Modal Will Move Agent Sandboxes Onto NVIDIA's Vera CPU -- Its Own Benchmark Shows Little Gain on a Quiet Host and a Lead Only When Containers Pile Up

On October 19, 2026, Modal engineer Erik Dunteman wrote that NVIDIA had "hand-delivered Modal's first Vera CPU server" the previous week, and that Modal will roll out Vera to power Modal Sandboxes, its isolated environments for coding agents, RL rollouts and untrusted code. Vera is NVIDIA's Arm-compatible CPU, built for the Vera Rubin platform but also sold as a standalone server. Supporting it means changes "across our infrastructure and image build system", because Modal comes "from a long history exclusively running x86 machines."

Why sandboxes stress a CPU differently

The post explains why agent sandboxes are an awkward scheduling problem. A sandbox can sit idle waiting on an LLM, then spike "10x above baseline, due to an npm install or expensive grep command", and its shape is only known at runtime. Modal lets users reserve capacity for P50 usage and burst above it, which saves money but raises noisy-neighbour risk. CPU time also costs more than the CPU: for background agents a slow sandbox delays the answer, and in synchronous RL training "there are GPUs sitting idle while sandboxed processes run."

The benchmark

Modal built "Agentless Agent Bench": 50 Terminal-Bench oracle solutions and verifiers with network calls removed, prewarmed in concurrent Docker containers, so only the CPU work agents do -- compiling, grepping, git, device emulation -- is measured. Run serially on an idle machine, the result was modest: "Vera often performed better than the other CPUs, but not always, and not by a particularly large margin." Under load the picture changed: "Counterintuitively, Vera gets better under heavy load." Modal credits Vera's memory bandwidth for letting more containers pack per core without throttling, while the other platforms plateau.

The chart compares Vera with AMD Zen 5, Intel Granite Rapids (bare metal and VM) and AWS Graviton 4 and 5 VMs. Our reading of its embedded data: Vera's highest value is 6,537 verified completions per 60 seconds at 176 logical CPUs, against 4,601 for bare-metal Granite Rapids at 155 -- about 1.4x -- while the Graviton VMs top out near 3,600 at 96 logical CPUs. The 2x figure in the title therefore depends on which comparison is chosen, and the VM results are not like-for-like with bare metal.

The commercial frame

Reuters, via The Economic Times, reported that Modal raised $750 million last week at a $15.75 billion valuation and that NVIDIA expects at least $20 billion from Vera sales this fiscal year. Chief executive Erik Bernhardsson stressed that Intel and AMD stay in a mixed fleet, so Modal "can always deploy whatever hardware performs best," and said of the rivals' newer chips: "We'll test them when they're available." The same day in Berlin, NVIDIA signed a joint declaration of intent with the German government on physical AI models ahead of its GTC Berlin keynote on October 21.

What remains uncertain. This is one customer's benchmark on one pre-production server, with no prices, power figures or server counts, and Modal measured current x86 parts, not the next generation. Arm porting costs and image-build changes are acknowledged but not sized.

Analysis: The finding that matters is the shape, not the multiplier: on a quiet host Vera is roughly a peer, and its advantage appears under contention. That fits agent workloads, which are bursty and memory-bound rather than compute-bound, and it suggests the right sandbox benchmark is throughput at realistic packing density, not single-task latency. Teams running their own agent or RL rollout fleets can reuse the method -- strip network calls from real task traces and sweep concurrency -- before deciding whether a CPU switch is worth an architecture port.