Amazon and the AI bubble: spending, returns and risk

Why is Amazon’s AI return a capacity business?

Amazon rents computing power, storage and AI services to companies of every size. When a customer trains or runs an AI model on its cloud, Amazon bills for the hours used. The product is capacity, and capacity has to be built before it can be sold, which makes the timing of demand central. A gap between the two dates is where the risk lives, and everything below measures the size of that gap.

The company also applies AI inside shopping, advertising and warehouse operations. Those uses lower cost or raise sales, but results rarely show them as separate lines. This page therefore concentrates on the cloud arm, where the return can be read in revenue, backlog and margin, and treats the rest as a supporting factor. If you follow Amazon for retail, the same capital spending reading still applies to the cost side.

What does Amazon’s contract backlog tell you about AI demand?

Backlog is the value of signed customer commitments that have not yet been billed. A growing backlog shows buyers lining up for future capacity, and the figure appears in quarterly results. Compare its growth with the growth of capital spending: backlog should rise at least as fast as the money spent to serve it.

Look also at who signs. A backlog spread across thousands of customers behaves like a utility. A backlog dominated by a few AI model builders behaves like a bet on their success, because those builders depend on outside funding. Ask management how concentrated the AI part of the backlog is. Remarks on contract length also help, since long commitments protect revenue while short ones can be cancelled when budgets tighten.

How do custom chips and power supply affect the return on spending?

Amazon designs its own AI chips alongside buying outside accelerators. Own chips can lower the cost of each task and give customers a cheaper option. The test is adoption: ask whether large customers run real workloads on them, and whether usage grows each quarter. A cheaper option wins customers only when the software around it works as smoothly as the alternatives.

Power is the other constraint. A data centre without electricity earns nothing, and grid connections take time to obtain. Capacity that waits for power sits on the balance sheet as cost without revenue. Follow comments on how much capacity is energised, and how long new sites wait for connection. A long wait means the spending of today pays back later than the plan assumed.

What risk comes from AI model builders buying Amazon’s capacity?

Amazon works with outside model builders that commit to large cloud purchases. These deals fill capacity, and they link a share of the revenue to partners whose funding comes from investors. If funding rounds slow, purchase plans can shrink.

The EQOS Asia AI study records 95% of AI projects failing in production (MIT 2025) and about $200B a year invested (April 2026). Those figures remind you that many customers still search for a working use case. Diversified demand from established companies lowers the risk, so watch how the customer mix evolves. A rising share of established companies running steady workloads is a healthy sign for the backlog.

Which margin signs show whether Amazon’s AI spending pays off?

Read the operating margin of the cloud arm. New data centres add depreciation and power cost before revenue arrives, so margin often dips during a build phase. A dip that recovers within a few quarters is normal. A dip that deepens while growth slows is the warning. Depreciation schedules matter here, since equipment bought now lowers profit for several years.

Compare the cloud margin with the revenue growth rate on the same chart. Rising growth with a steady margin shows demand absorbing capacity. Falling growth with a falling margin shows capacity ahead of demand. The shape of the two lines together gives a clearer reading than either alone. Plot them over at least 6 quarters so one unusual period does not mislead you.

Which five Amazon numbers should you follow each quarter?

Follow cloud revenue growth, cloud operating margin, backlog growth, capital spending over operating cash flow, and the share of AI customers that are model builders. They cover demand, profit, commitment, funding and concentration. Each number is public, and each comes from the quarterly report, so the routine takes minutes.

Two or more worsening together in a quarter is the signal to revisit your view. Keep the numbers in a dated table so you compare like with like. Add a column for the direction of change, so the table reads at a glance. The EQOS Asia method page describes how a table of this kind turns into a score.

How do you apply the Amazon reading when you buy cloud capacity?

If your company buys capacity from Amazon, the same logic helps your negotiation. Commit in stages, tie larger volumes to measured use, and keep data portable. Capacity contracts reward buyers who forecast their needs, so build a forecast before you sign. A buyer with a clear forecast can negotiate volume discounts and avoid paying for capacity that stays idle.

This page is general information, not investment advice. It frames risk and leaves prices open. For a larger AI purchase, the Λ simulator gives a first score in 10 minutes, and the 30-minute pre-diagnostic with EQOS Asia applies it to your case.

Frequently asked questions

About the author

Stéphane Jambu

President, EQOS Forecast Asia

Stéphane Jambu leads EQOS Forecast Asia and works as a consultant in e-commerce, search visibility, and AI. LinkedIn profile