Modal to run sandboxes on Nvidia’s Vera CPU

Modal announced on October 19 that it will begin running its AI sandbox product on Nvidia’s Arm‑based Vera CPU, its first move away from the x86 chips it has used since its founding. The company said it received a Vera server in San Francisco last week and will need to rework its infrastructure and image‑building pipeline to support the new architecture. Vera is part of Nvidia’s AI‑centric hardware lineup and uses Arm design, meaning Modal will have to adjust its build system to work with a different instruction set.

Benchmarking Highlights

In its own tests, Modal reported that a Vera‑powered server could handle up to twice as many concurrent tasks as a comparable Intel server when the workload was packed with Docker containers. The company built a benchmark of 50 coding tasks—such as compiling code, searching files, and performing Git operations—and found that, under heavy load, Vera maintained throughput while rivals such as Intel’s GNR, AMD’s Zen 5 and Amazon’s Graviton 4 and 5 hit early saturation. Both sources attribute Vera’s advantage to its high memory bandwidth and a design that allows more container workloads to fit on each core.

The Vera chip, launched at Nvidia’s March 2026 GTC conference, contains 88 custom Olympus cores that each can run two tasks through a feature called Spatial Multithreading. Its low‑power memory subsystem is based on LPDDR5X and can deliver up to 1.2 TB/s of bandwidth, allowing the system to sustain more than 22,500 concurrent CPU environments in a single rack. Modal will need to adjust its scheduler and image‑building process to accommodate the new architecture, and the move also affects how the platform handles noisy‑neighbor and burst‑capacity problems, as hundreds of sandbox instances share a single host.