Nvidia is positioning its Vera CPU as the host processor for large fleets of AI agents, where database queries, code execution and tool calls alternate with GPU-based model inference. The company says Vera can improve both response speed and the number of agents supported within a fixed power budget.
In Nvidia’s comparison for a 100-megawatt data center, Vera delivered twice the agent speed of the most energy-efficient x86 processor tested. Against the fastest x86 CPU in the comparison, Nvidia says Vera supported 1.6 times more agents per watt. The company did not present the results as an independent benchmark.
Several early users reported production-style tests. Perplexity said agentic coding tasks ran 1.5 times faster than on traditional CPUs, while DeepInfra reported up to 2.2 times the orchestration speed of its best x86 result. Daytona brought 1,000 agent sandboxes online in 10.8 seconds in its test.
The results highlight a less visible agent bottleneck: faster language-model inference does not help if CPU-side tools, sandboxes and databases cannot keep pace.