Amazon told its own engineers in May to stop wasting CPU cycles. The directive, delivered in internal meetings and reported by The Information, ordered EC2 users to reduce low-utilization instances to free capacity for paying customers. Provisioning time has stretched from hours to days. One multi-year Amazon engineer told The Information they had never waited this long for a compute instance.

The proximate cause: agentic AI workloads are consuming CPU at a rate the infrastructure was never designed to handle.

Traditional AI inference used one CPU for every eight GPUs. The CPU tokenized inputs, routed requests, and kept GPU queues fed—light work. Agentic workloads break that model. Each agent loop—plan, call tool, parse response, re-tokenize, route to the next sub-agent—executes almost entirely on CPU. A Georgia Tech and Intel paper from November 2025 measured it: tool processing on CPUs accounts for 50 to 90 percent of total latency in agentic workloads.

Agentic agent loop: steps 1 and 3–7 run on CPU; only LLM inference (step 2) uses GPU. Tool processing accounts for 50–90% of total latency.
FIG. 02 Agentic agent loop: steps 1 and 3–7 run on CPU; only LLM inference (step 2) uses GPU. Tool processing accounts for 50–90% of total latency. — Georgia Tech and Intel, November 2025

Intel's CFO David Zinsner confirmed on the Q1 2026 earnings call that the CPU-to-GPU ratio in data centers had moved from 1:8 to 1:4. He said agentic deployments could push it to 1:1. JP Morgan's analysts modeled an ideal ratio of seven CPUs per GPU.

ScenarioCPU:GPU RatioCPUs per GPUSource
Traditional AI inference1:80.125Industry baseline
Current data center shift (Q1 2026)1:40.25Intel CFO David Zinsner, Q1 2026 earnings
Agentic deployment projection1:11Intel CFO David Zinsner, Q1 2026 earnings
JP Morgan ideal agentic ratio7:17JP Morgan analyst model
FIG. 03 CPU-to-GPU ratio evolution in data centers: from traditional inference to agentic AI deployments — Intel Q1 2026 earnings call; JP Morgan analyst research

The practical arithmetic is brutal. On a standard 8-GPU node with 64 vCPUs, a GRPO rollout job calling a code-execution sandbox saturates all CPU cores at roughly 32 concurrent workers. GPUs sit idle waiting for scored trajectories. A 10-agent LangGraph or CrewAI pipeline making three tool calls per step needs approximately 30 simultaneous CPU threads per inference round. Spheron's infrastructure team recommends 16–24 vCPUs per GPU for pipelines running ten or more concurrent agents. Most existing EC2 instances are undersized by that ratio.

WorkloadSetupCPU Saturation PointRecommended vCPUs/GPU
GRPO rollout w/ code-execution sandbox8-GPU node, 64 vCPUsAll cores saturated at ~32 concurrent workers—
10-agent LangGraph / CrewAI pipeline3 tool calls per step~30 CPU threads per inference round—
Pipelines running 10+ concurrent agentsGeneral recommendation—16–24 vCPUs per GPU
FIG. 04 CPU requirements for specific agentic workloads on standard 8-GPU nodes — Spheron infrastructure team; article body

Demand has outrun supply. Dylan Patel at SemiAnalysis reported that Amazon and Microsoft have sold out their entire CPU inventory to AI firms including OpenAI and Anthropic. Amazon tripled its CPU server count year-over-year and still exhausted supply. OpenAI ported its entire codebase to ARM to access AWS Graviton instances—Graviton5, built on TSMC N3 with 192 cores, is currently Amazon's highest-density option.

Server CPU prices have risen 20 percent since March; average selling prices were up 27 percent in Q1 2026. Lead times for high-core-count server CPUs have stretched to six months. Intel confirmed Xeon demand exceeds supply by a "meaningful" margin and has shifted wafer capacity from consumer chips to data center production.

Server CPU price appreciation in 2026: spot prices +20% since March; average selling prices +27% YoY in Q1 2026.
FIG. 05 Server CPU price appreciation in 2026: spot prices +20% since March; average selling prices +27% YoY in Q1 2026. — Tom's Hardware / SemiAnalysis reporting, Q1 2026

The shortage is not uniform. Amazon's capacity crunch is concentrated in spot instances. The Information's sources confirmed that contracted reserved capacity has not experienced shortages. Architects who locked in reserved instance agreements before Q1 are insulated; teams on on-demand or spot are competing for what remains.

Cost risk compounds capacity risk. A coding agent inside Amazon blew through $1.8 million in token costs last month, exceeding its development budget by 860 percent. Agentic loops that spin unbounded sub-agents saturate CPU and generate token volume that overwhelms cost controls.

The takeaway for architects: CPU is the new GPU in agentic deployments. Reserved capacity is the only reliable capacity. Any agent loop without explicit concurrency limits and CPU-to-GPU ratio planning is a latency and cost incident waiting to happen.