skip to content
Stephen Van Tran
Table of Contents

The most consequential check written in technology this year did not buy a company. It bought a counterweight. On April 24, Alphabet committed up to $40 billion to Anthropic in a structure that pairs $10 billion of immediate cash with $30 billion in milestone-gated follow-ons and an expansion of Google Cloud’s tensor processing unit capacity to a scale that no AI laboratory has ever held under a single roof. TechCrunch’s reporting on the deal captures the structure cleanly: this is not a normal venture round. It is a hyperscaler underwriting a model lab’s entire next decade in a single document, with capital and silicon braided together in a way that makes the two companies harder to separate than any prior AI partnership.

The deal lands in a market that is already running on a different rhythm than the one most boards are budgeting against. Anthropic’s annualized revenue run rate crossed $30 billion in early April, up from roughly $9 billion at the end of 2025, according to PYMNTS. Venture capital offers reportedly value the company at $800 billion, more than double the $380 billion post-money mark from February’s funding round, as TechCrunch noted earlier this month. The Google check is the formal acknowledgment that those numbers are not speculative anymore. They are the floor on which a small handful of laboratories will divide the next phase of enterprise software.

The check that recalibrates the cap table

Google’s $40 billion commitment is the largest single backer pledge any AI laboratory has received from any source, and the structure of it is more revealing than the headline. The immediate $10 billion arrives at a stated $350 billion valuation, per CNBC — a touch below the $380 billion mark Anthropic carried out of February’s primary round, suggesting Google negotiated either a discount on entry or a richer set of compute commitments in exchange for matching the prior tape. The remaining $30 billion is contingent, tied to performance and commercial milestones that neither party has disclosed. That gating mechanism is the same architecture Amazon used in its own April expansion, in which the e-commerce giant committed $5 billion immediately and up to $20 billion more contingent on Anthropic spending against $100 billion of AWS capacity over a decade, as PYMNTS detailed. Two of the world’s largest cloud platforms now hold parallel options on Anthropic’s compute budget, and the laboratory has agreed to spend on both.

The compute layer is the part of the deal that is impossible to replicate at any other AI laboratory in the world. Google Cloud has agreed to deliver five gigawatts of new capacity over the next five years, on top of an earlier 3.5-gigawatt Broadcom-fabricated TPU order announced in early April. The companies’ joint announcement, posted on Anthropic’s site, lays out the underlying economics: Anthropic will use up to one million TPUs, with more than a gigawatt expected online during 2026. Google Cloud CEO Thomas Kurian’s quoted explanation — that “Anthropic’s choice to significantly expand its usage of TPUs reflects the strong price-performance and efficiency its teams have seen with TPUs for several years” — is the politest way of saying that Google’s custom silicon is, at this scale, a structural alternative to Nvidia. Anthropic CFO Krishna Rao framed the demand side: Fortune 500 customers and AI-native startups now depend on Claude for “their most important work,” and the new capacity is what keeps that work running.

What ties the two halves of the deal together is a structural feature of the AI lab business that almost no public analysis has yet metabolized: training and inference compute is now a more important constraint than capital. Cash without contracted gigawatts is a check no one can cash. By bundling $40 billion in capital with five gigawatts of guaranteed silicon — and by stretching delivery across a five-year horizon, with another batch starting in 2027 — Google has handed Anthropic the only thing more valuable than money in the present market. Anthropic, in turn, has delivered the only asset Google needed at this stage of its own competitive strategy: a flagship workload that justifies the multi-billion-dollar investment in TPU capacity Alphabet has been quietly making for two years. The deal is, in a precise sense, the consummation of a barter. Capital flows one way, optionality on the world’s most demanded compute flows the other, and both balance sheets get exactly what they could not buy on the open market.

The math behind a $350 billion handshake

Three numbers explain why Google was willing to write the largest single corporate check in AI lab history. The first is revenue. Anthropic’s $30 billion annualized run rate represents a roughly 230 percent quarter-over-quarter acceleration, with the company moving from $9 billion at year-end 2025 to a level that has reportedly overtaken OpenAI’s revenue trajectory, as SaaStr’s analysis put it. The second number is concentration: more than 1,000 enterprise clients now pay Anthropic over $1 million each annually, a figure that has more than doubled in recent months. That is not a chatbot business. That is a horizontal infrastructure layer being built into the cost-of-goods of large enterprises, where churn is structurally low and procurement decisions are reviewed at the CFO level rather than the AI-curious mid-tier. The third number is the gross-margin profile that follows from those two: a customer base that pays in seven-figure annual contracts is a customer base whose unit economics support the kind of capital expenditure Google’s compute bundle is now financing.

The competitive frame around the deal is the one most boards should be watching. Microsoft’s cumulative commitment to OpenAI sits above $13 billion and was, until very recently, the largest hyperscaler-lab pairing in the industry. Google’s commitment to Anthropic is now roughly three times that figure, and combined with Amazon’s $25 billion maximum exposure plus prior $8 billion infusion, the total committed to Anthropic from its two cloud partners has reached the high-$60-billion range. Across the corridor, OpenAI itself just raised $122 billion in fresh primary capital, according to its own announcement, and has begun publicly reframing its Microsoft relationship after a memo described the partner as having “limited our ability” to reach customers, per CNBC. The balance of power between AI labs and their hyperscaler patrons is no longer a single-axis question. It is a portfolio of relationships, each with its own dilution clock and compute schedule.

The most interesting cross-cut in the data is the implicit price of compute. Five gigawatts over five years is, on industry-average power-use-effectiveness assumptions, enough capacity to support continuous training and inference for a model class roughly an order of magnitude larger than Claude Opus 4.7. Combined with the earlier 3.5-gigawatt Broadcom order and Anthropic’s existing AWS Trainium and Nvidia GPU footprints, the laboratory has secured access to roughly 10 gigawatts of pledged AI compute — a figure that, by my own synthesis of the disclosed Google, Amazon, and Broadcom commitments, equals approximately one-third of the total global AI data center capacity Stanford’s HAI pegged at 29.6 gigawatts in its 2026 AI Index. One private laboratory now controls a third of the world’s contracted frontier AI compute. That is the part of the deal that is structurally unprecedented and the part that will draw the most regulatory attention before the year is out.

The IPO arithmetic is the second cross-cut worth pricing in. Anthropic is reportedly in early conversations with Goldman Sachs, JPMorgan, and Morgan Stanley about a public offering as early as October 2026, with target proceeds north of $60 billion. At an $800 billion implied valuation, the company would list at a multiple roughly 27 times its current run-rate revenue — rich, but not absurd by software comparable standards if the growth curve holds. Google’s $40 billion commitment, structured to convert and lock up at the immediate $350 billion mark, gives Alphabet a position in the laboratory that survives the IPO with substantial paper appreciation built in. That is not incidental. It is the financial mechanism by which Alphabet recovers part of the cost of its compute commitment through equity, rather than through cloud margin alone. The deal monetizes Anthropic’s capitalization at exactly the moment Anthropic needs the largest infusion of compute — and prices the compute at the friction point where Anthropic cannot easily counter.

Where the wires could melt

The simplest way the deal could break is the antitrust filing that has not yet been written. Google’s commitment, layered on top of Microsoft’s OpenAI position and Amazon’s parallel Anthropic exposure, means three hyperscalers now hold the controlling capital relationships over the three labs that account for nearly all frontier model revenue in the United States. The Federal Trade Commission’s 2024 inquiry into hyperscaler-AI lab investments closed without a formal action, but the underlying competitive concern — that compute access is the lever, and that the labs cannot exit the relationship without losing the silicon their products run on — has only intensified. The fact that Google competes directly with Anthropic via its own Gemini line, while simultaneously holding both investor and infrastructure-supplier positions, is the textbook case study in vertical integration that regulators on both sides of the Atlantic will not be able to ignore much longer. Anthropic’s own founder mythology, built around independence from any single corporate parent, is harder to reconcile with a cap table on which Google and Amazon together hold options on more than $65 billion of forward-looking capital.

The second risk is circularity, and it is the financial risk most likely to surface first. The $30 billion contingent tranche is gated on commercial milestones that almost certainly include Anthropic’s spend back into Google Cloud capacity. That is structurally identical to the disclosure-laden circular flows the SEC began asking questions about during last year’s Nvidia-Microsoft-OpenAI tangle. When a hyperscaler invests in a lab that then commits to spending the investment back on the hyperscaler’s compute services, the resulting revenue is real on both sets of books, but the underlying capital does not change hands in the way an arm’s-length investor would assume. Google and Anthropic are running an honest version of this transaction — the silicon is real, the workloads are real — but the optics will become a problem the moment Google’s cloud growth narrative needs to differentiate Anthropic-driven revenue from organic third-party demand. CFO disclosures about “commitments versus realized revenue” will be a recurring line on Alphabet earnings calls for the duration of the agreement.

The third risk is that the milestones do not get met. Anthropic’s run-rate growth is extraordinary, but it is also concentrated. Roughly half of the company’s enterprise revenue, by various analyst estimates, flows through the Claude API for coding and agentic workloads — a category in which OpenAI’s freshly released GPT-5.5, described by TechCrunch as bringing the company “one step closer to a super app”, has now reset the benchmark frontier. China’s DeepSeek V4, previewed on April 24 with TechCrunch coverage, is matching frontier reasoning performance at a fraction of the price. If Anthropic’s revenue growth slows in the second half of 2026 — or if Claude’s coding lead, reasserted with the recent Opus 4.7 release, erodes faster than expected — the contingent $30 billion may convert at a different valuation, or convert in part, or not convert at all on the original schedule. Anthropic’s deal-team has every incentive to keep the milestones private precisely because their disclosure would expose them to renegotiation pressure if performance disappoints.

The fourth risk is the one most likely to age the worst, and it is structural rather than legal. AI lab revenue at this scale is concentrated in a small number of large enterprise contracts that depend on continued, predictable improvement in model capability. The Stanford AI Index documented that capability is still accelerating, but the transparency and safety infrastructure has not kept pace, and Anthropic’s own Mythos model controversy from earlier this month — in which a frontier system designed to find security vulnerabilities reportedly leaked beyond its sanctioned access list — is the kind of incident that, repeated under different circumstances, would force enterprise customers to reassess deployment risk wholesale. A lab whose enterprise base is its valuation is a lab whose valuation can compress quickly if a single high-profile incident triggers procurement freezes across a meaningful fraction of the customer base. Google’s $40 billion is in many ways underwriting the bet that no such incident occurs at scale during the term of the agreement.

Wiring your stack for the multi-hyperscaler era

The Google-Anthropic deal is the third in a sequence — Microsoft-OpenAI, Amazon-Anthropic, Google-Anthropic — that has now hardened into a stable industry structure. The frontier of generative AI will, for the foreseeable future, be defined by three to four laboratories, each capitalized at hundreds of billions of dollars, each tightly bound to one or two cloud parents through compute commitments that span half-decades. That is the world that operators are now buying into, whether or not they have noticed. The deal also rewires a few specific assumptions that have been working in enterprise procurement decks since 2023, and the rewiring matters more than the headline deal size.

The first shift is that lab loyalty is no longer a moral or technical question. It is a procurement and risk-management question. An AI workload running on Claude through AWS Bedrock today may, in 2027, run more efficiently through Google Cloud’s TPU stack — and the customer’s ability to negotiate that move depends on whether their integration architecture treats lab and substrate as separable concerns. The second shift is that compute scarcity, not capital scarcity, is now the binding constraint on AI strategy. A $40 billion check matters because of the gigawatts attached to it; an AI strategy without a capacity plan is a strategy without a denominator. The third shift is that the IPO supply of frontier AI labs is about to become real. If Anthropic prices in October 2026 and OpenAI follows, the public market is about to absorb hundreds of billions of dollars of new AI lab equity — and the secondary effect on private valuations across adjacent categories will be substantial. The pre-IPO secondaries, the wealth-management product lines, the index inclusion debates: all of it is downstream of these primary structures.

The operator checklist that follows is what I would put in front of any executive team revisiting their AI strategy in the wake of the announcement.

  • Treat lab and infrastructure as decoupled procurement axes. The Google-Anthropic deal makes Claude available across more compute substrates, not fewer. Architect your AI workloads so that the model layer can be swapped without a substrate migration and the substrate can be swapped without a model migration. Vendor lock-in at the lab layer is the smaller risk; substrate lock-in to a single hyperscaler is the larger one, and the deal makes the latter easier to mitigate if you start now.

  • Audit your circular-spend exposure before the SEC does. If your organization is both a customer of and an investor in any AI lab through a hyperscaler partnership, document the commercial flows now. The disclosure regime around hyperscaler-AI investment circularity is going to tighten in the next twelve months, and procurement teams whose commitments are entangled with their cloud parents’ equity positions will need clearer separation than most currently maintain.

  • Pre-position for IPO supply. If Anthropic prices in October and OpenAI follows, secondary-market liquidity in adjacent private AI names will inflate rapidly, then compress as the public-market overhang sets in. Treasury and corporate development teams holding pre-IPO AI exposure should be re-pricing their holdings now against the implied public-market comparables, not against private-round marks from six months ago.

  • Map your power exposure end-to-end. Five gigawatts here, five gigawatts there: the AI-driven load on regional power grids is no longer an abstract sustainability question. Enterprise customers running high-volume inference workloads through Anthropic, OpenAI, or any frontier lab are now indirectly exposed to the same regional power and water constraints the labs are racing to secure. Build that exposure into your business-continuity planning explicitly, with named regional dependencies in your vendor risk register.

  • Watch the milestone calendar, not the headline number. The $40 billion is a maximum, not a guarantee. The contingent $30 billion converts on disclosed-only-to-the-parties performance markers. Expect Alphabet’s earnings disclosures over the next four quarters to be the closest thing to a public read on whether those milestones are tracking. Treat each conversion announcement as a signal of where Anthropic’s growth curve is actually landing relative to its private trajectory.

  • Pressure-test your AI vendor concentration. If your enterprise spends more than 20 percent of its AI budget with any single laboratory, the Google-Anthropic deal is a reminder that lab and hyperscaler are now correlated risks. Diversification at the lab level is the simplest hedge available, and the multi-lab routing infrastructure required to make it operational has matured to the point where it is no longer an experimental project.

  • Rewrite your AI vendor due-diligence checklist for the post-$40B world. Old checklists asked about model evaluation, training data provenance, and capability benchmarks. New checklists need to add capital structure, hyperscaler exposure, contingent-tranche schedules, and IPO timing — because the financial and operational stability of your AI vendor now depends as much on those variables as on the capabilities of the underlying model.

The deal is, in the end, a structural acknowledgment that the AI lab market and the cloud market are no longer separable. Google has paid $40 billion to make that point official. Operators who continue to plan for those two markets as if they were independent are now planning for a world that ended on April 24.

In other news

OpenAI ships GPT-5.5 with state-of-the-art agentic coding scores — OpenAI released GPT-5.5 on April 23, claiming an 82.7% score on Terminal-Bench 2.0 and 58.6% on SWE-Bench Pro, with the model rolling out across Plus, Pro, Business, and Enterprise tiers. The release reframes the company’s product line around agentic workflows rather than chat (CNBC).

DeepSeek previews V4 at frontier-class benchmarks and frontier-light prices — China’s DeepSeek unveiled V4-Pro and V4-Flash on April 24, with the larger model claiming 88.4% on MMLU and a one-million-token context window. V4-Flash is priced at $0.14 per million input tokens, an order of magnitude below US frontier alternatives (TechCrunch).

Meta cuts 8,000 staff to fund a $135 billion AI capex push — Meta announced a 10% workforce reduction on April 23, with the cuts taking effect May 20 and 6,000 open roles being closed simultaneously. The company’s 2026 capital expenditure guidance is $115 to $135 billion, up from $72.2 billion in 2025 (CNBC).

Google ships a fresh suite of AI agent tools to compete on automation — Alphabet introduced new agent-building tools at Cloud Next 2026 on April 22, including a Gemini Enterprise Agent Platform, the Agent2Agent protocol, and the new TPU 8t chip, formalizing Google’s pivot from model races to agent-orchestration as the next competitive surface (The Next Web).

DeepSeek’s flagship release puts US labs on direct cost notice — MIT Technology Review’s analysis emphasized that DeepSeek V4 arrived almost exactly a year after the company’s last breakthrough reset cost expectations across the industry, and that the new release runs natively on Huawei Ascend silicon, demonstrating that frontier-class models no longer require Nvidia hardware (MIT Technology Review).