Modal to Run Sandboxes on NVIDIA Vera CPUs After Early Tests

Modal said on 19 October 2026 that it plans to roll out support for NVIDIA Vera CPUs to power Modal Sandboxes — its first move beyond an all-x86 fleet — after early internal benchmarks showed up to double the task throughput versus an equivalent x86 Granite Rapids machine. For developers running agents and reinforcement-learning rollouts in Modal sandboxes, that could mean denser packing and less noisy-neighbour pain — if the company’s own tests hold up in production.

No public rollout date appears in the blog. The throughput figures are Modal’s early benchmarking, not an independent lab result. On unloaded, serial runs, Modal’s author noted Vera “often performed better than the other CPUs, but not always, and not by a particularly large margin.”

Why Modal cares about contention

Sandboxes host hundreds of concurrent containers and VMs per host. Workloads spike (an npm install or heavy grep after idle LLM waits), shapes are unknown until runtime, and bursting saves customers money while increasing noisy-neighbour risk. Modal also notes that sandbox CPU time often gates expensive external work: background agents waiting for answers, or GPUs idle until RL rollouts finish.

Vera is NVIDIA’s new Arm-compatible CPU, built for the Vera Rubin GPU platform but also sold as a standalone CPU server. Supporting it means infrastructure and image-build changes across Modal’s stack after a long x86-only history. NVIDIA hand-delivered Modal’s first Vera server to its San Francisco office, the post says.

What Modal’s own benchmarks say — and what they do not

Modal built Agentless Agent Bench: 50 Terminal-Bench oracle solutions and verifiers, stripped of network calls and pre-warmed in concurrent Docker containers, to stress compilation, grepping, git and similar agent/RL CPU work.

On a quiet machine running tasks serially, Vera’s per-task latency was often better than other CPUs, but not always, and not by a large margin — Modal’s own caveat. Under concurrent load, the company says early runs showed up to double the task throughput on Vera versus an equivalent x86 GNR machine, with individual task latency equal to or sometimes lower than x86. Modal attributes that to high memory bandwidth holding up under contention so more containers pack per core without throttling.

Those are Modal’s early tests on hardware it had for benchmarking. They are not third-party audits, and rivals’ next CPU generations are not in the comparison set.

Mixed fleet, no ship date

Reuters reported (Stephen Nellis) that Modal will run the NVIDIA chips alongside its current Intel and AMD offerings in a mixed fleet. CEO Erik Bernhardsson said a mixed setup means the company “can always deploy whatever hardware performs best,” and that newer Intel and AMD parts will be tested when available.

Modal’s blog says it is “very excited to roll out support for Vera CPUs to power Modal Sandboxes” and to work with NVIDIA and cloud providers to scale the fleet. It does not name a general-availability date for customers.

Our read

Vera’s pitch for Modal is contention tolerance in multi-tenant sandboxes, not a headline single-thread win — which is why the “not always” unloaded caveat matters as much as the 2× throughput line. The mixed-fleet stance, as Reuters frames it, keeps Intel and AMD in the scheduler. Watch for a customer-facing Vera region or SKU; until then this is an architecture commitment backed by vendor-site benches, not a dated launch.

Sources

1. Modal: “NVIDIA Vera CPUs double throughput for Modal Sandboxes” (Erik Dunteman, 19 October 2026).
2. Reuters / Investing.com: “Nvidia wins fast-growing AI startup Modal for new central processors” (Stephen Nellis, 19 October 2026).

Related stories

• Modal Raises $750 Million Series D at a $15.75 Billion Valuation
• DuckDB Posts Unaudited TPC-H Results for v2.0-dev on a Single NVIDIA Vera
• NVIDIA Opens Vera Rubin Rack Designs to OCP