Nvidia Vera CPUs Power Modal's AI Agents, Challenging Intel and AMD

A new front in the AI hardware race

Modal’s decision to begin supporting Nvidia’s Vera central processor is a small deployment story with much larger implications: a cloud platform built around AI developer workloads is preparing to add Nvidia CPUs beside the Intel and AMD processors it has used so far [2]. The immediate target is Modal Sandboxes, the isolated execution environments used by AI agents and reinforcement-learning workflows to run code, compile projects, search files, perform Git operations and execute other CPU-side tasks [1]. In practical terms, Nvidia is trying to move from being the default accelerator vendor for AI training and inference into the host CPU layer that schedules, feeds and coordinates those AI workloads [2].

The headline result is simple but important. Modal says early benchmarking showed Vera delivering up to twice the task throughput of an equivalent Intel Granite Rapids x86 system under concurrent sandbox workloads, while maintaining similar or sometimes lower per-task latency [1]. Tech Wire Asia’s account of the same development emphasizes that Modal is not presenting Vera as a universal winner across every CPU task; rather, the strongest result appeared when many agent-style workloads ran at the same time and contention became the bottleneck [3]. That distinction matters because AI agents do not behave like traditional steady-state web services. They may sit idle while waiting for a model response, then suddenly spike CPU demand when an LLM asks them to install dependencies, grep a repository, compile code or emulate a device [3].

Why Modal’s benchmark is different

Modal built its benchmark around the messy reality of agent execution rather than a narrow synthetic CPU test [1]. Its “Agentless Agent Bench” uses 50 tasks derived from Terminal-Bench oracle solutions and verifiers, with network calls stripped out and Docker containers prewarmed so the benchmark isolates CPU-side work agents actually perform [1]. The tasks include compilation, file grepping, Git operations and device emulation, which makes the test closer to the kind of sandbox activity produced by coding agents and reinforcement-learning rollouts than to a generic integer or floating-point benchmark [1].

Modal’s own engineering write-up says the first, quieter test was not the most revealing one [1]. On serial task latency, Vera often did well but did not dominate every task or produce a dramatic advantage across the board [1]. The decisive gap appeared when the company increased concurrent container counts and watched how each CPU behaved under load [1]. Modal reported that Vera kept improving as concurrency rose, while alternative systems hit throughput plateaus earlier [1]. Tech Wire Asia attributes Modal’s explanation partly to Vera’s high memory bandwidth, which helps it hold up when more container workloads share the host [3].

That is why this story is not just “another CPU is faster.” The more interesting claim is that Vera appears better suited to a particular AI-infrastructure pain point: dense, bursty, multi-tenant agent execution [4]. Modal’s sandboxes support bursting, meaning users can reserve capacity around typical usage rather than worst-case spikes, but that model also increases exposure to noisy-neighbor contention when many workloads surge at once [1]. If Vera can let Modal pack more sandbox workloads per host without harming latency, it could improve both user experience and infrastructure economics [1].

The CPU is becoming part of the agentic AI stack

The rise of AI agents has made the CPU more visible in systems that are usually discussed in terms of GPUs. GPUs still run the heavy neural-network computation, but agents spend much of their time calling tools, running code, manipulating files, launching containers and coordinating external processes [3]. Modal specifically identified reinforcement learning as a workload where CPU performance can affect GPU utilization, because synchronous rollouts may leave expensive GPUs waiting for sandboxed processes to finish [3]. In that setting, a faster or more contention-resistant CPU is not merely a cheaper host processor; it can raise utilization across the whole AI system [3].

This is the opening Nvidia wants. Reuters reported that Modal had previously run these hosting workloads only on Intel and AMD chips, and that adopting Vera is an early sign Nvidia’s push into a market long controlled by those rivals has traction [2]. Nvidia’s competitive argument is not simply that it can sell another component into data centers. It is that agentic AI changes the shape of CPU demand, putting a premium on memory bandwidth, concurrency and the ability to maintain per-task performance when many small, irregular jobs compete for the same machine [3].

The strategic significance also comes from where Modal sits in the market. Modal rents compute to AI developers and has seen rapid growth, with Reuters reporting that the company raised $750 million in a Series D round led by Accel at a $15.75 billion valuation and said it was on track for $1 billion in sales over the next year [2]. Reuters also reported that Modal’s service usage had tripled and revenue had grown more than 15 times year over year [2]. A design win at a fast-growing AI infrastructure company gives Nvidia a visible proof point for Vera beyond its own launch claims [2].

Not a clean break from Intel and AMD

Modal is not abandoning Intel and AMD. Reuters reported that CEO Erik Bernhardsson stressed the company will run Nvidia chips alongside its existing Intel and AMD offerings in a mixed fleet, assigning each job to whatever hardware runs it best [2]. He also said Modal would test newer Intel and AMD models as they become available, which keeps the competitive door open rather than declaring Vera the permanent winner [2].

That caveat is important for two reasons. First, Modal’s published benchmark is workload-specific: it tests concurrent agentic sandboxes, not every enterprise CPU application [3]. Second, switching to Vera requires Modal to support an Arm-compatible architecture after a long history of running exclusively on x86 machines [1]. Modal’s engineering blog says that supporting Arm means changes across its infrastructure and image-build system, so the performance upside has to justify real platform work [1]. Tech Wire Asia similarly notes that adopting Vera will require changes to Modal’s infrastructure and container image-building systems [3].

For Intel and AMD, the warning is not that one customer has replaced them overnight. The warning is that AI infrastructure buyers may increasingly evaluate CPUs by agent throughput, sandbox density and GPU-wait reduction rather than by traditional server benchmarks alone [3]. If customers begin to treat CPU choice as a workload-by-workload scheduler decision, incumbency becomes less protective. Modal’s mixed-fleet approach means Nvidia can win slices of demand where Vera performs best without requiring a wholesale migration away from x86 [2].

Why Vera’s architecture matters

Vera is described by Modal as an Arm-compatible CPU built for the Vera Rubin GPU platform but also available as a standalone CPU server [1]. Tech Wire Asia reports that Vera uses Nvidia’s custom Olympus CPU architecture, with 88 cores and an LPDDR5X memory subsystem capable of up to 1.2 terabytes per second of memory bandwidth [3]. The same report says Nvidia positions the architecture around maintaining per-thread performance while supporting concurrent workloads, which aligns closely with Modal’s contention-heavy benchmark results [3].

The memory-bandwidth point is central to Modal’s explanation. Under heavy load, many agent sandboxes may be active at once, and their work can involve unpredictable mixes of file operations, compilation, process scheduling and container activity [1]. Modal says Vera’s “killer feature” is its ability to hold up under contention, allowing significantly more container workloads to pack per core without CPU throttling [1]. If that behavior persists in production rather than only in early tests, it could let Modal offer higher absolute scale and throughput without sacrificing latency [1].

This also shows why AI-agent infrastructure is pulling hardware competition into less familiar territory. Traditional cloud CPU procurement often emphasized broad compatibility, price-performance and general-purpose fleet management. Agentic AI adds a new profile: unpredictable, multi-step, tool-using workloads that may make the CPU a bottleneck even when the most expensive hardware in the rack is the GPU [3]. RuntimeWire framed the issue succinctly: Modal is testing whether Nvidia’s Arm CPU can relieve CPU bottlenecks that slow agents and leave reinforcement-learning GPUs idle [4].

Nvidia’s broader CPU ambition

Reuters reported that Nvidia has told investors it expects at least $20 billion from Vera chip sales this fiscal year and views the relevant CPU market as roughly $200 billion [2]. Those figures underline why this design win matters beyond Modal. Nvidia’s GPU dominance already gives it deep relationships with AI labs, cloud providers and infrastructure buyers. Vera extends that relationship into the CPU socket, where Intel and AMD have historically defined the server roadmap [2].

Modal gives Nvidia a particularly useful reference customer because the company’s workload maps directly onto the “CPU for agents” thesis. Modal’s sandboxes are used for coding agents and similar systems that run long-lived, tool-heavy tasks [1]. Reuters described agents as AI systems that can carry out long-running tasks for human users, which is exactly the category Nvidia is trying to optimize Vera around [2]. If more AI platforms find that agent workloads are limited by host-side throughput and contention, Nvidia can argue that Vera is part of the AI factory rather than a commodity server component [3].

Still, the evidence should be read carefully. Modal’s results are early, the company is still preparing infrastructure support, and the mixed-fleet strategy means workload placement will remain empirical [1]. Intel and AMD are also developing newer chips that Modal has not yet tested, according to Reuters [2]. The current state of the story is therefore not a final CPU market reshuffle, but a credible opening move: Nvidia has won a meaningful AI-infrastructure customer for Vera, Modal has identified a real agent-workload bottleneck, and Intel and AMD now face a competitive test in a part of the AI stack they cannot afford to treat as secondary.

The industry shift

The significance of Modal’s Vera adoption is that it reframes the CPU as an active competitive lever in AI infrastructure. For years, the most visible AI hardware battles centered on accelerators. Modal’s benchmark suggests that as agents become more common, the host CPU’s ability to sustain many irregular tasks under contention may influence cost, latency and GPU utilization [3]. That gives Nvidia a path to challenge Intel and AMD not by replacing every x86 server, but by winning the workloads where agentic AI changes the rules [2].

For Modal, the next test is production reality: making Arm support work across its platform, scaling the fleet with cloud providers and proving that the benchmark advantage translates into customer-visible throughput and latency gains [1]. For Nvidia, Modal is a validation point in a campaign to make Vera part of the default AI data center architecture [2]. For Intel and AMD, it is a signal that the AI era’s CPU contest will be fought not only on cores, clocks and compatibility, but on how well a processor keeps agents moving when hundreds of sandboxes collide on the same host [1].

Developments

1. Nvidia Vera CPUs Will Power Modal’s AI Agents, Competing with Intel and AMDTech Wire Asia · Oct 20, 2026, 6:01 AM · 8/10

Sources from the last 72 hours

1. [1]NVIDIA Vera CPUs double throughput for Modal SandboxesOct 19, 2026, 2:00 AM
2. [2]Nvidia wins fast-growing AI startup Modal for new central processorsOct 19, 2026, 4:11 PM
3. [3]Nvidia Vera CPUs to power Modal’s AI agents, challenging Intel and AMDOct 20, 2026, 2:00 AM
4. [4]Modal says NVIDIA's Vera CPU doubles sandbox throughput under loadOct 19, 2026, 5:47 PM

AI-generated article based on recent web research, then preserved as a dated editorial snapshot.