SAN FRANCISCO, California — AI infrastructure company Modal has reported that NVIDIA's new Vera central processing unit delivered up to twice the task throughput of a comparable Intel-based server in early testing, as the chipmaker expands into a market traditionally dominated by Intel and AMD.
The results come as Modal prepares to introduce Vera processors into its Sandboxes platform, which provides isolated computing environments for artificial intelligence agents and other automated workloads. The company said NVIDIA delivered its first Vera CPU server to Modal's San Francisco office the previous week.
Modal's findings suggest the processor could handle a greater number of simultaneous AI-related tasks without substantially increasing the time needed to complete individual jobs. The performance gains were most pronounced when the server was operating under heavy demand, rather than when tasks were run individually.
The announcement gives NVIDIA another opportunity to demonstrate the capabilities of its server processors beyond the graphics processing units that have driven much of its growth in AI computing.
Modal developed a benchmark to evaluate how different processors handle the workloads generated by AI agents, which can execute software commands, search files, compile code and interact with computing environments while completing assigned tasks.
The company tested Vera against processors based on Intel and AMD architectures, as well as Amazon's Graviton chips. Its testing framework, called Agentless Agent Bench, uses 50 tasks adapted from Terminal-Bench, a benchmark designed around software development and command-line activities.
Rather than measuring only how quickly a processor completes one operation, Modal examined what happens when many tasks run simultaneously. The company increased the number of active containers on each machine and measured how many jobs could be completed within a fixed period.
Initial tests involving individual tasks showed a narrower performance difference. Vera completed many operations faster than competing processors, but it did not consistently outperform every alternative. Modal reported a more substantial advantage when it increased the number of concurrent workloads.
Under those conditions, Vera achieved up to twice the task throughput of a comparable Intel Xeon 6 server based on the Granite Rapids architecture, according to Modal. The company said individual task completion times remained broadly comparable even as the workload increased.
Much of the investment in AI computing has centered on graphics processing units, which perform the intensive calculations required to train and run large models. Yet the systems supporting AI applications also depend on conventional processors to manage software, coordinate tasks and handle operations that do not require a GPU.
AI agents introduce additional demands because their workloads can be difficult to predict. An automated coding assistant, for example, might spend time waiting for a language model to respond before suddenly executing commands that place substantial pressure on the CPU.
Modal's Sandboxes platform is designed to run these activities in isolated environments. Multiple customers and applications can share underlying computing resources, allowing infrastructure providers to use server capacity more efficiently.
That arrangement also creates competition for resources. A sudden increase in activity from one application can affect the performance of other workloads running on the same machine, particularly when the server approaches its processing limits.
Modal said Vera's performance under heavy demand could allow more workloads to operate on a single server without the same degree of slowdown observed in its tests of alternative processors.
The potential financial implications extend beyond the cost of CPU processing. In some AI training workflows, expensive GPUs may remain idle while supporting software tasks finish. Reducing those delays could improve the use of existing computing capacity.
NVIDIA introduced Vera in May as a processor designed for AI agents, reinforcement learning and data processing. The chip uses an Arm-compatible architecture and can operate in standalone servers or alongside NVIDIA's Rubin graphics processors.
The company said at the launch that Vera could complete certain tasks 1.8 times faster than x86 processors.
Vera also places NVIDIA in more direct competition with Intel and AMD, whose processors have long been central to enterprise servers and cloud computing infrastructure. Amazon's Graviton family represents another alternative, reflecting growing interest in Arm-based chips for data-center applications.
NVIDIA has identified several technology companies and cloud providers planning to adopt Vera, including Oracle Cloud Infrastructure and CoreWeave. Hardware manufacturers such as Dell Technologies, Hewlett Packard Enterprise, Lenovo and Supermicro are also developing systems incorporating the processor.
The expansion gives NVIDIA a way to address more of the computing infrastructure surrounding AI applications. While GPUs remain essential to many model-training and inference workloads, CPUs perform the supporting operations needed to keep those systems running.
Modal's testing offers an early example of how NVIDIA's processor strategy could translate into demand from specialized AI infrastructure providers. Its significance will depend partly on whether the reported advantages remain consistent when Vera is deployed across larger production environments.
Adopting Vera would require changes to Modal's existing systems because the company has historically operated on x86 processors. Moving to an Arm-compatible architecture involves adapting software infrastructure and image-building processes to support a different instruction set.
Modal said it intends to introduce Vera CPU support for Sandboxes and work with NVIDIA and cloud providers to expand the available computing capacity. The announcement describes a planned rollout rather than a completed migration of the company's infrastructure.
The startup's interest reflects the growing importance of infrastructure designed specifically for AI agents. Unlike traditional applications with relatively predictable resource requirements, agent-driven systems may execute different combinations of tasks depending on the instructions they receive.
Modal has also expanded its Sandboxes offering beyond lightweight containers. On October 1, the company announced the general availability of virtual machine Sandboxes, providing AI developers with isolated Linux environments capable of running more complex software operations.
The Vera tests build on that infrastructure strategy by examining how additional processing capacity could support large numbers of simultaneous automated tasks.
