NVIDIA published a November 5, 2026 update describing Vera CPU performance for AI-agent workloads. Its argument centers on execution capacity: an agent may generate a plan on an accelerator, then depend on CPU resources to run code, search, or manage an isolated workspace.
What the reported tests cover
NVIDIA reports a data-center comparison in which Vera supported 1.6 times more agents per watt than the fastest x86 configuration in its comparison. It also cites partner tests covering coding, orchestration, databases, and sandbox provisioning. These are vendor and partner results under particular test conditions.
Daytona’s own October 7 report provides supporting context rather than a new November announcement. It describes bringing 1,000 sandboxes online in 10.8 seconds on Vera, compared with 32.3 and 88.4 seconds on two specified AMD EPYC systems. The company separately measures completed jobs under concurrent load. Startup time and completed-job throughput are different metrics.
Measure the work between model responses
Our interpretation is that an agent benchmark should include the parts of a task that a user actually waits for. Fast generation alone cannot describe the experience if an environment takes a long time to start or a test suite waits for execution resources. A sensible evaluation records each interval and the final task outcome.
Concurrency deserves the same care. A lightly loaded machine can hide contention that appears when many workspaces start together. Teams should test the expected busy period, record how long the slowest tasks take, and check whether those tasks still finish correctly. A large number of active sessions is useful only when they can make progress.
Keep comparisons reproducible
An evaluation should identify the processor configuration, memory, operating system, sandbox image, task set, and concurrency. If a model service is part of the test, hold its version and request pattern steady. Otherwise, a change elsewhere in the system can be mistaken for a processor improvement.
Power comparisons also need a clearly defined boundary. Decide whether the measurement covers the CPU, the server, or a broader facility estimate. Record that choice beside the result so the number can be compared with another setup honestly.
What the November update changes
The update broadens the discussion of agent infrastructure beyond model serving. It offers concrete vendor and partner measurements for teams preparing evaluations. Those measurements are useful starting points; a production decision still needs a representative workload, a complete cost picture, and evidence that the selected system sustains the required response times.