Anthropic signs biggest compute deal yet with Google and Broadcom as revenue run rate hits $30bn

Anthropic agreed to access ~3.5 GW of next‑generation Google TPU capacity via Broadcom starting in 2027 (on top of ~1 GW in 2026) — its largest compute commitment — while disclosing a revenue run‑rate exceeding $30bn (from ~$9bn at end‑2025). The deal cements Broadcom as a critical intermediary (and long‑term Google silicon partner through 2031), supports Anthropic’s $50bn US infrastructure pledge, and comes as Mizuho estimates Broadcom could earn ~$21bn from Anthropic in 2026 and ~$42bn in 2027. Anthropic’s Feb 2026 Series G raised $30bn at a $380bn post‑money valuation and enterprise customers paying >$1m/year have doubled to >1,000, driving urgent multi‑gigawatt capacity needs.

Analysis

This deal effectively converts a chip-and-interconnect supplier into a strategic bottleneck for next‑gen AI infrastructure; Broadcom’s role as the middle layer creates persistent gross margin visibility and optionality that is underpriced by markets still focused on accelerator OEMs. By owning design and supply linkages rather than model IP, Broadcom captures a utility‑like cash flow profile with lower churn risk and higher switching costs once racks are qualified and installed.

The supply‑chain ripple favors capital‑intensive component suppliers (advanced packaging, high‑speed optics, datacenter power and cooling) and domestic construction/real‑estate partners that can host hyperscale racks under US jurisdiction. Conversely, pricing dynamics for on‑demand GPU cloud services will bifurcate: bespoke TPU/ASIC stacks will command a premium for latency‑sensitive enterprise inference, while commodity GPU spot markets will see episodic oversupply pressure as customers arbitrage price vs performance.

Principal execution risks are front‑loaded: manufacturing yield, logistics for rack integration, and qualification cycles could slip on a 6–24 month horizon and compress near‑term estimates of incremental revenue. Policy shocks — export controls, subsidies withdrawal, or unexpected antitrust scrutiny of long‑dated exclusivity arrangements — are 12–36 month tail risks that would re‑open bargaining leverage to hyperscalers and OEMs.

Strategically, labs that maintain true multi‑vendor deployments will be advantaged in negotiating pricing and capacity; firms locked into single‑vendor stacks face escalating renewal bills. That creates a tactical window for vendors that enable rapid cross‑platform porting (toolchains, compilers, orchestration) to monetize migration services and capture follow‑on revenue streams.

Key Decisions for Investors

• Long Broadcom (AVGO) — buy AVGO or 12–18 month LEAP calls: play for durable margin expansion and recurring revenue from custom rack supply. Risk: execution/yield misses or FCF reinvestment delays; reward: asymmetric if multi‑year contracts reprice to higher ASPs. Target timeframe: 6–18 months.
• Pair trade — long AMZN / short NVDA (equal notional) for 3–9 months: AMZN benefits as a diversified cloud/accelerator provider and owner of Trainium-based capacity; NVDA is priced for perfection and vulnerable to rate-of-adoption rotation. Risk: NVDA can outcompete on software and retain enterprise share; reward: captures re‑rating to infrastructure suppliers if hyperscalers monetize proprietary stacks.
• Long select infra enablers (NBIS or datacenter network / optics exposure) — buy 9–18 month calls or small cap exposure: these firms will see order flow and upsell opportunities from rack deployments and domestic buildouts. Risk: project timing slippage; reward: multiple expansion as backlog converts to revenue.
• Hedge catalyst risk with short-dated AVGO puts / buy protection — purchase 3–6 month downside protection sized to 20–30% of AVGO long: protects against near‑term execution or policy headlines while retaining upside exposure to the structural thesis.