Inference Is the Constraint Now. Anthropic's 3.5GW TPU Lock-In Proves It.

The Inversion No One Priced In

For three years the canonical constraint on frontier AI was training compute: who could rent enough GPU clusters to run the next big run. That constraint is gone. Inference workloads now dominate accelerator cycles and demand different architectural tradeoffs — more on-chip SRAM, lower interconnect diameter, faster collective operations — while training chips need aggregate FP4 flops and HBM capacity. A single monolithic design compromises both. The market has inverted, and Anthropic just placed the largest public bet on that inversion.

Anthropic has secured 3.5 gigawatts of Google TPU capacity via Broadcom, tripling its October 2025 deal, as its revenue run rate has surpassed $30 billion. The figure comes from a Broadcom SEC filing, not a press release. Anthropic did not disclose specifics for its compute expansion, but a recent Broadcom SEC filing shows the deal includes 3.5 gigawatts of compute. That distinction matters: Broadcom had a legal obligation to file. Anthropic chose silence.

What the Structure of This Deal Actually Says

The deal breaks into three parts. First, volume. Amazon Web Services remains Anthropic's primary cloud and training partner under Project Rainier, and this new Google-Broadcom capacity sits alongside that arrangement rather than replacing it. The new capacity is in addition to the 1 GW already coming online in 2026 under the Google Cloud agreement announced last October. Total committed TPU capacity for Anthropic hits 4.5 GW.

Second, timing. The multi-gigawatt capacity is expected to come online in 2027. Booking capacity two-plus years out signals how tight supply is expected to get. Labs that haven't made equivalent commitments won't be able to acquire comparable capacity at spot pricing in 2027.

Third, a contingency clause. A Broadcom regulatory filing notes that the financial arrangements represent sufficient risk to require disclosure, and that the consumption of such expanded AI compute capacity by Anthropic is dependent on Anthropic's continued commercial success. Broadcom is protecting itself, but the clause also reveals the deal hinges on Anthropic's revenue trajectory staying intact. Growth stalls, and the commitment becomes a millstone.

The Revenue Figure Is a Distraction — But Not Irrelevant

Anthropic's revenue run rate has exceeded $30 billion, up from $9 billion at the end of 2025, representing a more than threefold increase in roughly four months. Run-rate revenue is projection, not audited fact. Run-rate revenue extrapolates current revenue to a full-year equivalent. Annual recurring revenue tries the same but restricts itself to customer subscriptions normalized over 12 months, excluding other potential revenue streams. A stronger signal: the company's enterprise customer base of businesses spending more than $1 million annually has simultaneously doubled from 500 to more than 1,000 in roughly five weeks. That cohort expansion is harder to game than a run-rate number.

Recent data shows that Anthropic is now capturing more than 73% of all spending among companies buying AI tools for the first time, while its rival OpenAI is down to around 27%. If accurate, demand isn't hype. It's the reason the capacity deal exists.

Single-Hyperscaler Risk Is Now Material

The infrastructure play has a built-in flaw. Anthropic is deepening its operational dependence on one silicon ecosystem. Anthropic trains and runs Claude on AWS Trainium, Google TPUs, and Nvidia GPUs, arguing that platform diversity translates to better performance and greater resilience. Amazon remains the primary cloud provider and training partner, with continued work on Project Rainier. The multi-vendor framing is real, but it masks the concentration: 3.5 GW of a single vendor's custom silicon, unavailable on the open market, arriving from one supply chain in 2027.

Google and Broadcom have formalized that supply chain. Broadcom has entered a long-term agreement to develop and supply custom Tensor Processing Units for Google's future generations of TPUs. Google has committed to a roadmap stretching from Ironwood shipping today through two purpose-built eighth-generation chips at 2 nanometres in 2027, backed by $185 billion in annual infrastructure spending, a four-partner supply chain, and an anchor customer in Anthropic scaling to 3.5 gigawatts.

A regulatory action touching Google's cloud business, a Broadcom supply chain disruption at TSMC, or a shift in Google's internal TPU allocation strategy would ripple directly into Anthropic's unit economics. TSMC concentration, export controls, and antitrust scrutiny of cloud hyperscalers are all live risks in 2026.

What Google Gets That Everyone Misses

This deal matters most from Google's angle. Google isn't offloading excess capacity. It's using a committed anchor customer to justify the capital expenditure required to scale TPU production. Google projects 4.3 million TPU shipments in 2026, scaling to 35 million by 2028 — a 700% three-year growth trajectory that requires TSMC to allocate a meaningful share of its 2nm and 3nm capacity to Google. Anthropic's commitment de-risks that trajectory. In return, Google locks in long-term revenue, ecosystem stickiness, and an external validator for TPU performance at scale.

Google procures TPUs through Broadcom and pays a margin, but significantly less than Nvidia earns on its GPUs and the entire downstream system—CPUs, switches, NICs, memory, cabling, connectors. The all-in TCO per Ironwood chip in full 3D Torus configuration comes to roughly 44% lower than a GB200 server. That cost structure is the competitive moat Google is building — and Anthropic is the proof point it needs to sell it elsewhere.

What to Watch

1. Watch whether OpenAI or xAI announce equivalent multi-gigawatt commitments with a hyperscaler by Q3 2026. Silence is a concrete signal of capacity scarcity, not neutrality.

2. Track Anthropic's inference pricing against AWS Bedrock and Azure OpenAI pricing through mid-2026. Flat or dropping prices while GPU spot rates rise suggests Google is effectively subsidizing Anthropic's market position through favorable TPU pricing. Rising prices signal cost recovery, not strategic subsidy.

3. Monitor the Broadcom-Google TPU 8t/8i production timeline. TPU 8i triples on-chip SRAM to 384 MB, doubles inter-chip interconnect bandwidth to 19.2 Tb/s, and introduces a new Boardfly topology. Both the 8t and 8i chips run on Axion hosts, support native PyTorch via TorchTPU in preview, and introduce bare-metal access to TPUs for the first time. Any slip in the 2nm ramp at TSMC delays the 2027 capacity date and pushes back Anthropic's inference scaling.

4. Watch for exclusivity language to surface. Current disclosures are silent on whether Anthropic faces restrictions on Nvidia GPUs or AWS Trainium at scale. Exclusivity clauses would sharpen the structural dependency.

5. Observe whether other frontier labs — Mistral, Cohere, any well-capitalized open-weight effort — secure comparable hyperscaler capacity deals. The next twelve months will tell whether multi-gigawatt commitments are available only to labs at Anthropic's revenue scale, or whether they're becoming table stakes for production inference.