Modal's first Vera CPU servers recently arrived at its San Francisco office.
Until that shipment every box in the fleet was x86, AMD and Intel, so the move to Arm reached through the infrastructure layer and into the image-build systems that each agent sandbox starts from.
Modal's own word for the port is painstaking.
The number that came out of it is a 2x increase in AI throughput across agentic sandboxes, and the company calls Vera an incredibly compelling option.
Why that number is believable comes down to how they got it.
Agent workloads spike inside a container and spike again across the cluster, so a scheduler has to pick a packing density and accept what latency pays for the choice.
Instead of tuning that by feel, Modal built Agentless Agent Bench, a collection of 50 Terminal-Bench oracle solutions and verifiers, and ran representative sandbox workloads on the new chips.
A smaller team can copy the method without copying the capital spend.
Rent capacity for the week, benchmark the port on the hardware you are actually considering, then commit to the rebuild.
RTX 3090s run at $0.15/hr, and Buy Now Pay Later means zero upfront with payment after one month.
What was the last migration where a benchmark changed your mind before you shipped it?
https://ai.bhkcloud.com/?utm_source=linkedin&utm_medium=social&utm_campaign=bhk_social_2026w43&utm_content=evening
Originally posted on linkedin_personal