Latest in AI
Amazon Q3 2026: AWS Growth, Bedrock Usage and the AI Spending Gap
AWS revenue is accelerating, while Amazon's cash spending shows how much infrastructure has to be built before that demand pays back.
Amazon reported October 29 that AWS revenue grew 39.3% year over year in the third quarter, reaching about $46 billion. Alongside that growth, Amazon reported a trailing-year free-cash-flow outflow of $42.4 billion, driven mainly by higher infrastructure spending that it says primarily reflects AI investment.
The figures describe two different parts of the AI business. Customers are buying more cloud capacity now, while the provider is paying to build capacity intended to serve demand over a much longer period.
The important AWS and spending numbers
Amazon's third-quarter results provide the following figures. Dollar amounts are US dollars.
| Measure | Reported figure | Scope |
|---|---|---|
| AWS quarterly revenue | $46.0 billion | Q3 2026 |
| AWS revenue growth | 39.3% year over year | Q3 2026 |
| AWS operating income | $17.8 billion | Q3 2026 |
| Amazon free cash flow | Negative $42.4 billion | Trailing twelve months |
| Expected cash capital expenditure | About $220 billion | Amazon-wide 2026 guidance |
The capital-expenditure number is company-wide guidance, not a separately disclosed AI-only budget. Similarly, AWS revenue includes the broader cloud business; it should not be relabelled AI revenue.
Strong operating results and negative free cash flow can coexist. Equipment purchases consume cash when the investment is made, while their expense is generally recognised over time through depreciation. The two measures answer different questions about the business.
What Bedrock usage tells us
In its quarterly product highlights, Amazon says Bedrock's total token volume grew nearly 17 times year over year, while Fortune 500 customers' spending grew nearly seven times. It also says customers deployed twice as many AgentCore agents in Q3 as in the preceding three quarters combined.
Those are company-reported adoption indicators. The spending and token figures refer to different populations, so dividing seven by seventeen would not calculate a valid change in price per token. Nor does an agent deployment count tell us how many agents remain active or complete useful work reliably.
Still, these are more concrete signals than a count of companies experimenting with AI. They describe consumption and deployment, which are closer to the actual cloud business. The missing step is connecting that consumption to outcomes for the customer.
Why the provider and customer see different economics
For a cloud provider, a larger inference workload can be a source of revenue. For the company running that workload, it is a cost that needs to produce something useful. Both businesses can be adopting AI successfully while judging it through very different numbers.
Consider an enterprise agent that researches a customer issue and prepares a proposed resolution. More model calls might mean the agent is handling more cases. They might also mean it is retrying failed searches, re-reading context or correcting its own output. A token-growth chart cannot distinguish those explanations by itself.
The same applies to capacity. Building more infrastructure can make a service easier to obtain, but it does not guarantee a lower customer bill. Hardware utilisation, model efficiency, service pricing and the customer's own workflow all sit between new investment and the cost of a completed task.
The useful question after the growth headline
Amazon's results support a clear conclusion: cloud demand and infrastructure investment are both substantial. They do not settle whether every AI application has attractive economics, or whether today's investment will earn the return its builders expect.
For an AI team, the useful response is to track its own cost per successful task alongside total usage. If volume rises because more people use a reliable product, the extra bill can be a sensible trade. If volume rises while completion quality stays flat, the spending may be exposing a workflow problem.
The next phase of AI infrastructure will be judged on how much useful work the new capacity supports, not simply how quickly it can be filled.
I'm Paras. I work at the intersection of AI, engineering and distribution.
I help enterprise AI teams turn complex product economics and technical decisions into clear stories for developers and business leaders.