Google to invest up to $40B in Anthropic

Read past the $40 billion headline and the structure looks less like a venture investment and more like a forward purchase order for compute. Only $10 billion is upfront cash, struck at a $350 billion valuation; the remaining $30 billion is contingent on Anthropic hitting undisclosed development milestones, and the entire deal is paired with a commitment from Google Cloud to deliver 5 gigawatts of capacity over five years, including access to up to one million seventh-generation Ironwood TPUs. In other words, Google is wiring capital to Anthropic that will largely flow back through Google Cloud invoices when Anthropic trains and serves Claude on those TPUs. The economic substance is closer to a multi-year, milestone-gated supply agreement than to a passive equity stake.

That framing matters because it changes how to evaluate the price tag. eMarketer's Gadjo Sevilla argues Google is "boosting its investment in Anthropic by up to $40 billion—not to win the model race, but to lock in infrastructure dominance." The point of the structure is to convert balance-sheet capital into Cloud revenue and to validate Ironwood TPUs as a credible Nvidia alternative at frontier scale. The contingent tranche additionally lets Google scale its exposure with Anthropic's actual delivery: if Claude continues compounding, Google grows alongside it; if not, Google never writes the next check. The shape of the deal — cash plus compute plus options — is what makes it interesting, not the gross number.

The financial detail buried in Anthropic's own announcement is arguably more shocking than Google's check. Run-rate revenue went from roughly $9 billion at the end of 2025 to more than $30 billion by April 2026 — a roughly threefold increase in about four months. Anthropic also disclosed that the number of enterprise customers spending more than $1 million per year on its products had passed 1,000, doubling in less than two months. That curve is the actual reason Google is pre-paying for compute and Amazon is simultaneously expanding its commitment to up to $25 billion: every quarter Anthropic doesn't have access to additional gigawatts is a quarter of revenue it cannot serve.

This is also why the deal is structured around physical capacity rather than cap-table mechanics. A 5 GW, five-year commitment, with the bulk sited in the U.S., anchors major utility, real-estate and supply-chain investment through 2030. CFO Krishna Rao described the partnership as "a continuation of our disciplined approach to scaling infrastructure," and the language is deliberate: Anthropic is publicly framing multi-hyperscaler buildout — Google TPUs for inference and a growing share of training, AWS Trainium2 via Project Rainier as the primary training base, plus Nvidia GPUs — as a logistics problem rather than a fundraising one. When demand triples in a quarter, capital stops being the bottleneck; megawatts are.

Up to one million Ironwood TPUs and 5 GW of capacity earmarked for a single customer is one of the largest single-tenant accelerator deals on record, and the first time a non-Google frontier lab has committed to consuming TPUs at that magnitude. Broadcom's own securities filing pegs the buildout at roughly 3.5 GW starting in 2027, suggesting the capacity ramp is real enough to be material to its co-designer's revenue guidance. Reporters covering the announcement noted that the market is now willing to price in TPU competition against Nvidia's incumbent position at all — an outcome that would have been unthinkable two years ago.

The second-order effect is on Anthropic's chip portfolio strategy. Anthropic continues to use Nvidia GPUs and remains anchored on AWS Trainium2 for primary training via Project Rainier, but the Google deal makes TPUs a first-class platform alongside the other two. That multi-platform stance is itself a hedge: it gives Anthropic negotiating leverage with each hyperscaler, but it also forces it to invest in the engineering work of keeping Claude performant across heterogeneous silicon. For Google and Broadcom, the validation that custom silicon can rival Nvidia at frontier scale is the prize that makes a $40 billion outlay legible to public-market shareholders.

Strategically, the most important consequence of this announcement may be structural: Google-Anthropic now functions as a clean second axis to Microsoft-OpenAI. Each pairing combines a hyperscaler willing to write multi-tens-of-billions checks with a frontier lab whose models drive that hyperscaler's cloud demand. The symmetry is sharpened by the fact that Anthropic has reportedly overtaken OpenAI on run-rate revenue, which inverts the usual narrative that OpenAI is the commercial leader and Anthropic the safety-focused challenger.

This duopoly framing also explains Google's apparent willingness to invest in a direct competitor to its own Gemini models. As The Next Web's Ioanna Lykiardopoulou puts it, "Google wins whether Anthropic succeeds spectacularly or merely succeeds. The only scenario in which Google loses is the one in which it did not invest." If Gemini wins enterprise share, Google captures the upside directly; if Claude wins, Google captures the inference spend through Cloud and the equity appreciation through this stake. Investor-facing community discussion has settled on the same logic — that Google is essentially "assuring you win even if you lose" — and that framing has now become the default analyst read on the deal.

The contrarian reading, loud both in analyst notes and in community discussion, is that the AI industry's capital flows are increasingly self-referential. Google invests in Anthropic; Anthropic spends that capital on Google Cloud TPUs; Google books the Cloud revenue and the equity appreciation. Amazon's parallel up-to-$25 billion commitment, paired with AWS Trainium2 consumption, follows the same template. Reporting on this round explicitly flags capital flowing from cloud providers into AI labs that pay it back through compute purchases as a potential regulatory and accounting concern. Retail-investor and developer communities have settled on the meme of a "self-circulating economy" — five companies passing the same money back and forth — and that framing is no longer confined to skeptics.

The rebuttal is that the underlying demand is real: Anthropic's revenue growth and its more than 1,000 customers spending over $1M per year are independent evidence that the compute is being used to serve actual workloads, not just to inflate booked revenue. International coverage, including DW Espanol's framing of the cross-investment web among OpenAI, Amazon, Google, Anthropic, and Broadcom as a self-financing AI ecosystem, treats the circularity less as a scandal and more as the defining shape of this cycle. Either way, the deal will sharpen the debate: if Anthropic's growth flattens, the contingent $30 billion never gets written and the circularity argument looks prescient; if growth continues, the structure looks like one of the most efficient capital recyclings in recent corporate history.