NVIDIA pitches Vera CPU for agentic fleets
NVIDIA published telemetry from 163,594 agentic sessions to argue that agent fleets are bottlenecked by single-thread latency rather than core count, and that its forthcoming Vera CPU is designed for that operating point.
In a developer-blog post, NVIDIA makes the case that CPU choice — not just GPU capacity — sets the economics of "AI factories" running agentic workloads, where CPUs handle orchestration, tool execution, and sandboxed computation.
The supporting data is the more interesting part. NVIDIA says telemetry from 163,594 agentic sessions showed over 97% with unique trajectory profiles, which it argues makes right-sizing a fleet across multiple specialized CPU designs impractical. It characterizes trajectories by length (reasoning steps, tool calls, retries) and width (concurrent tool calls, retrieval, sandboxes, sub-agents), and reports that sessions spend most wall-clock time on a strictly latency-bound sequential chain, with only transient parallel bursts. A cited real-world Claude Code session ran 33 minutes, with sub-agent fan-out punctuating a long dependency chain during which the main agent idles.
From that, NVIDIA argues the optimization target is completed user sessions rather than raw core count, and that high-core-count parts trade away the single-thread performance the critical path needs. It also flags a memory cost to core-parking strategies: up to 8 GB per core left unused.
The performance claim is thinner. NVIDIA cites estimated SPEC CPU 2026 results — compiler, static analysis, and Python benchmarks — showing Vera, built on its Olympus cores, delivering "up to 1.5x the agentic performance of the latest competition." The results are vendor-estimated rather than independently verified, and the post gives no absolute scores, configurations, or availability date.