The Modal post is by Erik Dunteman and dated Oct 19, 2026. The page says that last week NVIDIA handed Modal its first Vera CPU server at the San Francisco office. That is one machine already in hand, not a fleet already deployed. Vera is described as a new Arm-compatible CPU, built for the Vera Rubin GPU platform and also offered as a standalone CPU server.
[1]
The test is named Agentless Agent Bench. The page says it collected 50 Terminal-Bench reference solutions and checks, stripped out every network call, and prewarmed them in concurrent containers, so the run would show the work actually done on the CPU. On a quiet host with no load, running one task after another, Vera was often better than the other CPUs across those 50 tasks, but not on every task, and not by a particularly large margin.
[1]The result under load carries two limits. The page says early runs showed up to double the task throughput on Vera compared with an equivalent x86 GNR machine, while latency on a single task measured equal to that x86 machine, or sometimes lower. Up to double is not a fixed multiple on every task. The early runs are also not yet a service users can use. The page ends by saying they are eager to put Vera CPUs behind Modal Sandboxes and to add machines with NVIDIA and cloud providers. Eagerness is not a launch.
[1]