AWS CEO Matt Garman says AI inference is still under 1% of its eventual scale

AWS CEO Matt Garman thinks the AI boom is still in its opening minutes. He says AI inference currently sits at less than 1% of the scale it will eventually reach.

That is a striking claim from the head of a cloud business already running at roughly $169-170 billion a year. It amounts to the view that most of the demand is still ahead.

What Garman actually said

Garman made the comments in an interview on November 10, 2026. His central point concerned AI inference, the stage where a trained model is put to work answering questions, writing code or completing tasks.

Inference already makes up more than half of AWS’s AI revenue, according to figures from late October 2026. Running models, in other words, now earns AWS more money than building them.

Garman also described a pricing dynamic where the cost of each inference keeps falling. Every time it drops by 50%, usage climbs significantly.

The numbers behind the confidence

AWS is growing at 37% year over year. The AI-specific slice of the business is smaller but moving fast. AWS’s AI revenue run rate is projected to land between $15-25 billion.

The more telling figure may be the backlog. As of September 30, 2026, AWS reported a long-term backlog of about $578 billion. That is committed future business, signed but not yet delivered.

Garman’s long-range target is even bigger. He forecasts AWS could reach $1 trillion in annual revenue within five to ten years.

The bottleneck is physical

Demand is not the constraint. Supply is.

AWS installs an AI accelerator approximately every 14 seconds. Capacity remains limited by shortages of power, data center space and chips. Those constraints are expected to persist for a couple of years.

Garman also pointed to Trainium, AWS’s in-house AI chip, which the company plans to sell externally.

Why AWS thinks its customer mix matters

Garman leaned on the breadth of AWS’s customer base as a competitive advantage. The company serves millions of enterprises and startups. He contrasted that with competing cloud providers, which he characterized as more concentrated in their focus.

Garman described inference as a foundational “Lego block” of future digital infrastructure, crucially enabling new task-oriented AI agents. Agents are software that can take actions rather than just respond to prompts.

AWS is also investing in AI safety measures alongside the infrastructure buildout, according to Garman.

What this means for the cloud market

If inference truly sits below 1% of its eventual scale, Garman is telling customers, rivals and shareholders that today’s spending spree is a down payment, not a peak. Falling prices are pulling in more usage, a large backlog is already booked, and AWS cannot install hardware fast enough.

The supply constraints cut both ways. A shortage of power and chips protects pricing in the near term. It also means AWS may be unable to capture all the demand it sees for a couple of years.

The things to track from here are fairly concrete. Watch whether the AI revenue run rate lands at the upper or lower end of its projected $15-25 billion range. Watch whether the roughly $578 billion backlog keeps growing. And watch whether the power and chip crunch eases on the timeline AWS expects.

The $1 trillion target is a forecast, not a guarantee, and it spans a window of five to ten years. Still, Garman has put a clear marker down: AWS believes the inference market it is building for is overwhelmingly still in the future.