Modal says NVIDIA hand-delivered its first Vera CPU server to the company and that it was running in a rack. In early tests, Modal reported up to 2× task throughput on Vera compared with an equivalent x86 GNR system under concurrent load. The company describes Vera support for Modal Sandboxes as a planned rollout.
The server has arrived; Sandbox support is planned
Modal says the server was running in a rack after NVIDIA delivered it. That is a step in Modal’s Vera work, while support for Vera-powered Sandboxes remains a plan. Modal says it needs to adapt its infrastructure and image-build system for Arm, and plans to continue using x86 CPUs alongside Vera.
What Modal measured
Modal built Agentless Agent Bench from 50 Terminal-Bench oracle solutions and verifiers. It removed network calls and ran tasks in prewarmed, concurrent Docker containers. The selected work included compilation, file searches, Git operations, and device emulation.
The benchmark separates two measures. Serial tests on an unloaded machine measured individual task latency; concurrent tests increased the number of containers and measured task throughput. Modal’s result of up to 2× refers to throughput in the concurrent test, not to the time needed to finish one task. Modal says individual task latency under load was equal to or sometimes lower on Vera than on the x86 system.
The comparison is with one x86 GNR system
Modal names an equivalent x86 GNR machine as the comparator for its up-to-2× concurrent-throughput result. That is a workload-specific comparison, not a general result for AMD and Intel CPUs as a class.
Modal attributes its result to Vera’s ability to sustain work under contention and to its memory bandwidth. The company says higher task density could help with bursty, mixed workloads in multi-tenant Sandboxes, where many customers’ tasks share host resources.
NVIDIA’s separate performance claim
NVIDIA makes a different claim in its Vera presentation, saying the CPU completes agentic tasks 80% faster than x86 processors. That is a task-completion claim from NVIDIA, distinct from Modal’s reported task-throughput result. NVIDIA published the video on June 1, 2026.
