Following its recent $15.75B valuation, Modal is making its first venture into Arm-based computing by integrating Nvidia's Vera CPU into its infrastructure. The shift marks a significant departure from the x86 architecture that has dominated the company's operations to date.
Modal's Sandboxes product—which generated $300M in annual revenue at the time of the company's latest funding round—will run on Vera chips. According to tests conducted by Modal itself, the Arm-based processor handled up to twice as many containerized tasks as comparable Intel servers, though these results remain unverified by independent parties.
The transition to Vera represents Modal's first move away from x86 chips manufactured by Intel and AMD. Since Vera operates on Arm's instruction set architecture, Modal will need to overhaul its infrastructure and rebuild its software image compilation systems. Nvidia delivered Modal's first Vera server to the company's San Francisco headquarters last week.
Why Vera matters for sandbox workloads
Sandboxes function as isolated execution environments where AI agents write and run code. These workloads present unique scheduling challenges: agents frequently alternate between idle periods—waiting for model inference—and resource-intensive bursts that can demand ten times the baseline compute. With hundreds of sandbox instances running concurrently on each server, one customer's peak usage can degrade performance for others.
Modal's benchmark testing involved 50 coding tasks spanning compilation and file searching operations. When executed sequentially, Vera matched or exceeded performance from Intel, AMD, and Amazon chips, though gains were modest and inconsistent. The performance gap widened substantially when Modal densely packed the server with containers: Vera maintained consistent task completion rates while competing processors reached saturation limits. Modal attributes this advantage to Vera's superior memory bandwidth.
The throughput improvements prove particularly valuable for two scenarios, according to Modal: background coding agents and reinforcement learning workloads. In reinforcement learning pipelines, expensive GPU resources frequently remain underutilized while waiting for CPU-bound tasks to complete.
The broader Vera ecosystem
Nvidia introduced Vera in March as a purpose-built CPU for AI agent workloads, featuring 88 cores. The company has already secured commitments from Anthropic and OpenAI. Perplexity has announced plans to deploy agent workloads on the chip after reporting 1.5 times faster coding task performance. Meanwhile, competitors are expanding their own sandbox capabilities, with Baseten acquiring Blaxel in September.
Modal has not disclosed a timeline for when Vera-powered sandboxes will become available to its customer base.
Source: The Next Web
