Amazon disclosed $200 billion in planned AI infrastructure spend during its latest earnings guidance, the largest single compute commitment by any hyperscaler. The capital will flow into data centers, custom silicon, and power infrastructure over a multi-year horizon. The company did not specify end-date, but capex velocity suggests $75 billion deploys in the next eighteen months alone. AWS already runs the tightest margin profile among the big three clouds; this spend compresses near-term free cash flow in exchange for structural moat.
The announcement follows Amazon's $33 billion compute commitment to Anthropic, formalized as a multi-year chip supply and inference hosting agreement. That deal secures one of the largest known AI workload contracts and guarantees AWS occupancy for Trainium and Inferentia2 chips through at least 2027. The two moves are not separate. Amazon is building the picks-and-shovels layer for foundation models while locking in the anchor tenant. Anthropic gains dedicated capacity without building its own data centers. Amazon gains proof-of-concept at scale for its custom silicon, which costs 40% less per inference than Nvidia H100s in AWS's internal benchmarks.
The capital deployment forces a reckoning for Microsoft and Google, both of which guided $80 billion and $75 billion respectively in AI capex for the current fiscal year. Amazon's figure dwarfs both and signals that compute capacity, not model performance, will define competitive position through 2026. Hyperscalers are no longer competing on API elegance or developer tooling. They are competing on who can provision 500,000-node clusters fastest and keep them thermally stable at 95% utilization. The spend also implies Amazon expects enterprise AI workloads to grow 3x-4x annually through 2028, a faster adoption curve than cloud migration saw in its first five years.
The infrastructure build reshapes power and real estate markets in Virginia, Oregon, and Ohio, where AWS operates its largest availability zones. Data centers at this scale require 2-3 gigawatts of continuous power per region, equivalent to a mid-sized city. Amazon has already signed 15-year power purchase agreements with six nuclear and solar operators, locking in electricity at fixed rates through 2040. That hedging protects gross margin if energy prices spike, but it also creates $12 billion in off-balance-sheet liabilities that don't appear in quarterly capex figures. Allocators should model the full infrastructure stack, not just the reported capital line.
Operators should watch three follow-on effects. First, AWS pricing for GPU instances will likely hold flat through mid-2025 despite capacity constraints, as Amazon uses Anthropic's committed workload to subsidize margin compression. Second, Nvidia's data center revenue guidance for Q2 will indicate whether hyperscaler chip orders are pulling forward or spreading across fiscal quarters. Third, Amazon's next earnings call in late April will clarify whether the $200 billion includes land acquisition and power infrastructure or purely compute hardware. The difference matters for free cash flow timing.
The spend is not speculative. Amazon does not guide capex this size without signed contracts in hand. The infrastructure is already being built.