AWS grows 39.3% as AI consumption and equipment spending expand

AWS third-quarter revenue rose 39.3% to $46 billion. Amazon’s October 29 results also disclosed how enterprise customers are expanding their use of AI services, giving cloud customers a view of spending and consumption behind the infrastructure buildout.

The cloud business accelerated from the preceding quarter. Amazon’s July 30 release put second-quarter AWS sales at $42.2 billion and growth at 36.7%. Revenue therefore rose by $3.8 billion between the two quarters. The latest presentation also shows quarterly AWS operating income of $17.8 billion, compared with $11.4 billion a year earlier. The July announcement separately put its AI business above a $25 billion annualized revenue run rate, a narrower business measure than total AWS revenue.

Amazon said Fortune 500 spending on Bedrock grew nearly sevenfold year over year, while total token volume grew nearly seventeenfold. The comparisons cover different populations: large-company spending and service-wide processing.

Bedrock is a managed service for building and scaling generative AI applications with foundation models. Its current documentation lists more than 100 supported models and several API options for requesting responses, or inference. The application connects to a model through the service; AWS manages the model-serving environment. That is the service whose enterprise spending and processing volume Amazon reported.

The operating arrangement also separates a model supplier from access to customer requests. According to AWS’s data-protection documentation, Bedrock copies a supplier’s inference and training software into deployment accounts owned and operated by its own service team. Model suppliers cannot access those accounts, the document says, and consequently cannot access Bedrock logs or customers’ prompts and completions.

Customers remain responsible for control of hosted content and the security configuration of services they use. AWS protects the underlying global infrastructure under its shared-responsibility model. Those existing responsibilities describe how enterprise data is managed when applications consume the service.

Pricing also affects the relationship between processing volume and expenditure. AWS lists different rates by provider, model and modality. It offers batch inference for selected models at half the on-demand price. Together with the different customer populations, those distinctions prevent the two growth multiples from establishing a service-wide change in price per token.

For example, pricing in Northern Virginia, Ohio and Oregon for Gemma 4 31B lists $0.14 per million input tokens and $0.40 per million output tokens. Published rates differ from realized averages.

Cash flow after equipment purchases

For the twelve months through September, Amazon’s presentation lists $168.3 billion of operating cash flow and $210.7 billion of property and equipment purchases, net of sales proceeds and incentives. Subtracting the latter produces a $42.4 billion free-cash-flow outflow.

That reconciliation covers all of Amazon over twelve months. It is not an AWS-only cash-flow figure or a quarterly net loss.

At the second-quarter release, trailing free cash flow was a $7.6 billion outflow. Amazon already attributed increased equipment purchases primarily to AI investment. The preceding-quarter figures provide a baseline for both the cloud acceleration and the company-wide investment bill.

Amazon maintained its forecast of approximately $220 billion in cash capital expenditure for 2026. That annual figure remains a management forecast.