AWS Commits $200B in 2026, Deprecates 12+ Services, and Launches 'AI Factories' for On-Premises Deployments

Amazon confirmed a record $200 billion capital expenditure plan for 2026, the majority directed at AWS data center expansion, custom AI chip development, and infrastructure to support its Bedrock foundation model platform and Amazon Q services. The announcement lands alongside two complementary moves that signal where AWS sees its future: aggressive catalog pruning and a new on-premises deployment model.

$200 Billion and What It Buys

The $200B figure represents a meaningful step up from AWS’s 2025 capex and dwarfs any single-year infrastructure commitment in Amazon’s history. The allocation breaks down into three focus areas:

• Data center expansion across North America, Europe, and Asia-Pacific for core compute and inference workloads
• Custom silicon — continued investment in Trainium and Inferentia chips to reduce dependence on NVIDIA hardware for AI workloads
• AI platform infrastructure for Bedrock (foundation model APIs) and Amazon Q (enterprise AI assistant)

The custom chip push is the most strategically significant element. AWS paid significant NVIDIA GPU costs throughout 2024–2025 while training customers on Bedrock. Trainium 3, expected to enter production later in 2026, targets a 2–3x improvement in training throughput per dollar over its predecessor.

12+ Services Pruned Simultaneously

AWS deprecated more than a dozen services in a coordinated cleanup — an unusually aggressive pruning for a platform that has historically been reluctant to remove anything. The deprecated services span legacy analytics, networking, and developer tooling categories.

The signal is deliberate. AWS has long been criticized for catalog sprawl — hundreds of services with overlapping functionality creating decision paralysis for enterprise architects. The coordinated deprecation tells customers: concentrate on the core services, the rest is being wound down. Resources freed from maintenance overhead go toward the AI platform.

AI Factories: The On-Premises Play

The most architecturally interesting announcement is AI Factories — a managed on-premises deployment model that brings the full AWS tooling stack, the same APIs, the same operational model, into customer-owned data centers.

This is not hybrid cloud in the traditional sense. AI Factories is specifically designed for organizations that cannot send data to a public cloud endpoint — regulated industries, government agencies, defense contractors, financial institutions with strict data residency requirements. The model lets those customers run Bedrock-compatible workloads on infrastructure they physically control, managed by AWS.

The strategic move here is clear: capture the segment of enterprise AI spend that was previously locked out of cloud offerings. Sovereign computing is a real constraint for a large portion of AWS’s potential market, and AI Factories removes that constraint without requiring AWS to compromise on its management model.

What This Means for the Cloud Market

AWS’s $200B commitment sets a benchmark that Google Cloud and Azure will have to respond to. Google confirmed $75B for 2025 earlier in the year; Microsoft has been more guarded about exact figures but has indicated comparable scale. The gap between stated commitments and AWS’s number will pressure both companies at their next earnings calls.

For developers and architects: the service deprecations create short-term migration work but long-term clarity. The AI Factories announcement is worth reading closely if you operate in regulated industries — the pricing model and SLA structure will determine whether it’s a viable alternative to building your own ML infrastructure or just expensive managed hosting with an AWS logo.

Amazon’s infrastructure bet is that AI workloads scale like no computing category before them. $200 billion says they are prepared to be right.