Inference.(2)

All For CTOs For Executives For Policymakers For Researchers

Agent infrastructure Agent payments Identity and mail Machines and edge Inference Cloud market Operations

19 Oct 2026 · 5 min

CPUs are the agent bottleneck

An agent spends most of its time running tools, sandboxes and browsers, not generating tokens. The measurements and the chip makers now agree, and the CPU-to-GPU ratio in data centres is moving with them.

CTOsResearchersAgent infrastructureInference

19 Oct 2026 · 6 min

Speculation pays on structure, not prose

Our measurements show speculative decoding speeding up SQL by 3.5x and code by 1.5x on one GPU, and doing nothing for prose. The useful decision is whether to speculate at all, and memory bandwidth sets the ceiling on everything else.

ResearchersCTOsInference