AWS recorded its fastest growth in more than four years during the second quarter, but Amazon said it still lacked enough computing capacity to meet customer demand despite raising its 2026 capital spending forecast to $220 billion.
AWS revenue rose 37% year over year to $42.2 billion in the quarter ended June 30. The result exceeded the 31.21% growth expected by analysts surveyed by LSEG and marked the cloud unit’s strongest performance in 18 quarters.
“AWS is booming right now,” Amazon CEO Andy Jassy said during the company’s earnings call. AWS added more than $4.6 billion in revenue from the previous quarter, about 80% more than its previous largest quarterly increase.
Amazon increased its forecast for 2026 cash capital expenditure from approximately $200 billion to $220 billion. Jassy said the additional spending would still not provide enough infrastructure to meet all the demand Amazon expected during the year.
“Even at that amount, we will still not have enough capacity to meet all of the demand we have in 2026,” Jassy said. He expected the capacity constraint to continue in 2027.
The “lion’s share” of AWS computing capacity planned for 2027 had already been reserved by customers, Jassy said, while some capacity scheduled for 2028 had also been committed. AWS backlog reached $496 billion at the end of the quarter, up from $364 billion three months earlier.
The reservations provide Amazon with advance information about customer requirements as it plans data centres and computing infrastructure. The company did not identify the customers involved or disclose how much capacity would remain for organisations procuring resources closer to deployment.
Customer commitments extend beyond current capacity
AWS offers products that allow customers to reserve specified computing resources in advance. EC2 Capacity Blocks provide access to some GPU-based instances for defined periods, while On-Demand Capacity Reservations secure selected EC2 capacity within an Availability Zone.
These products are separate from the longer-term customer commitments that Amazon discussed for 2027 and 2028. The company did not say whether those commitments would affect on-demand availability, pricing, or allocation for smaller enterprise customers.
Most of Amazon’s current AI capacity is contracted for terms of at least five years, according to Jassy. Those commitments give the company customer-demand visibility before it purchases equipment, although Amazon did not disclose how the contracts are distributed across customer groups.
Amazon’s disclosures confirm that infrastructure availability is limiting the demand AWS can currently serve. They do not establish that capacity has become a more important competitive factor than pricing across the wider cloud market.
Jassy said Amazon starts spending on data centres about two years before the facilities open. The company therefore incurs construction and equipment costs before customers can use the new capacity and before the facilities begin generating revenue.
Expanding cloud capacity requires sites, electricity connections, buildings, servers, and networking equipment. The development timeline means a higher capital expenditure budget does not immediately translate into additional AWS capacity.
A US Department of Energy advisory report said grid-connection requests for hyperscale facilities requiring between 300 MW and 1,000 MW or more can involve lead times of one to three years. It also identified constraints involving transformers, switchgear, transmission infrastructure, and other equipment required to supply large data centres.
The International Energy Agency expects global electricity consumption from data centres to double between 2025 and 2030, with electricity use by AI-focused facilities expected to triple. It also identified power equipment, advanced chips, and other component constraints as factors limiting how quickly new capacity can enter service.
AI workloads also require more than access to accelerators. Large training and inference clusters depend on memory, storage, cooling, power distribution, and high-bandwidth networking.
Amazon said software and process improvements, custom chips, and internally developed network equipment were helping it use server and network capacity more efficiently. The company recorded $53.1 billion in cash capital expenditure during the second quarter, primarily related to AWS and generative AI.
Memory and infrastructure costs increase spending
Jassy said higher memory costs were the main reason Amazon increased its 2026 capital expenditure forecast by approximately $20 billion. Amazon did not disclose how the higher prices would affect the amount of computing capacity delivered by that spending.
Memory is used throughout AI servers. High-bandwidth memory supplies data to accelerators during training and inference, while conventional server memory supports the wider computing system.
Amazon did not specify whether the increase related mainly to high-bandwidth memory, conventional server memory, storage, or a combination of those categories. It also did not provide a breakdown of spending on processors, networking, storage, cooling, and electrical infrastructure.
Amazon said its AI and semiconductor businesses had each exceeded an annualised revenue run rate of $25 billion. AWS reached an annualised revenue run rate of $169 billion during the quarter.
Jassy said AI adoption was also increasing customers’ consumption of core AWS services. These include computing, storage, databases, and networking used to support AI applications and their associated data.
AWS operating income rose to $16.6 billion from $10.2 billion a year earlier. The division recorded a 39% operating margin during the quarter, alongside the 36.7% increase in revenue.
Capital expenditure weighs on free cash flow
Amazon recorded a trailing 12-month free cash flow outflow of $7.6 billion, compared with an inflow of $18.2 billion a year earlier. Purchases of property and equipment increased by $66.1 billion over the same period, primarily reflecting AI investment.
Amazon’s trailing 12-month operating cash flow rose 33% to $161.4 billion. The increase in property and equipment purchases was nevertheless large enough to move free cash flow into negative territory.
The gap reflects the timing of Amazon’s infrastructure programme. Data centre construction begins about two years before opening, while servers and networking equipment are generally purchased several months before they enter service.
Jassy said Amazon has stronger visibility into customer requirements before purchasing servers and networking equipment because those assets are installed closer to deployment. He said Amazon would not make those equipment purchases if the corresponding customer demand was absent.
Jassy said data centres can generate revenue for more than 30 years without requiring Amazon to repeat the initial construction expenditure. Servers and networking equipment operate on shorter investment cycles.
On average, servers and networking equipment take slightly less than three years to reach break-even, Jassy said. He added that the equipment has a useful life of at least five to six years, leaving another two to three years to generate free cash flow after the initial investment has been recovered.
(Photo by Martin Woortman)
See also: Microsoft expands Azure AI with Databricks and Mistral
Want to learn more about Cloud Computing from industry leaders? Check out Cyber Security & Cloud Expo taking place in Amsterdam, California, and London. The comprehensive event is part of TechEx and is co-located with other leading technology events, click here for more information.
Cloud Computing News is powered by TechForge Media. Explore other upcoming enterprise technology events and webinars here.