# Nvidia Vera Doubled Sandbox Throughput Under Load, Modal Says

Published 2026-10-19 · AI.info Team · AI.info

> Nvidia’s Arm-based Vera CPU completed up to twice as many agent-sandbox tasks as an equivalent Intel server in Modal’s early tests. The benchmark measured concurrent coding and reinforcement-learning workloads; Modal has not announced when Vera-powered sandboxes will reach customers.

> “Vera does not slow down as jobs accumulate but rips through each one at the original pace.”
>
> Erik Dunteman, author of Modal’s engineering blog post

[Nvidia](https://ai.info/companies/nvidia)’s Vera CPU delivered up to twice the task throughput of an equivalent Intel server in Modal’s early tests of agent sandboxes under heavy load. The infrastructure company announced on October 19, 2026, that it plans to add support for Vera to Modal Sandboxes, marking its first move beyond an all-x86 fleet.

The claim comes from Modal’s own benchmark, not an independent evaluation. Nvidia hand-delivered the company’s first Vera server to its San Francisco office the previous week, according to Modal. The company still needs to adapt its infrastructure and image-building system for Arm processors, and gave no customer launch date.

## Why Modal tested crowded sandboxes

Agent sandboxes run code generated by AI systems, and their CPU demand can swing sharply. A sandbox may spend time waiting for a model, then spike during tasks such as installing software or searching files. Hundreds of containers and virtual machines can share a cloud host, where one workload’s surge can affect another tenant, Modal says.

Modal’s Sandboxes allow users to reserve capacity for typical usage and burst above that level when demand rises. That arrangement can reduce the need to reserve for peak capacity, the company says, but puts more pressure on the host to absorb simultaneous spikes without slowing down individual jobs.

Those delays also matter beyond the sandbox itself. Modal points to background coding agents, where faster execution can reduce task time and CPU costs, and reinforcement-learning runs, where GPUs in synchronous training setups can wait for sandboxed work to finish before moving to the next step.

## Fifty coding tasks, tested under contention

For its tests, Modal created Agentless Agent Bench from 50 Terminal-Bench solutions and verifiers. The company removed network calls and ran the workloads in prewarmed Docker containers to focus on CPU tasks such as compilation, file searches, Git operations and device emulation.

Modal tested both serial task latency on an otherwise idle machine and throughput as it increased the number of concurrent containers. Vera was often faster in the quiet, one-task-at-a-time tests, but not consistently and not by a large margin, the company’s post says. The larger gap appeared under contention, when many workloads competed for the host at once.

Modal compared Vera with AMD Zen 5, Intel Granite Rapids, and [AWS](https://ai.info/companies/amazon) Graviton 4 and 5 processors, including virtual-machine results for some chips. Vera delivered up to twice the task throughput of an equivalent x86 Granite Rapids machine, Modal reports, while individual task latency was equal to or sometimes lower than on that Intel system. The post does not describe an independent audit of those results.

## Arm support is the next engineering step

Modal attributes Vera’s performance under load to its memory bandwidth, which the company says lets more container workloads share each core without CPU throttling. Its published results show competing processors reaching throughput plateaus as CPU counts rise, while Vera continues to improve. That pattern suggests the chip may have more headroom than the single benchmark server it tested, Modal says.

That benchmark advantage does not mean Vera is already part of Modal’s customer fleet. Supporting Arm requires changes across infrastructure and software-image builds after years of running x86 machines, the company says. Its stated next step is to work with Nvidia and cloud providers to scale up the Vera fleet; for now, the twofold figure describes an early, company-run test rather than a generally available service.

## Sources

1. Original source · Blog post [Modal](https://modal.com/blog/nvidia-vera-sandboxes) modal.com

## Companies in this story

- [Modal Labs](https://ai.info/companies/modal-labs)

## Related stories

- [Modal Raises $750 Million at a $15.75 Billion Valuation](https://ai.info/news/modal-raises-750-million-series-d-15-75-billion-valuation)
- [NVIDIA Launches RTX Spark Laptops for Local AI and Gaming](https://ai.info/news/nvidia-launches-rtx-spark-laptops-local-ai-gaming)
- [NVIDIA Opens Nemotron Models for Always-On AI Agents](https://ai.info/news/nvidia-nemotron-models-always-on-ai-agents)
- [MiniMax Publishes M3.1 Flash Weights Under Community License](https://ai.info/news/minimax-publishes-m31-flash-weights)
- [NVIDIA’s Nemotron 3.5 Super VL Targets Agentic Workflows](https://ai.info/news/nvidia-nemotron-3-5-super-vl-agentic-workflows)

## Explore

[More articles](https://ai.info/news)

---

Canonical: https://ai.info/news/modal-nvidia-vera-doubles-sandbox-throughput
