Share with your CTO
Amazon is spending $220 billion on AI infrastructure in 2026, up from $200 billion, and CEO Andy Jassy says it still won’t build fast enough to meet demand through 2027, with contracted commitments already extending into 2028. AWS posted $42.2 billion in Q2 revenue, its fastest growth in 18 quarters, and carries a $496 billion backlog growing at a triple-digit rate. The capacity constraint driving the increase is high-bandwidth memory costs, not a construction slowdown. Both AWS’s AI and custom silicon businesses now run above $25 billion in annualized revenue.
What this means for your business
If your organization has workloads queued for AWS AI infrastructure, the supply picture isn’t improving on any timeline that fits a 2026 planning cycle. The companies already holding multi-year commitments, Anthropic and OpenAI on Trainium, enterprises with reserved 2027 capacity, are the ones who moved early. That backlog number, $496 billion, isn’t just a revenue metric; it’s a queue, and every enterprise that hasn’t locked capacity is further back in it than they were a quarter ago.
The $20 billion spending increase wasn’t Amazon deciding to build more data centers. It was the cost of equipping the ones already planned, specifically high-bandwidth memory, the specialized chips that sit alongside GPU accelerators and determine how fast data moves during inference and training. When the bottleneck shifts from concrete to silicon, the lead times and vendor dynamics change entirely. Concrete responds to capital; memory supply responds to fab capacity, which Samsung, SK Hynix, and Micron control on multi-year cycles. Amazon can’t outspend that constraint the way it outspends a zoning delay.
Jassy’s framing of Trainium adoption deserves scrutiny here. Anthropic is substantially Amazon-funded, so its multi-gigawatt Trainium commitment reflects a relationship more than an open-market accelerator evaluation. OpenAI’s commitment is more signal, but the potential for Amazon to sell Trainium outside AWS, which Jassy called genuinely possible, would require building field engineering, public roadmaps, and channel infrastructure that currently don’t exist. The vendors who’ve done that, Nvidia most obviously, spent a decade on it. That timeline matters if you’re a CTO evaluating whether Trainium is a credible alternative to Nvidia for on-premises or co-location workloads in the next two years. It isn’t yet, and the honest answer is that 2027 is the earliest that changes.
The vendor renewal worth reconsidering isn’t your GPU contract. It’s your AWS commitment structure. Jassy’s own framing, data centers as 30-year assets, servers reaching break-even inside three years, tells you exactly how Amazon thinks about pricing power over time. Organizations that lock capacity now trade short-term rate certainty for long-term flexibility. If your AI workload profile is still shifting, which it is for most enterprises mid-deployment, the cost of that inflexibility compounds. The leading indicator to watch is whether AWS starts offering shorter-tenure reserved capacity at a premium rather than only discounting multi-year. That would signal demand softening before any earnings call does.
Concept deep-dive: High-bandwidth memory
High-bandwidth memory, or HBM, is a type of RAM stacked in layers directly alongside an AI accelerator chip, think of it as the chip’s short-term working memory, allowing data to move at speeds a standard memory module can’t match. AI training and inference are bottlenecked by how fast data moves, not just how fast it’s computed. HBM supply is controlled by three manufacturers and tied to semiconductor fabrication cycles measured in years, which is why a $20 billion spending increase can’t simply buy its way out of the constraint.
Based on reporting from Amazon Lifts 2026 AI Capex to $220B, Still Capacity-Constrained, originally published 2026-07-31 15:11:00.
