Technology · NEWS EXPLAINER
Modal plans NVIDIA Vera rollout for AI agent sandboxes after early throughput test
The cloud startup says its own loaded benchmark reached up to twice the task throughput of an equivalent x86 system, but it will keep Intel and AMD hardware in a mixed fleet.
Modal plans to add NVIDIA’s Arm-compatible Vera CPU to the infrastructure behind its AI agent sandboxes, marking its first move away from an exclusively x86 fleet. Reuters independently confirmed the planned adoption on October 19.
The headline performance number needs a boundary: Modal says its early, concurrent sandbox benchmark reached up to twice the completed-task throughput of an equivalent x86 Granite Rapids machine. The company will still test newer rival processors and keep a mixed fleet.
Modal is adding Vera, not abandoning x86
Modal says it plans to roll out support for NVIDIA Vera CPUs to power its Sandboxes. Reuters reports that the cloud startup currently relies on Intel and AMD CPUs and intends to run Vera alongside those offerings, assigning work to the hardware that fits it best.
The two-times figure comes from Modal’s own early load test
Modal built Agentless Agent Bench from 50 verified Terminal-Bench tasks with network calls removed. In early loaded runs, it reported up to twice the completed-task throughput of an equivalent x86 Granite Rapids machine, while individual task latency was equal or lower. This is a company benchmark for a specific concurrent sandbox workload, not a universal CPU comparison.
Contention matters more than an idle speed test
Modal says an unloaded serial test showed Vera was often faster than the alternatives it tested, but not always and not by a large margin. The larger gap appeared when many containers competed for resources, which the company attributes to Vera’s memory bandwidth and ability to maintain throughput as jobs accumulate.
Arm support requires infrastructure work
Vera is Arm-compatible, while Modal has historically operated an all-x86 fleet. Modal says the rollout requires changes across its infrastructure and image-build system. That means the announcement is a deployment plan supported by early testing, not proof that every Modal workload has already moved to Vera.
Why agent sandboxes are a strategic CPU workload
Agent sandboxes run compilation, code search, package installation, tool calls and other CPU work around AI models. NVIDIA positions Vera as a standalone or host CPU for such tasks. Modal says faster sandbox execution can shorten agent response times and reduce periods when GPUs wait for reinforcement-learning rollouts to finish.
