Modal, a cloud infrastructure platform for AI workloads, has reported promising performance results from testing NVIDIA's new Vera CPUs, demonstrating up to twice the task throughput compared with an equivalent x86 server. Announced on October 19, 2026, the findings highlight the processor's ability to handle large numbers of concurrent workloads without significantly increasing latency. The results could support Modal's plans to introduce Arm-based infrastructure for its Sandbox environment, which currently operates on x86 machines.
NVIDIA Vera CPU Performance
NVIDIA Vera is an Arm-compatible processor developed for the Vera Rubin computing platform and also available for standalone server deployments. Modal recently received its first Vera CPU server at its San Francisco office, allowing the company to evaluate the hardware against existing infrastructure. The testing represents a potential architectural shift for Modal, requiring modifications to its infrastructure and container image-building systems to accommodate Arm processors.
The company's initial evaluations showed that Vera performed competitively across individual computing tasks, although its advantages were not consistently substantial. More significant improvements emerged when multiple workloads were executed simultaneously, reflecting the conditions typically encountered in shared cloud environments. Under these conditions, Vera delivered up to twice the throughput of a comparable Intel Granite Rapids x86 system while maintaining similar or sometimes lower task latency.
Addressing Infrastructure Challenges
Modal Sandboxes support applications that execute code in isolated environments, including AI agents and reinforcement learning systems. These workloads often experience unpredictable demand, with brief periods of intensive processing followed by extended intervals of limited activity. When hundreds of containers share the same physical server, sudden demand increases can create resource contention and affect overall performance.
To manage these fluctuations, Modal allows workloads to temporarily exceed their guaranteed CPU allocations instead of requiring customers to reserve capacity for maximum potential usage. This approach helps reduce unnecessary resource reservations but also exposes workloads to interference from other applications operating on the same server. Vera's ability to sustain performance under heavier concurrent demand could make this resource allocation model more efficient.
Benchmarking Agentic AI Workloads
To assess the processor's capabilities, Modal developed Agentless Agent Bench, a benchmarking framework based on 50 Terminal-Bench tasks and their verification procedures. The framework removes network dependencies and executes workloads inside prewarmed Docker containers, allowing researchers to isolate CPU-intensive operations commonly performed by AI agents. These operations include software compilation, file searching, Git commands, and device emulation.
The benchmark evaluates both individual task completion times and the number of tasks completed under increasing levels of concurrency. Modal observed that competing processors reached throughput plateaus earlier as additional containers were introduced, while Vera continued delivering higher aggregate performance. The company attributed this advantage largely to Vera's high memory bandwidth, which helps accommodate more simultaneous workloads without substantial performance degradation.
Implications for AI Applications
The findings could have implications for organizations operating autonomous AI agents and reinforcement learning infrastructure at scale. Faster CPU execution may reduce response times for background agents while lowering the computing resources required to complete their tasks. In reinforcement learning environments, improved CPU performance could also reduce GPU idle time when training systems must wait for groups of computational tasks to finish.
For Modal, these improvements could translate into greater server utilization, higher workload capacity, and more predictable performance during demand spikes. However, the reported results reflect early internal benchmarks rather than independently verified improvements across all customer workloads. The company has not announced a specific timeline for making Vera-powered Sandboxes generally available.
Modal's early testing suggests NVIDIA Vera CPUs could offer meaningful infrastructure advantages for cloud platforms supporting increasingly complex AI workloads. By combining higher concurrent throughput with relatively stable latency, the processors may help address resource contention challenges associated with large-scale agent execution. Modal plans to continue working with NVIDIA and cloud infrastructure providers to introduce Vera support and expand its computing fleet.