Amazon Bedrock usage rose sharply in the third quarter of 2026, with the volume of tokens processed by the AI service growing nearly 17 times from a year earlier, according to figures Amazon disclosed on October 29. The company also said spending on Bedrock by Fortune 500 customers was nearly seven times higher. These are measures of customer activity and spending, not independently audited proof that AI projects are generating comparable business returns.
The numbers arrived alongside Amazon’s latest earnings. Amazon Web Services (AWS) recorded $46.0 billion in quarterly revenue, an increase of 39.3% from the year-earlier period and its fastest growth in almost five years. The combination makes the results noteworthy for developers and enterprise technology buyers: demand for hosted AI systems is rising, but so is the scale of the infrastructure needed to deliver them. The figures appear in Amazon’s October 29 earnings announcement.
What nearly 17x growth in Bedrock tokens actually means
A token is a unit of text or other information that an AI model processes. An application consumes tokens when it sends a prompt, supplies supporting material or receives a generated response. More tokens can reflect more users, longer interactions, richer documents, more agent steps or a combination of these factors.
Amazon did not supply the absolute number of Bedrock tokens processed in either comparison period. Without that baseline, the near-17-fold increase establishes the rate of growth, not Bedrock’s total scale relative to rival AI platforms. The nearly sevenfold rise in Fortune 500 customer spending is also a separate metric: it does not mean all Bedrock revenue or every customer’s individual bill grew at that rate.
The distinction matters because token volume is not the same as revenue, and neither figure establishes a project’s return on investment. An AI assistant that reads a large document before every answer may consume substantially more tokens than a simpler workflow while delivering no additional customer value. Enterprises need to connect usage statistics to resolved requests, accuracy, latency and cost per successful task.
AgentCore and managed agents push AWS beyond model access
Amazon’s update also points to growing interest in AI agents: software that can work through a sequence of tasks rather than produce a single response. AWS said tens of thousands of customers were using Bedrock AgentCore, a managed service for operating such agents, and that customers deployed twice as many agents during Q3 as in the preceding three quarters combined. Those figures are Amazon’s own reported adoption measures.
For developers, the relevant product changes extend beyond the headline usage numbers. Amazon announced a public preview of Bedrock Managed Agents, described as an OpenAI-powered managed offering intended to reduce the engineering work needed to deploy AI agents. It also introduced built-in web search for supported OpenAI models on Bedrock, an option for applications that require current information rather than relying solely on a model’s existing knowledge.
Separately, AWS released the Strands agent harness and previewed Strands Box, a sandbox designed to place boundaries on what agents can do. A sandbox may help restrict access and actions, but the existence of a control is not evidence that a deployment is risk-free. Production teams still need to test permissions, log actions, define recovery procedures and check whether human review is necessary for sensitive workflows.
Cloud growth provides the financial backdrop
The wider AWS business remains much larger than Bedrock alone. Revenue reached $46 billion for the three months ended September 30, while AWS operating income increased to about $17.8 billion from $11.4 billion a year earlier. AWS also sells conventional computing, databases, networking, storage and other services. Its 39.3% growth therefore cannot be attributed exclusively to generative AI.
Amazon described the latest AWS sales level as an annualized $184 billion revenue run rate. That number is simply the latest quarter multiplied by four; it is not a reported full-year revenue total or a guarantee of future sales. Chief executive Andy Jassy also projected growth above 40% for AWS in 2027 and 2028, a management forecast rather than an independently established outcome.
Competitors are facing their own trade-offs between AI demand and data-center expansion. BitCodeSolution recently covered Google Cloud’s strong revenue growth and Alphabet’s infrastructure spending. Comparing headline growth rates alone can mislead, however, because the companies report different businesses and operate from different revenue bases. Businesses choosing a provider should use workload-specific pricing and service requirements instead of treating an earnings growth rate as a performance benchmark.
Rising AI investment creates a second question
Amazon expects capital expenditure of approximately $220 billion during 2026. This is a company-wide spending outlook, not a claim that the entire amount is being spent on Bedrock or even exclusively on AWS. The company said increased purchases of property and equipment, primarily related to artificial intelligence, helped drive trailing-12-month free cash flow to a $42.4 billion outflow, compared with a $14.8 billion inflow a year earlier.
Independent Associated Press coverage of the results also highlighted the contrast between rapid cloud growth and the cash demands of expansion. The spending is not equivalent to a quarterly operating loss: Amazon reported company-wide operating income of $28.1 billion. It does, nevertheless, show how much cash can be required before infrastructure investments have time to produce returns.
Amazon also described TorchNeuron, a tool intended to help developers run existing PyTorch workloads on its Trainium AI accelerators without extensive rewriting. Making it easier to use another chip architecture could widen developers’ options, but actual savings will depend on model compatibility, memory needs, performance tuning and the price paid for compute capacity.
What enterprises should check before increasing AI spending
The practical lesson from these figures is not that every organization needs to expand its AI budget. A customer-support platform might track the price of a correctly resolved ticket, while a software team may measure the time spent reviewing generated code. Both should monitor error rates, data exposure, response speed and the number of model calls required to complete a useful result. Teams comparing cloud computing platforms should also account for transfer costs, regional availability and how easily an application can move between providers.
Several important details remain unpublished. Amazon’s announcement does not provide Bedrock’s total token count, the share of AWS revenue attributable to Bedrock, or a standardized measure of enterprise customer outcomes. Public previews such as Bedrock Managed Agents may also have different access conditions and operational capabilities from generally available services. These gaps limit how far the adoption claims can be interpreted.
For now, the October 29 results show two trends moving together: substantial reported growth in AI workload activity and exceptionally heavy investment to support cloud demand. The next test for AWS customers will be whether those expanding capabilities make individual applications more reliable and economical, rather than merely increasing their token consumption.
