Anthropic Is No Longer a Model Vendor

The Structural Shift Nobody Should Miss

The number that matters is not $30 billion. It is 3.5 gigawatts.

Anthropic has signed a deal with Broadcom and Google for the supply of 3.5GW-worth of Google Tensor Processing Units (TPUs), slated to come online from 2027. Paired with a $30 billion annualized revenue run rate, up sharply from roughly $9 billion at the end of 2025, driven largely by enterprise demand, what emerges is not a model company posting strong quarters. It is a company betting that inference is a utility business and building the supply chains to prove it.

Model differentiation expires fast. Capability gaps close within months. Compute reservations measured in gigawatts do not. That asymmetry drives the thesis.

How the Compute Deal Is Structured

Broadcom will supply Anthropic with roughly 3.5 gigawatts of Google TPU capacity starting in 2027, in addition to the 1 GW already coming online in 2026 under the Google Cloud agreement announced last October.

The architecture reveals intent. Google owns both the TPU architecture and software stack, with Broadcom converting that architecture into a manufacturable ASIC layout while supplying high-speed SerDes, power management, and packaging. TSMC handles fabrication. Anthropic sits at the top as the committed anchor customer—not buying spot capacity, but locking in a production floor.

Broadcom's filing noted: "The consumption of such expanded AI compute capacity by Anthropic is dependent on Anthropic's continued commercial success." The language matters. It is not a traditional take-or-pay contract, yet the commitment proves material enough that Broadcom disclosed it to regulators.

Broadcom separately signed a Long Term Agreement with Google to develop and supply future TPU generations, plus a Supply Assurance Agreement to provide networking and components for Google's next-generation AI racks through 2031. That locks the supply chain for five years, with Anthropic as the named downstream customer.

For context: Broadcom's $10 billion custom silicon program with OpenAI, announced as a 10 GW co-development effort last October, makes Broadcom the implementation layer for two of the three largest U.S. frontier model developers. Broadcom has positioned itself as the critical bottleneck between Google's TPU design and anyone seeking scale.

Revenue Trajectory and What Is Driving It

Anthropic's growth has been relentless: $87 million run rate in January 2024, $1 billion by December 2024, $9 billion by end of 2025, $14 billion in February 2026, $19 billion in March, and $30 billion in April.

This is not broad-based. Claude Code, the company's agentic coding tool launched publicly in mid-2025, became the fastest-growing product in company history. It hit $1 billion in annualized revenue within six months. By February 2026, it was generating over $2.5 billion in run-rate revenue.

Anthropic now counts more than 1,000 enterprise clients paying over $1 million annually, more than double recent prior levels, signaling a shift from experimentation to full-scale deployment. Customers spending over $100,000 annually grew 7x in the past year. These are production deployment metrics, not chatbot metrics.

Eight of the Fortune 10 are now Claude customers. When a customer cohort is that concentrated and that large, inference reliability and capacity guarantees become operational dependencies. An outage or latency spike is not a UX problem—it is a business continuity event. That is precisely why the TPU reservation matters operationally.

What This Does to the Competitive Map

The 3.5 GW reservation accomplishes three things.

First, it removes Anthropic from spot compute markets, where prices spike during demand surges and availability is uncertain. Dedicated TPU capacity converts a variable cost line into something predictable. That matters for pricing stability and enterprise SLAs.

Second, it shrinks available Google TPU capacity for competitors. Anthropic is responding to a surge in enterprise usage while shifting toward long-term, utility-scale compute commitments that resemble energy procurement rather than traditional cloud infrastructure. The energy analogy is exact. An industrial buyer locking in a power purchase agreement for a decade is not just hedging price—it is denying that capacity to others.

Third, it deepens Anthropic's dependency on Google infrastructure. Claude remains the only frontier AI model available on all three major cloud platforms: Amazon Web Services (Bedrock), Google Cloud (Vertex AI), and Microsoft Azure (Foundry). That multi-cloud posture is real, but the TPU deal shifts center of gravity toward Google in ways AWS and Azure should monitor. Anthropic trains and runs Claude on AWS Trainium, Google TPUs, and NVIDIA GPUs. Training diversification does not neutralize inference concentration at this scale.

For smaller vendors without hyperscaler backing, the calculus is unforgiving. Mistral, Cohere, and similar operators cannot commit compute at this scale. That is structural fact, not criticism. The arms race has moved to a level requiring either hyperscaler equity or hyperscaler contracts.

The Broadcom Position

Broadcom's role warrants closer attention. Mizuho analysts estimated that Broadcom would record $21 billion in AI revenue from Anthropic in 2026 and $42 billion in 2027. Whether those figures hold is speculative, but the trajectory confirms Broadcom is a primary beneficiary of Anthropic's growth, not a commodity supplier. Hock Tan has stated publicly that hyperscalers lack internal capability to build custom accelerators at scale. Combined with the Anthropic anchor deal and the Google TPU long-term supply agreement, this gives Broadcom a durable position in the AI silicon supply chain through 2031.

Anthropic is not just consuming silicon. It is aligning supply chains around its roadmap.

What to Watch, in Sequence

1. Q3 2026, AWS and Azure: Watch whether either announces a comparable long-term compute commitment with Anthropic. Silence within six months signals the Google relationship has shifted from partnership to primary infrastructure dependency.

2. Broadcom earnings, late 2026: The Mizuho estimates ($21B for 2026, $42B for 2027) are the inflection point. Reported figures will confirm whether the Anthropic deal is translating into recognized revenue or remains contingent.

3. Anthropic IPO timing: Bloomberg reported the company is weighing an IPO as early as October 2026, with Goldman Sachs, JPMorgan, and Morgan Stanley already in discussions. The S-1 will be the first document requiring GAAP revenue, not run-rate. The gap between reported and recognized figures will be the first real test of whether these numbers survive scrutiny.

4. 2027 TPU capacity online: When 3.5 GW comes online, watch Anthropic's inference pricing. A drop signals cost pass-through and competitive pressure. Flat pricing signals margin capture. That single data point reveals more about the long-term business model than any earnings call.

5. Independent model vendors: If a mid-tier company announces a hyperscaler-backed multi-gigawatt deal within six months, the arms race is confirmed. If not, the field is stratifying into two tiers with different cost structures and no bridge between them.