Nvidia has won a conspicuous foothold in a part of the data center it has never owned: the general-purpose CPU. AI cloud provider Modal, which until now ran its workloads only on Intel and AMD processors, plans to adopt Nvidia's Arm-based Vera CPU to run AI agents, Reuters reported on October 20. The decision follows early tests in which Vera delivered up to twice the task throughput of an equivalent Intel Granite Rapids system when many workloads ran at once.
Modal's business makes that distinction important. The company provides isolated sandboxes where AI coding agents can execute shell commands, compile software, search files, manipulate repositories and run other tools. Those environments are unpredictable, spending long stretches idle before suddenly demanding CPU resources, and Modal's economics depend on packing as many of them as possible onto each server without one customer's work slowing another's.
To measure that, Modal built the Agentless Agent Bench, a set of 50 tasks drawn from Terminal-Bench covering compilation, file searching, Git operations and device emulation, with network calls removed to isolate CPU work. Its throughput test does not simply time a single job; it counts how many verified tasks a system can complete in a fixed period as concurrent workloads are added. Under that contention, Vera's advantage grew as the machine got busier.
Modal attributes much of the result to memory bandwidth. Vera offers up to 1.2 TB/s of LPDDR5X bandwidth, compared with approximately 844 GB/s for Intel Granite Rapids and 576 GB/s for AMD's Zen 5-based EPYC platform in Modal's comparison — roughly 42 percent more than Granite Rapids and more than twice the EPYC figure. Modal says the extra bandwidth lets Vera accommodate more concurrent sandbox workloads per core without the performance degradation seen on competing systems.
The caveats matter as much as the headline number. On individual tasks run on an otherwise quiet server, Vera's advantage was smaller and more mixed, and it did not win every test. Modal also developed its benchmark with close contact with Nvidia, including visits to Nvidia headquarters, so these are a prospective customer's own workload-specific results, not independent third-party benchmarks. Vera is also the newest chip in the comparison, and it is specifically optimized for the low-latency, high-concurrency work associated with AI agents and reinforcement learning, while Granite Rapids and EPYC are broader general-purpose server CPUs.
Modal is not abandoning x86. Chief executive Erik Bernhardsson said the company will keep its Intel and AMD hardware and assign jobs to whichever processors perform best, adding that both rivals are developing newer chips that could narrow the gap and that Modal will test them when they are available. That hedge is well timed: AMD's next-generation EPYC Venice processors are arriving with up to 1.6 TB/s of memory bandwidth, and Intel has confirmed its Diamond Rapids Xeons for 2027 with a 16-channel memory configuration expected to deliver a substantial increase over Granite Rapids.
The commercial stakes explain the attention. Nvidia reportedly expects at least 20 billion dollars in Vera sales in its first fiscal year, targeting a CPU market it values at approximately 200 billion dollars — a target that, for scale, would exceed the 16.6 billion dollars AMD's entire Data Center segment generated in fiscal 2025. Modal, which recently raised 750 million dollars at a 15.75 billion dollar valuation, gives Nvidia a reference customer built entirely around agent workloads. If other AI-native clouds see the same throughput under contention, the CPU socket — long a two-vendor contest — becomes a three-way fight.