Amazon Web Services internally told engineers in May to reduce EC2 CPU waste, restricting instance access after years of abundant capacity. Engineers who previously spun up instances within hours now wait days, a stark signal of CPU constraint. The root cause: as agentic AI workloads shift from GPU-bound training to CPU-intensive orchestration—tool calls, API management, memory coordination, conditional logic—the traditional 4:1 to 8:1 GPU-to-CPU ratio has collapsed toward parity. One engineer told The Information they had never waited this long for an instance in years of employment at Amazon.
AMD and Intel have both confirmed severe CPU scarcity. Intel warned Chinese customers of six-month lead times; AMD informed major customers its entire 2026 server CPU capacity is already allocated. The demand stems from agentic AI's architecture: unlike training (GPU-centric), agents orchestrate sequential tasks, branch on conditions, call external functions, and manage state—all CPU work. AMD CEO Lisa Su stated on Q1 earnings that agentic adoption is generating CPU demand far exceeding projections, with the GPU-to-CPU ratio tightening toward 1:1 in some scenarios. Even Amazon, which tripled annual CPU server purchases, cannot keep pace.
For architects, this signals a structural repricing of compute: CPUs, treated as commodity plumbing for decades, are now bottleneck-critical. Spot and flexible capacity are sold out; only contracted long-term capacity survives. Nvidia has entered the x86 CPU market with Vera; OpenAI reportedly ported its entire codebase to ARM for Graviton. This is not cyclical scarcity but a permanent shift in how agentic systems partition work between orchestration (CPU) and inference (GPU/ASIC). Organizations locking in CPU capacity now avoid a 2H2026 supply crunch.