Modal Adopts Nvidia Vera CPUs After Tests Show 2x Agent Throughput
Internal benchmarks showed the Arm-based processor doubled throughput under heavy container load, though Modal will retain its x86 fleet.
In short
- Modal reported that Nvidia's 88-core Arm-based Vera CPU delivered up to twice the throughput of Intel Xeon and AMD EPYC processors in dense sandbox environments.
- The startup received its first Vera server at its San Francisco office in October 2026, marking its first step away from an exclusively x86 fleet.
- Modal will operate a heterogeneous server architecture, retaining existing AMD and Intel processors because porting all software to Arm immediately is impractical.
- Nvidia projects at least $20 billion in Vera CPU revenue during its current fiscal year, aiming to sell 4 million units.
Cloud infrastructure startup Modal announced on October 19, 2026, that it plans to deploy Nvidia's Arm-based Vera processors across its server fleet after internal benchmark tests showed the chip delivered twice the throughput of existing x86 processors in AI agent workloads. The decision marks the company's first architectural shift away from relying exclusively on Intel and AMD chips.
According to a Modal blog post, Nvidia hand-delivered the company's first Vera server to its San Francisco office in mid-October 2026. The processor is being evaluated specifically to power Modal Sandboxes, isolated execution environments where autonomous software agents execute code, install dependencies, and run commands.
Solving Contention in Multi-Tenant AI Sandboxes
AI agents executing multistep workflows place unique stresses on data center processors. Unlike traditional web services with predictable traffic patterns, sandbox environments run hundreds of concurrent containers on a single host. In these settings, an agent may sit idle for several seconds while awaiting a model completion, followed by an immediate ten-fold surge in CPU usage during tasks such as compilation, file searches, or package installations.
Modal reported that these spiky, heterogeneous workloads frequently trigger noisy-neighbor issues on shared machines. When multiple containers experience simultaneous resource spikes, traditional server processors encounter contention that slows down adjacent tasks. In reinforcement learning pipelines, these CPU bottlenecks become particularly expensive because costly GPU clusters must sit idle while waiting for environment rollouts to complete.
To evaluate server behavior under these conditions, Modal created a testing suite dubbed Agentless Agent Bench, comprising 50 Terminal-Bench oracle tasks and verifiers stripped of network dependencies. In single-task tests running on an unloaded server, Modal found Vera was often faster than comparable chips from Intel, AMD, and Amazon, though not by a wide margin. The performance gap widened substantially once servers were packed with concurrent containers.
Benchmark Performance Under Heavy Contention
As reported by Wccftech, Vera sustained its task completion rate under heavy container packing while rival processors reached throughput ceilings. In Modal's tests against Intel Xeon "Granite Rapids" processors, Vera yielded up to twice the task throughput per core. Against AMD Zen 5 systems running 150 concurrent containers across 124 enabled logical CPUs, Vera demonstrated more than six times the task throughput.
Modal attributed this scaling behavior to Vera's memory subsystem. As noted by Tom's Hardware, Vera pairs 88 custom Arm v9.2 Olympus cores—capable of decoding and executing up to 10 instructions per cycle—with 1.2 TB/s of memory bandwidth. Modal claimed the architecture allows servers to pack more active containers per core without experiencing the CPU throttling typical of standard multi-tenant setups.
However, Tom's Hardware noted several caveats in Modal's disclosed data. Modal did not release detailed system configurations, specific baseline processor SKUs, or performance-per-watt measurements. Additionally, the benchmarks evaluated Vera against currently shipping x86 chips rather than upcoming processor generations planned by AMD and Intel.
Heterogeneous Fleet Strategy
Adopting Vera requires Modal to adapt its infrastructure, operating environment, and container image build pipeline to support the Arm instruction set architecture. Modal Chief Executive Erik Bernhardsson stated that the company cannot port its entire software catalog to Arm overnight and will continue purchasing and running x86 processors from AMD and Intel.
Instead of transitioning entirely to Arm, Modal will maintain a heterogeneous server fleet, routing individual workloads to whichever CPU architecture performs most efficiently, according to Tom's Hardware. The startup also intends to benchmark future generations of x86 processors as AMD and Intel release them.
The deployment comes during rapid commercial growth for Modal. The startup recently closed a $750 million funding round led by Accel at a $15.75 billion valuation, as reported by The Next Web. Modal's sandbox product currently generates $300 million in annual revenue, with total company revenue growing more than 15-fold over the prior year. Modal expects its overall annual revenue to reach $1 billion within the next year.
Nvidia's Push Into Server CPUs
For Nvidia, Modal's adoption represents a notable customer win in its effort to establish standalone data center CPUs alongside its graphics accelerators. As reported by Seeking Alpha, Nvidia unveiled Vera in March 2026 as a processor tailored for agentic AI workloads, both as a component of the Vera Rubin GPU platform and as an independent server offering.
Nvidia has previously secured Vera commitments from major AI organizations including OpenAI, Anthropic, and Perplexity, with Perplexity reporting 1.5 times faster execution on coding tasks, according to The Next Web. Nvidia projects that Vera sales will generate at least $20 billion in revenue during its current fiscal year, targeting shipments of 4 million units into an estimated $200 billion total server processor market.
Modal has not yet disclosed an exact release date for when customers will be able to run production sandbox workloads on Vera hardware.
Questions people ask
Why is Modal testing Nvidia Vera CPUs?
Modal tested Nvidia Vera to evaluate its performance in multi-tenant AI sandboxes, where background coding agents and reinforcement learning rollouts require heavy CPU resources and high memory bandwidth.
What performance increase did Modal observe?
Modal reported up to twice the throughput per core compared to Intel Xeon processors and more than six times the throughput of AMD Zen 5 systems under heavy container contention with 150 concurrent tasks.
Is Modal replacing all of its Intel and AMD processors?
No. Modal plans to maintain a heterogeneous fleet, continuing to run x86 processors from Intel and AMD alongside Vera while evaluating future processor releases from all three vendors.
What architectural modifications does Modal need to make for Vera?
Because Vera uses the Arm v9.2 instruction set architecture rather than x86, Modal must modify its cloud infrastructure and container image build pipeline to support Arm binaries.
Sources
- Modal taps Nvidia Vera CPUs after firm's internal tests show 2x throughput over Intel and AMD chips — AI agentic workflows said to favor Arm chips, but firm will keep x86 chips deployed in heterogene…tomshardware.com, Oct 20, 2026
- Modal says Nvidia’s Vera CPU doubles throughput for its AI sandboxesTNW | Artificial-intelligence, Oct 19, 2026
- NVIDIA Vera CPUs double throughput for Modal SandboxesModal, Oct 19, 2026
- Modal gets first Nvidia Vera CPU server, breaking from Intel and AMDseekingalpha.com, Oct 19, 2026
- Modal Sees A 2x Performance Jump After Moving To NVIDIA’s Vera CPU From Intel & AMD, Says “Vera Gets Better Under Heavy Load”Wccftech, Oct 20, 2026
How this story was made: the newsroom picked it up from Techmeme, Google News and tomshardware.com, gathered the full text of the sources above, and drafted it with AI assistance. Every factual claim was then checked against those sources before publishing (41 claims checked). Illustrations marked as AI-generated are not photographs. Spotted an error? Tell us.
Published October 21, 2026 at 02:08 UTC
