Amazon Web Services has instructed its internal engineering teams to reduce CPU waste in EC2 instances as demand for compute capacity surges, according to a report from The Information. The directive, delivered in a May meeting, reflects a growing strain on Amazon's data center resources driven by the rise of agentic AI workloads. Engineers now report wait times of days for virtual machines that previously took hours to provision.

What You Need to Know

AWS engineers face days-long waits for EC2 instances that formerly required hours to spin up. Agentic AI applications demand more balanced CPU and GPU usage, shifting away from traditional GPU-heavy ratios. The internal scarcity signals that external customer demand for EC2 compute is rising, which could lead to tighter allocation or higher costs for users of Amazon's cloud.

The CPU Crunch Behind the Crackdown

EC2 instances form a significant portion of the modern internet and are deployed extensively in private environments. Historically, AWS engineers had the flexibility to spin up instances for development work, taking advantage of relatively low CPU utilization in web infrastructure. That ease has evaporated. One engineer told The Information the current wait times are unprecedented in their years at Amazon.

  • Engineer access: AWS engineers must now compete for CPU cycles that were previously plentiful for internal use.
  • CPU-GPU shift: The traditional 8:1 or 4:1 ratio of GPUs to CPUs is approaching parity, amplifying overall CPU demand.
  • Processor diversity: Amazon relies on AMD, Intel and its own Graviton5 chips — all in high demand for EC2 customers.

Amazon Web Services deploys a range of processors in EC2 instances, including offerings from AMD and Intel as well as its homegrown Graviton5 chip. That Arm-based CPU represents Amazon's most powerful in-house design and competes with Nvidia's Vera CPU and Arm's own data center designs.

Agentic AI Reshapes Infrastructure Demands

The CPU strain is closely tied to the rise of agentic artificial intelligence. Traditional AI inference workloads rely heavily on GPUs, with CPUs functioning mainly as a support layer to feed the accelerators. Agentic workloads, however, are more complex. They involve tool calls that execute on CPUs and orchestration logic that requires additional compute outside the GPU pipeline. This structural change has elevated CPUs from a supporting role to a central resource.

The scale of the demand is visible in recent incidents at Amazon. One coding agent blew through $1.8 million in token costs, exceeding its budget by 860%. Such episodes underscore the resource intensity of agentic AI and its impact on cloud capacity.

An updated internal directive from AWS makes clear that the company must prioritize customer workloads over internal development. The Information report noted that an updated version of the message reinforced the need for efficiency. CPU demand now rivals GPU demand in many deployments, forcing a rethinking of infrastructure planning.

Why This Matters

The crackdown at Amazon Web Services is a leading indicator for the entire cloud industry. If a company with AWS's scale is struggling to manage internal CPU consumption, smaller providers will face even greater challenges. For AWS customers, the tightening supply of EC2 instances could translate into longer provisioning times, higher spot pricing or stricter limits on instance types. The shift also accelerates the market for CPUs: AMD, Intel and Nvidia are all competing aggressively for data center sockets, with AMD launching its Zen 6 'Venice' series and Nvidia pushing its Vera CPU. The era of cheap, abundant CPU cycles in the cloud may be ending, replaced by a new equilibrium where every core counts.