 40 percent of Google’s total future cloud backlog. That number is hearsay from Stebbings’s vantage point, not a confirmed Google disclosure. But the framing it introduces is worth taking seriously on its own terms: if a single customer relationship represents that share of a hyperscaler’s committed revenue pipeline, the dependency runs in both directions. Anthropic is not simply a tenant on Google’s infrastructure. The two companies are, on this reading, financially entangled in a way that constrains the choices of both.

What makes Rao’s account particularly striking is the operational texture he added. He noted that compute consumes 30 to 40 percent of his time as CFO even now. That figure is the tell. Most CFO time allocations at technology companies at Anthropic’s stage cluster around fundraising, financial controls, and investor relations. When a single procurement and financing category commands that share of a CFO’s calendar, it has become the company’s primary financial operating challenge, not a line item. Rao is not describing a capital allocation decision that was made and can now be managed. He is describing an ongoing and consuming set of obligations that require continuous attention.

We have another $50 billion that'll come in into the future from the Amazon and Google deals that we that we inked last month. Krishna Rao · Invest Like The Best

The emergence of this framing across two networks, rather than one, is the signal. It is too early to call it a settled picture of how the AI infrastructure financing story resolves. The specific numbers differ enough between sources to warrant caution. But the structural claim, that a frontier AI lab has entered into compute obligations of a scale and duration more typically associated with national infrastructure projects or utility concessions, is now appearing independently in conversations oriented toward investment and finance rather than engineering. That audience tends to follow the money more carefully than it follows the model benchmarks.

There is a version of this story in which the scale is simply the price of staying at the frontier. Training and inference at the level Anthropic operates requires hardware commitments years in advance, and locking in capacity through multi-year deals with Google and Amazon is the rational response to that reality. The commitments are large because the compute requirements are large, and the compute requirements are large because that is what the product demands. On this reading, the CFO time allocation is high because the problem is genuinely hard to manage, not because something has gone structurally wrong.

But there is another version worth watching. Commitments of this size and duration create obligations that are largely fixed regardless of how the revenue side develops. If Anthropic’s revenue growth does not track the pace implied by the commitment schedule, the company will be servicing capacity it cannot fully monetize. The deals as described by Rao include both Google and Amazon, on separate infrastructure stacks, which adds operational complexity alongside the financial exposure. The $50 billion still incoming from last month’s deals means the commitment base is still growing, not contracting.

Neither Rao nor Stebbings framed any of this as a problem. Rao presented the commitments as a deliberate and considered set of decisions. The concern, if there is one, is not stated in the evidence. It is structural, and it is the kind of thing that looks fine until it does not. A sophisticated reader of these Multiple speakers should come away not with alarm but with a clearer picture of what kind of company Anthropic has become: one where the CFO’s primary job, for 30 to 40 percent of each working week, is managing compute finance at a scale that most companies will never touch. That is a different animal than the research lab that most of the industry’s narrative still defaults to.