Cloud Arbitrage for Agent Swarms

Problem

How to Scale Globally with a Small Team?

Meso AI's platform monitors the social media world for major record labels and global brands. The workload is bursty by design: a quiet baseline can explode into a viral moment in hours, and customers expect the platform to be ready when it happens.

Under the hood, Meso runs a dynamically scaling, massively parallel multi-agent architecture — many independent workloads, each performing its own data-processing task against a high-volume, high-velocity, multi-source data stream. The demand isn't uniform: different workloads need very different scaling models, from real-time on-demand processing to steady daily batches to always-on services.

On top of that, Meso wanted cloud flexibility — committed-use compute on AWS, fresh discounts on GCP, and they had a sense the math would keep changing. Doing this on raw Kubernetes across both clouds would have meant two infrastructure stacks, no shared identity, and at least one full-time engineer pulled off the product.

Solution

The Right Cloud and Scaling for Each Workload

Meso first migrated from EC2 fleets on AWS to Control Plane regions on AWS. A few months later, they moved from AWS to GCP regions on Control Plane with no major complications. Meso's environment runs on a Bring Your Own Kubernetes location: Control Plane manages the Kubernetes layer inside Meso's own GCP project, preserving committed-use pricing and native networking while keeping everything portable to any other cloud. On top of that platform, two choices get made per workload — how it scales, and where it runs. Both are push-button. Both can change without rebuilding anything.

“Beautifully” Handled Scaling Profiles

Meso's architecture spans many independent workloads, each with its own scaling profile — high-priority work that scales on demand, a steady daily baseline, dynamic workloads that spin up for peak load and wind down after, and a reserve held ready for surges. Always-on services run active-passive so updates roll through without downtime, and request-driven services scale with demand. Every workload gets the model that fits its shape.

Hyperscale Clouds without the Lock-in

Today, Meso runs primarily on GCP, but their AWS EKS cluster is connected via Control Plane's Bring-Your-Own-Kubernetes to their virtual cloud, making A/B testing of workloads on alternate clouds a push-button exercise. That enables Meso to optimize its processing across varying cloud pricing and capability landscapes. That same routing flexibility extends to regions: data can be sourced from Japan, Italy, Germany, or wherever a customer needs. Doing this by hand would mean maintaining Kubernetes on both clouds and a dedicated engineer; with Control Plane, the same Universal Cloud Identity setup extends to AWS the moment a workload flips over.

A Vendor That Plays on Your Team

With other cloud providers, getting help often turns into a tug-of-war over whose problem it is. Control Plane's team worked the opposite way — alongside Meso's engineers, treating Meso's problems as shared ones. It showed during the migration, when Control Plane stood up Meso's workloads on GCP within days despite minimal lead time, and it shows in the day-to-day.

Summary

Meso AI is a small engineering team running a real-time, massively parallel multi-agent platform for the world's largest record labels and major brands. With Control Plane, two decisions get made per workload, both push-button: how it scales, and where it runs. That lets a lean team keep shipping features for hundreds of users while the infrastructure adapts to whatever the social world and the cloud pricing landscape does next.