Modal Says Vera Doubles Sandbox Throughput Under Load

Two results in one post

Modal's post contains two findings, and they point in different directions. The first is the quiet one. Run one task at a time on an empty machine and Vera is, in Modal's words, often better than the other processors tested, but not always, and not by a particularly large margin. The post's author writes that he was initially skeptical of the results. That is an unusual sentence to find in a vendor-adjacent launch post, and it is the reason the second finding deserves a hearing.

The second finding appears when the machine fills up. Modal's throughput test keeps adding concurrent containers and counts verified task completions in a fixed time. There, the company reports up to double the throughput of the x86 Granite Rapids machine it chose for comparison, and says the higher throughput did not require worse individual task latency in the configurations it tested. The rival processors plateaued early as concurrency rose. Vera, on one benchmark node, did not.

The distinction is the whole story. Nobody at Modal is claiming your job runs twice as fast. The claim is about a host running hundreds of strangers' sandboxes at once finishing more of them before contention slows everything down — completed tasks per host, on machines whose core counts differ: 88 physical cores for Vera, 128 for the Intel GNR machine, 96 for Graviton. It is not a per-core, per-watt or per-dollar measurement. Modal attributes the under-load behaviour to Vera's memory bandwidth, which it says lets significantly more container workloads pack onto each core in its tests. NVIDIA's stated figure for that memory subsystem is up to 1.2 TB/s; a specification is not by itself proof of what caused the gap.

One more boundary on the benchmark itself. Agentless Agent Bench was built to isolate the CPU work agents actually do — compilation, file grepping, git operations, device emulation — with network calls stripped out and containers prewarmed. That makes it informative about host-level contention. It does not measure cold starts, full agent-session latency or real-world network delays, and 50 tasks is a small stand-in for production traffic.

Why a CPU result matters for agents

Sandboxes are where agents stop talking and start working: running code, installing packages, searching files. Modal names its two dominant uses as background agents and reinforcement-learning rollouts. The second one prices CPU time in GPU money. Most training setups are still synchronous, so a fan-out of CPU rollouts has to finish before idle GPUs can take the next step. A CPU bottleneck there is not a CPU problem. It is a GPU bill for waiting. Modal's benchmark did not measure an end-to-end training-speed improvement, and it publishes no cost-per-completed-task comparison; whether the throughput result lowers anyone's bill is a separate, unanswered question.

That is also the workload Vera was designed for. NVIDIA launched the chip in March as a processor purpose-built for agentic AI and reinforcement learning, and it has told investors it expects at least $20 billion in Vera sales this fiscal year against a market it sizes at roughly $200 billion. Modal is one name on a longer list NVIDIA has already announced, which includes OpenAI, Anthropic, Oracle and Perplexity — announced plans to use the chips, not confirmed production deployments. What Modal adds is not another logo. It is a published benchmark, with the method and the benchmark code public, from a buyer whose entire fleet was x86 until this week.

The discounts to apply

Four, specifically. First, the benchmark is Modal's own, built for this evaluation, run on early hardware — a single Vera node the company says NVIDIA hand-delivered the week before. "Up to double" is the best result, not the average. Second, the comparison set is the Intel and AMD chips already on the market. Modal's chief executive, Erik Bernhardsson, told Reuters both rivals have newer models in development that could narrow the gap, and that Modal will test them when they arrive. Third, this is a plan, not a deployment: Modal says it intends to roll out Vera support and scale the fleet with NVIDIA and cloud providers, and that supporting an entirely different CPU architecture is a huge change across its infrastructure and image build systems. No date, fleet share or pricing was given. Fourth, the same CEO was explicit that Intel and AMD are not being replaced. Vera enters a mixed fleet, and each job goes to whatever hardware runs it best.

None of that cancels the result. A second sandbox operator, Daytona, published its own Vera tests on 7 October against AMD's EPYC Zen 5 processors and found the same direction under load: Vera sustained about 13 completed agent jobs per second at 2,000 concurrent sandboxes, ahead of both EPYC chips it ran. Daytona is also an interested party selling concurrency, and two self-run tests under different protocols are not independent verification. But it is a pattern worth watching: the operators who sell concurrency are the ones finding the gap.

What to watch

Whether Modal puts a date and a fleet share on the rollout, and whether the throughput gap survives contact with production traffic rather than a benchmark built from 50 tasks. The next useful data point is not another launch post. It is an Intel or AMD response tested the same way, or a Modal follow-up once customers are running on Vera hosts.

Sources

[1] Modal (primary), "NVIDIA Vera CPUs double throughput for Modal Sandboxes", company post, 19 October 2026 (published 14:01 UTC) — benchmark method and results, serial vs under-load findings, Arm migration scope, hand-delivered first server, rollout intentionRead source
[2] Reuters, Stephen Nellis, "Nvidia wins fast-growing AI startup Modal for new central processors", 19 October 2026 (published 14:12 UTC) — mixed-fleet plan, Bernhardsson on testing against chips already on the market and on rivals' newer models, NVIDIA's $20B Vera expectation and ~$200B market sizing, Modal funding and growth figuresRead source
[3] Unite.AI, "Modal Plans NVIDIA Vera Rollout for Sandboxes in First Arm Move", 19 October 2026 — synthesis of the Modal post (benchmark design, core-count boundaries 88/96/128, use cases) and NVIDIA's 16 March 2026 Vera launch specifications (88 Olympus cores, Spatial Multithreading, up to 1.2 TB/s LPDDR5X)Read source
[4] Daytona, "When 2,000 Agents Share One CPU", company post, 7 October 2026 — Daytona's own Vera tests against AMD EPYC Zen 5: Vera sustained about 13 completed agent jobs per second at 2,000 concurrent sandboxes, ahead of the EPYC 9755 and 9575F; self-run by an interested operatorRead source
[5] ALFW, "Modal Crossed $1 Billion in Run Rate. That Is a Pace, Not a Year.", 16 October 2026 (our earlier coverage of Modal's Series D and run rate, cited for context only)Read source

THE ALFW VERDICT

Modal's early tests suggest Vera could be particularly valuable for sandboxes running many tasks at once: in its own 50-task benchmark Vera often — not consistently — outperformed rival processors on individual tasks, and under heavier contention Modal reported up to twice the task throughput of an Intel Granite Rapids server, with comparable task latency in the configurations it tested. That is a host-throughput result, not a 2x speedup for every agent job, and it establishes no cost-per-task advantage — the tests used early hardware, prewarmed containers and a benchmark designed by the prospective buyer, against currently available rivals whose newer designs remain untested. Modal's willingness to undertake an Arm migration it calls a huge change, while keeping its existing suppliers in a mixed fleet, is a meaningful adoption signal, but the production rollout and its economics remain unproven.