NVIDIA publishes Vera CPU results claiming more AI agents per watt

More AI agents could run within the same electricity budget without sacrificing speed, according to NVIDIA’s latest Vera CPU results. Published November 5, 2026, the findings compare its processor with selected x86 chips in a 100-megawatt data-center comparison. NVIDIA also cited partner tests covering coding, databases and isolated environments for running agent software.

Two comparisons, not one universal multiplier

The argument starts with work agents perform between model responses. NVIDIA describes GPUs handling reasoning while CPUs execute searches, database queries and code tests. Its pitch addresses two demands: finishing those steps quickly and supporting many agents within a facility’s power limit.

These are company findings, not an independent verdict on competing CPUs. NVIDIA’s summary does not name the x86 models or provide full test configurations, limiting how precisely operators can apply the results to hardware choices.

Partner tests measure different pieces of agent work

NVIDIA also attributed these production-style test results to early adopters:

• Perplexity ran agentic coding tasks 1.5 times faster than on traditional CPUs and plans to deploy Vera in an upcoming production system.
• DeepInfra achieved up to 2.2 times the speed of its best x86 processor on agent orchestration—coordinating agent work—and led four tested architectures.
• Daytona started 1,000 agent sandboxes, isolated software environments, in 10.8 seconds versus 88.4 seconds with x86. That measures startup, not completion of an entire assignment.
• ClickHouse’s standard software scaled near-linearly across 176 cores without Vera-specific tuning and topped the ClickBench hardware leaderboard.

Modal’s earlier test explains the crowded-server effect

In its October 19 benchmark post, Modal reported a twofold sandbox throughput increase on Vera. Its benchmark used 50 Terminal-Bench solutions and verifiers, removing network calls and prewarming concurrent Docker containers to isolate CPU work.

On an unloaded machine running tasks serially, Modal found Vera often performed better, but not always and not by a particularly large margin. Under concurrent load, its results changed substantially. Modal describes how sudden CPU demand from one customer’s workload can slow another on a shared server.

Adoption would carry an implementation cost. Modal said supporting Arm-compatible Vera would require infrastructure and image-building changes after exclusively using x86 machines. Its post evaluated an option, not a completed migration.