Anthropic ups compute deal with Google and Broadcom amid skyrocketing demand

Anthropic signed expanded compute agreements with Google and Broadcom that, according to a Broadcom SEC filing, include 3.5 gigawatts of TPU compute capacity to come online in 2027 and will be primarily housed in the U.S. This extends Anthropic’s prior >1 GW deal and ties to its $50 billion commitment to U.S. compute infrastructure. The company also closed a $30 billion Series G valuing it at $380 billion and reports a $30 billion run-rate revenue (up from $9 billion at end-2025) with >1,000 enterprise customers spending more than $1M annually.

Analysis

Proprietary, long-term compute commitments shift margin and bargaining power away from generic component sellers toward (a) cloud operators that can amortize fixed data‑center costs and (b) silicon OEMs that supply vertically integrated, workload‑specific ASICs. That dynamic raises long‑run customer stickiness — once an enterprise’s production models are trained and deployed on a given accelerator stack, migration costs (retraining, tooling, compliance) create 3–5 year lock‑in windows that favor incumbents and increase predictable recurring revenue for the cloud owner.

Second‑order infrastructural effects matter: large, centralized compute builds compress procurement cycles for server houses and DC power suppliers while stressing local grid capacity and real‑estate supply (US data center permits, transmission upgrades). This amplifies value for companies sitting at the intersection of chips, networking, and data‑center engineering — and it creates an avoidable latency for smaller cloud competitors who can’t match upfront capex commitments.

Key risks are concentrated and lumpy: regulatory/defense procurement actions, a macro hit to enterprise AI spend, or execution slippage in chip ramping can flip the outlook quickly. The most actionable catalyst cadence is quarter‑to‑quarter guidance revisions (near term) and commit-to-deploy schedules around 2026–2028 (medium term). Market pricing today underweights the durability of committed racks but overweights near‑term headline risk from policy/defense scrutiny.

Key Decisions for Investors

• Long AVGO (preferred equity exposure) — buy AVGO shares or buy Jan 2028 LEAP calls ~15–20% OTM; position size 2–4% notional. Rationale: upside from capture of system‑level silicon economics and sticky OEM relationships. Target 12–18 month total return 35–50%; hard stop: cut to -30% from entry on semiconductor end‑market softness or missed design wins.
• Relative trade: long AVGO / short GOOGL (net long AVGO overweight) — target size 2:1 (AVGO:GOOGL) for 6–12 months. Mechanism: expresses view that silicon OEMs reprice faster to the upside while cloud gross margins are more gradually realized. Take profits when AVGO outperforms GOOGL by 15–25%; tighten or flip if GOOGL reports materially higher incremental cloud ARR or regulatory headwinds escalate for AVGO.
• Options funded spread: buy AVGO 12–18 month call (10–20% OTM) funded by selling 30–60 day OTM calls on GOOGL around non‑event windows. This reduces premium decay while keeping directional upside; risk is assignment on short calls—size at 1–2% notional. Exit on IV collapse below 25% or if catalyst timing slips beyond 18 months.
• Event hedge: maintain a 0.5–1% notional put hedge on GOOGL (3–6 months) to protect against rapid sentiment reversal driven by regulatory or defense procurement announcements. This caps downside tail risk without materially diluting upside exposure in the core long positions.