Amazon Web Services (AWS) asked internal engineers to conserve EC2 CPU capacity, The Information reported on August 7, 2026. According to the report, AWS stepped up efforts to reclaim idle instances and right-size them to match actual usage in order to secure computing resources for customers. AWS denies that this was a directive prompted by new capacity constraints. However, CEO Andy Jassy stated explicitly in his 2025 shareholder letter that AWS has capacity constraints leading to unmet demand. The internal conservation request comes as the AI-era scramble for computing resources spreads from GPUs to general-purpose CPUs.
The Reported May Request and AWS's Rebuttal
According to The Information, AWS executives met with engineers in May and told them to "conserve capacity by any means necessary" in order to ensure sufficient capacity for EC2 customers going forward. This applies to both servers equipped with AI chips and those equipped with CPUs. Following the report, some internal engineers said the wait time to secure CPU servers has grown longer than before.
In its response to Tom's Hardware, AWS explained that reclaiming idle instances, right-sizing, and scaling up or down based on demand are practices that have been ongoing for some time. It stated that treating such efficiency measures as evidence of new capacity constraints would be a mistake, and when asked whether there had been a time-limited reduction request, it countered that the narrative around EC2 is inaccurate. In other words, what AWS denied was not the efficiency measures themselves, but the causal claim that they stemmed from a special directive triggered by new supply difficulties.
The Unmet Demand AWS Itself Has Acknowledged
Amazon itself has not claimed that supply capacity is abundant. In his 2025 shareholder letter, Jassy laid out a plan for AWS to add 3.9GW of power capacity in 2025 and to double its total power capacity by the end of 2027. At the same time, he acknowledged that there is demand AWS is failing to capture due to capacity constraints. He noted that two large customers had requested to purchase AWS's entire 2026 Graviton instance capacity, but AWS could not accommodate this because it also needs to supply other customers.
Demand has continued to grow since then. AWS revenue for the second quarter of 2026 rose 37% year-over-year to $42.2 billion, putting it on an annualized run rate of $169 billion. Both the AI business and the chip business—including Graviton, Trainium, and Nitro—exceeded an annualized $25 billion each. Capital expenditures over the trailing twelve months increased by $66.1 billion year-over-year, and Amazon attributes this increase primarily to AI investment. Even so, it takes time to convert power, buildings, and servers into revenue-generating capacity. In his shareholder letter, Jassy stated that AWS's capital expenditures are required 6 to 24 months before billing begins.
The PR rebuttal and the shareholder letter address different things. The former denies the explanation that links this particular internal efficiency drive to newly established capacity measures. The latter, meanwhile, acknowledges that AWS as a whole faces supply constraints and that extreme large-scale demand has been directed at Graviton. There is no evidence to definitively conclude that this particular wait time is a result of AI demand. Even so, the backdrop that raises the value of internally reclaimable CPU capacity can be confirmed from Amazon's own public disclosures.
Why AI Agents Want CPUs
AI's computing demand does not end with GPU or Trainium workloads. Agents execute code before and after inference, call search functions and APIs, and run multiple work environments in parallel. While accelerators handle model computation, CPUs continue to manage the control and data processing of each environment. AWS markets Graviton5 as a CPU built for agentic AI—one that keeps accelerators running while processing many concurrent execution environments.
Graviton5, which became generally available on June 10, 2026, packs 192 cores into a single chip. According to Amazon's published figures, it delivers up to 25% higher compute performance than Graviton4, with 33% lower inter-core latency. More than 120,000 customers use Graviton, and Meta has signed a contract to deploy tens of millions of cores for agentic AI. New demand from AI agent execution environments is now layering on top of the CPU demand that has long come from web servers and databases.
In this structure, simply increasing the number of GPUs does not complete the overall processing pipeline. Along with the CPUs and memory that invoke inference, networking and storage need to be provisioned at matching speed. The existence of large customers seeking all of Graviton's capacity plainly shows that general-purpose computing resources can no longer be treated as a supporting player to AI infrastructure.
A System That Doesn't Label Under-5% Usage as "Waste"
AWS Compute Optimizer analyzes up to 93 days of data and proposes right-sizing by looking not just at CPU but also memory, network, and disk. The criteria for idle candidates are also a composite condition: peak CPU usage below 5% over 14 days, combined with network I/O below 5MB per day. The reason low utilization alone doesn't trigger a shutdown decision is to reclaim capacity while still protecting against sudden load spikes and preserving availability.
Spot Instances, offered to external customers, similarly provide surplus EC2 capacity at up to 90% off on-demand pricing. If regular demand increases or supply decreases, AWS can reclaim instances, typically with two minutes' notice, and availability is not guaranteed. Spare capacity inherently comes with the property of being offered cheaply in exchange for the ability to reclaim it when needed elsewhere. In this internal review as well, how much capacity can be reclaimed without halting development environments will determine its practical effectiveness.
AWS has set a goal of doubling power capacity by the end of 2027. Until then, whether internal wait times for securing capacity improve will serve as an early indicator of whether the pace of capacity expansion is keeping up with both AI and ordinary cloud demand. If reclaimed idle CPUs can be redirected to customer demand without slowing down development, then efficiency measures can effectively buy time until new infrastructure comes online.