What Modal actually announced
Modal said in an October 19, 2026 blog post that it plans to add support for Nvidia's Arm-based Vera CPU to power Modal Sandboxes. That would be its first departure from an all-x86 fleet. Nvidia hand-delivered Modal's first Vera server to its San Francisco office the week before. Modal describes Vera as an Arm-compatible CPU built for the Vera Rubin GPU platform that is also sold as a standalone server.
Reuters reported that Modal intends to keep using Intel and AMD processors alongside Nvidia's chips, selecting hardware based on workload performance. In Reuters' account, CEO Erik Bernhardsson said the tests ran against Intel and AMD chips already on the market. He added that Modal will test newer models when they arrive.
Customers can't buy this yet. Modal has not announced when customers could use Vera.
What "2x" means
Some coverage of this story implies Vera beat both Intel and AMD by 2x. Modal's own account is narrower.
- The benchmark: Modal built Agentless Agent Bench, a collection of 50 Terminal-Bench oracle solutions and verifiers. The workloads were stripped of network calls and prewarmed in concurrent Docker containers. They isolate CPU-side work such as compilation, file grepping, git operations and device emulation.
- The 2x result: It comes from the concurrency test. Modal found up to twice the task throughput of an equivalent Intel GNR x86 machine under concurrent load. GNR is Granite Rapids.
- The quiet-host result: Modal's serial, unloaded tests were far less dramatic. The company wrote that across the 50 tasks, Vera often performed better than the other CPUs, but not always, and not by a particularly large margin. The serial comparison also included AWS Graviton 4, Graviton 5 and AMD Zen 5.
- Latency: Modal said individual tasks took about the same time or sometimes less time on Vera. In its own words, Vera "gets better under heavy load."
So the strongest claim is that Vera keeps completing more work as containers pile up, while the comparison chips plateau. It is not a claim that every task runs twice as fast.
Why a sandbox host is a good fit for this test
Modal says sandbox hosts run hundreds of concurrent containers and VMs. Agents often sit idle waiting on an LLM, then spike. Modal cites CPU surges 10x above baseline from something like an npm install or an expensive grep.
Modal Sandboxes support bursting. Users reserve guaranteed capacity for typical P50 usage rather than the worst-case P100 moment. That improves utilization, but it also raises exposure to "noisy neighbor" contention. A chip that holds up under heavy packing would let Modal run denser hosts and absorb bursts more comfortably. Modal says this unlocks greater packing density and more forgiving burst protection, and lets it offer higher absolute scale and throughput without sacrificing latency.
Modal names two uses: background coding agents, and reinforcement-learning rollouts. For RL, expensive GPUs often sit idle until the CPU work finishes. The company has not quantified customer-level savings.
Why the memory-bandwidth explanation is unproven
Modal credits Vera's high memory bandwidth. Nvidia's engineering blog gives the design context. Vera has 88 cores, per Nvidia's published material. Nvidia says its SOCAMM2 LPDDR5X memory delivers up to 1.2 TB/s of aggregate bandwidth, or up to 14 GB/s per core. Its Olympus cores add a neural branch predictor and 10-wide decode. Nvidia also says Vera's design partitions resources between two hardware threads to reduce thread-to-thread interference.
Those are vendor design claims. They fit Modal's story, but they don't prove it. Nvidia says Vera delivers up to 1.8x on agentic workloads in SPEC CPU 2026 results it measured internally in July 2026. That is a separate vendor measurement on a different test.
What's missing
- The full test configurations and exact Vera server configuration are not published.
- There is no performance-per-watt or price/performance comparison.
- The results have not been independently verified. They are Modal's internal benchmark findings, not measured gains from customer deployments.
- The test measures CPU-side work, not complete live AI agent sessions.
- The comparison used current x86 parts. Newer Intel and AMD generations could narrow the gap, as Reuters noted.
The test node may also understate Vera. Modal said that in its test the alternatives hit early throughput plateaus while Vera kept improving, suggesting it could push further than the node it had access to. That is Modal's inference, not a measured result.
The Arm porting cost
For admins and developers, the practical story is the cost of changing architecture. Modal said it must change its infrastructure and image build system to support an entirely different CPU architecture. The ISA constraint is also why Modal is going heterogeneous. Anything that assumes x86 binaries, container images or build pipelines needs a multi-architecture path before it can run on Arm. Routing each job to the best-fit hardware only works if the software stack can target both.
This is general industry knowledge, not something Modal detailed. The same pattern applies to teams weighing Arm instances in any cloud. The performance gain has to outweigh the engineering work, and Modal's decision suggests it thinks it does.
Business context
Reuters reported that Modal's adoption brings Nvidia into closer competition with Intel and AMD, whose processors Modal has relied on. Modal recently raised $750M at a $15.75B valuation, per daily.dev's summary. Reuters said Modal reported tripled usage and more than 15x year-over-year revenue growth. Nvidia has told investors it expects at least $20 billion in Vera sales this fiscal year, in a market it values at roughly $200 billion. These are company figures, not audited results.
Vera isn't Modal-only. Nvidia named Alibaba Cloud, ByteDance, CoreWeave, Lambda, Nebius, Nscale and Oracle Cloud Infrastructure among cloud providers planning to deploy it.
Bottom line
This is a credible early signal, not a verdict. Modal built a test that favors dense, bursty, CPU-contended workloads, which is the exact shape of its business. Vera did well on that test against Intel's current Granite Rapids. The gap vanished on a quiet host. Modal plans to keep its x86 machines, and Intel and AMD have newer chips coming. Without published configurations or independent replication, treat "2x" as a promising workload-specific number rather than a general server-CPU ranking.
Update: Additional details (October 20, 2026)
Wccftech reports that Modal’s concurrent-load testing also showed Vera exceeding AMD EPYC throughput by more than six times at 150 concurrent containers. The report says each test system had 124 logical CPUs enabled.
That AMD figure is distinct from Modal’s headline comparison with Intel Granite Rapids: it applies to a specific high-concurrency point, not to unloaded performance or every sandbox task. Wccftech also characterizes the tested configuration as using 124 logical CPUs per system, a detail Modal had not publicly specified in its broader summary.