Modal Plans NVIDIA Vera Rollout for Sandboxes in First Arm Move

Modal said in a company blog post on October 19, 2026 that it plans to roll out support for NVIDIA’s Arm-based Vera CPU to power Modal Sandboxes, the company’s first departure from an all-x86 fleet, after benchmark results it published showed up to double the task throughput of an equivalent x86 machine under load.

NVIDIA hand-delivered Modal’s first Vera CPU server to the company’s San Francisco office in the week before the announcement, Modal said. Vera is NVIDIA’s Arm-compatible processor, built for the Vera Rubin GPU platform but also available as a standalone CPU server. Bringing Arm into the fleet would require changes across Modal’s infrastructure and image build system, which the company described as a huge change after a long history of running exclusively x86 machines.

Multi-Tenant Noise and Burst Capacity

Orchestrating a Sandbox product means running hundreds of concurrent containers and virtual machines on each cloud host, and Modal said the resulting multi-tenant clusters are noisy. Resource usage can be extremely spiky: a sandbox may sit mostly idle while waiting on a language model and then see CPU usage surge 10x above baseline because of an npm install or an expensive grep command. The shape of each workload is only known at runtime, so resource sizing is often a guess, and heterogeneous tenants packed into a single pool of compute create noisy-neighbor problems, where one customer’s behavior degrades another’s performance.

Rather than require fixed reservations sized to maximum usage, Modal Sandboxes support bursting, which lets users reserve guaranteed capacity for their P50 usage and trust that padding inside the host node will absorb transient spikes. Modal said the approach has saved users millions of dollars, but acknowledged that it raises exposure to the noisy-neighbor effects that strict limits would otherwise avoid.

Modal identified two prevalent Sandbox use cases: background agents, citing Ramp Inspect and agents running on Niteshift, and reinforcement-learning rollouts. Most training setups are still synchronous, the company said, so an entire fanout of CPU-bound rollouts must complete before idle GPUs can move to the next step, which means time saved on CPU is time saved on GPU. For background agents, the company said, faster execution means faster answers and lower CPU time, driving down the cost of goods sold.

The Agentless Agent Bench

To measure Vera, Modal built and released Agentless Agent Bench, a collection of 50 Terminal-Bench oracle solutions and verifiers with all network calls stripped out, prewarmed in concurrent Docker containers. Modal said building an agentic benchmark is difficult because most of the work done on its sandbox servers is not CPU-bound, so it assembled a concurrent and heterogeneous mix of the tasks agents and reinforcement-learning processes actually perform on the CPU, including compilation, file grepping, git operations and device emulation. According to the public repository, the benchmark is built on the terminal-bench-2.1 task dataset with Dockerfiles ported for Arm compatibility, and a run counts as completed only if the task verifier passes.

In serial testing on an unloaded machine, which measures straight-line per-task latency, Modal found Vera was often faster than the other processors tested — Intel GNR, AMD Zen 5, and AWS Graviton 4 and 5 — but not always, and not by a particularly large margin, and the post’s author described being initially skeptical of the results.

Under load, Modal reported, the results diverged. The throughput test increases the number of concurrent containers on the machine and loops as many verified task completions as possible in a fixed time. Modal said its early benchmarking runs showed up to double the task throughput on Vera compared with an equivalent x86 GNR machine, while individual task latency measured equal to or lower than on x86. In most systems, throughput and latency trade off as packing density rises until the CPU saturates, but Modal said Vera did not slow down as jobs accumulated and kept completing tasks at its original pace.

Modal reported that Vera kept improving as logical CPU counts rose, while the other tested processors hit early throughput plateaus. The physical core boundaries of the tested machines were 128 for Intel GNR, 96 for Graviton and the GNR virtual machine, and 88 for Vera. Because the alternatives plateaued early, Modal said its results suggest it could push Vera even further than the single benchmark node it had access to.

The company attributes Vera’s advantage under contention to its high memory bandwidth, which it said allows significantly more container workloads to pack onto each core without CPU throttling. For Modal’s scheduler, that means greater packing density and more forgiving burst protection; for Sandbox users, the company said, it means higher absolute scale and throughput without sacrificing latency.

Vera’s Design and Production Status

NVIDIA launched Vera on March 16, 2026 at its GTC conference, describing it as a processor purpose-built for agentic AI and reinforcement learning. The chip has 88 custom NVIDIA-designed Olympus cores, each able to run two tasks through a feature NVIDIA calls Spatial Multithreading, and a second-generation low-power memory subsystem built on LPDDR5X that delivers up to 1.2 TB/s of bandwidth. NVIDIA claims Vera delivers results with twice the efficiency and 50% faster than traditional CPUs.

At the launch, NVIDIA also introduced a Vera CPU rack integrating 256 liquid-cooled CPUs, which it said can sustain more than 22,500 concurrent CPU environments, each running independently at full performance. NVIDIA said Vera is in full production, with partner availability in the second half of 2026, and named Alibaba Cloud, ByteDance, CoreWeave, Lambda, Nebius, Nscale and Oracle Cloud Infrastructure among the cloud providers planning to deploy the CPU.

Modal said it intends to work closely with NVIDIA and cloud providers to scale up its Vera fleet as the rollout proceeds.