Anthropic and Amazon expand collaboration for up to 5 gigawatts of new compute

Anthropic and Amazon signed a new long-term agreement that secures up to 5GW of compute capacity for Claude, with more than $100 billion committed over the next ten years and $5 billion in new Amazon investment plus up to $20 billion more potential funding. The deal expands Trainium2/3/4 infrastructure, adds inference capacity in Asia and Europe, and should deliver meaningful new capacity within the next three months and nearly 1GW by year-end. Anthropic also said run-rate revenue has surpassed $30 billion, up from about $9 billion at the end of 2025, underscoring rapid demand growth.

Analysis

AMZN is the clear structural winner because this agreement does more than monetize spare capacity: it deepens switching costs at the model, tooling, and procurement layers. The strategic edge is that enterprise buyers prefer a single-cloud, governed path to frontier models, so AWS can capture incremental spend not only from inference but from adjacent workloads like storage, networking, and proprietary accelerators. That creates a second-order flywheel where AI usage lifts core AWS attach rates even if headline margins on custom silicon remain thinner than on pure software.

The supply-chain implication is more interesting than the model narrative. A multi-GW commitment effectively gives Amazon a demand anchor for Trainium iterations, which should compress payback periods for future silicon generations and improve unit economics versus NVIDIA-dependent deployments. The risk is that this also exposes Anthropic to concentration risk: any delay in Trainium3/4 ramps, interconnect bottlenecks, or power delivery issues would hit product reliability before it hits revenue, and consumer-facing usage spikes make that fragility more visible over the next 1-2 quarters.

Consensus may be underestimating how this changes cloud competition. If AWS proves it can deliver frontier-model performance at scale with custom silicon, the real competitive pressure shifts to Azure and Google, whose AI economics rely more heavily on third-party accelerators. That is bearish for the notion that hyperscaler AI capex automatically flows to NVDA; over time, more of the value may migrate into cloud platforms that can internalize inference economics and lock in enterprise governance workflows.

The market may still be underpricing the option value embedded in the deal: if Anthropic keeps growing at this pace, AMZN gets a rare mix of infrastructure demand, strategic equity upside, and platform monetization. The near-term catalyst is AWS disclosures on AI-related revenue acceleration and capex efficiency; the main tail risk is that model demand growth outpaces power and chip deployment, forcing either margin dilution or service degradation before the next capacity tranche arrives.

Key Decisions for Investors

• Long AMZN vs. short MSFT or GOOGL over 1-3 months: own the cloud platform with the strongest AI demand lock-in and most visible custom-silicon monetization path; size for 2:1 upside/downside if AWS AI attach rates re-rate on next print.
• Buy AMZN 6-9 month call spreads on any post-news consolidation: the story is multi-quarter, and volatility should decay once the market digests that this is not a one-off investment but a recurring infrastructure annuity.
• Short NVDA basket hedge against hyperscaler custom-silicon adoption over 3-6 months: use a partial hedge rather than outright short, since the risk is not lost AI spend but margin migration from merchant silicon to cloud-owned chips.
• Watch ANET and power/electrical infrastructure names for a secondary trade: if AMZN meaningfully accelerates AI capex, networking and data-center power bottlenecks become the binding constraint; enter on confirmation from capex guidance revisions.
• Avoid chasing Anthropic private-market exposure at this level unless you have direct visibility into utilization metrics: the better risk/reward is in AMZN, where infrastructure economics and enterprise distribution are already public and scalable.