NVIDIA Reports Vera CPU Agent-Processing Performance, Comparing Speed and Power Efficiency With x86

NVIDIA released comparison results on Nov. 5, 2026, for its Vera CPU against x86 CPUs in a hypothetical 100-megawatt data center. The company said Vera processed 1.6 times as many agents per watt as the fastest x86 CPU and processed agents twice as fast as the most energy-efficient x86 CPU. The comparison assumed agents moved between inference on GPUs and execution on CPUs, and showed processing speed and performance per unit of power.

In NVIDIA's comparison, Vera CPU demonstrated both higher throughput per watt than the fastest x86 CPU and greater speed than the most energy-efficient x86 CPU for agent processing. The examples show that in agent platforms involving CPU-intensive stages such as search, code execution and database processing, CPU performance and power efficiency need to be evaluated separately from GPU performance.

The focus is less on a general ranking of standalone CPU performance than on balancing capacity and speed for the CPU execution stages of agent workloads in power-constrained AI infrastructure. NVIDIA says agents do more than run inference on GPUs: They also search, query databases, write and test code, and perform operations, with CPU processing in between. The figures for throughput per watt and speed in the 100-megawatt comparison are notable because they evaluate these execution stages. However, the figures were released by NVIDIA and cannot be treated as the results of third-party verification under standardized configurations and power-measurement conditions. A related article from Physical AI News, NVIDIA Explains Full-Stack Design to Address AI Memory Constraints, examined pressure on GPU memory from larger models and longer context windows. In the same AI infrastructure, NVIDIA's Vera announcement focuses on CPU execution capacity. Addressing memory constraints and improving CPU processing capacity are separate design challenges, and the performance results for either one alone cannot evaluate the overall processing capacity of an agent. Similarly, the factory-agent design covered in NVIDIA Introduces Agent Collaboration Design to Investigate Factory Production Problems is not evidence that the CPU benchmark will directly deliver benefits in deployment. The broader industry implication is that evaluating agent platforms requires looking beyond GPU inference performance to include CPU-based orchestration and sandbox execution. The results cited by NVIDIA also cover different types of workloads, including code processing, databases and sandboxes. For example, ClickHouse's scaling at 176 cores and Daytona's startup time and completed jobs are different metrics and should not be combined into a single measure of “agent performance.” By presenting multiple workload examples, NVIDIA appears to be positioning Vera to handle diverse CPU workloads with one product rather than focusing on a specific compute task. Next, it will be important to examine each test's CPU configuration, actual power-measurement conditions, comparison settings and workload fluctuations in real-world operations. NVIDIA's published material does not provide detailed conditions or pricing that would allow these tests to be compared across the board. The company also mentioned plans to deploy Vera in Perplexity's production system, but the reported test results need to be assessed separately from processing throughput and power efficiency in real-world operations.

In the tests cited by NVIDIA, Perplexity ran agentic coding tasks 1.5 times faster than on previous-generation CPUs. DeepInfra said Vera was the fastest of the four architectures it tested for its production agent workloads, reporting speeds up to 2.2 times those of the best x86 processor. ClickHouse said its unoptimized, standard binary scaled nearly linearly to 176 cores on Vera and ranked No. 1 in the ClickBench hardware rankings. In Daytona's test, 1,000 sandboxes started in 10.8 seconds, and 2,000 concurrent sandboxes completed 12.7 jobs per second. NVIDIA said it took 88.4 seconds to start 1,000 sandboxes on x86. Modal reported up to twice the sandbox throughput per core compared with equivalent x86 hardware, while DuckLabs reported about 1.5 times the median performance on TPC-H SF1000, using Intel Xeon 6 as the baseline.