Modular as AI infrastructure control plane
Modular
This pushes Modular up from selling faster model serving to selling control over the whole GPU fleet, which is where enterprise budgets get much larger. Mammoth is a Kubernetes-native orchestration layer that routes jobs across clusters, keeps utilization above 90%, and supports NVIDIA, AMD, and CPU environments. That makes Modular look less like a point inference vendor and more like a data platform style control plane that can charge as usage grows across teams and workloads.
-
The practical product shift is from single endpoint serving to cluster management. Enterprises are not just buying tokens or one model endpoint, they are buying scheduling, routing, batch processing, reliability, and quality of service across shared infrastructure. Those are broader, stickier workflows that usually support larger annual contracts.
-
Mixed hardware matters because big companies rarely run one clean stack. Mammoth is designed to place workloads across different chips and clouds, while Modular also offers an OpenAI-compatible batch API. That lets a central platform team standardize operations without rewriting apps every time hardware availability or pricing changes.
-
The closest economic analogy is Snowflake and Databricks, where the vendor becomes the layer that meters and optimizes shared compute. Snowflake emphasizes consumption based pricing, and Databricks has built a large enterprise business with many customers spending six and seven figures annually. Modular is aiming for a similar budget owner, but for AI infrastructure instead of data warehousing.
If Modular keeps expanding from serving into orchestration, it can become the operating layer enterprises use to allocate scarce AI compute, enforce performance tiers, and add new hardware without retraining every engineering team. That is the path from an infrastructure tool to a core platform with larger deals, deeper lock in, and expanding usage revenue.