Mithril aggregates GPU, CPU, and storage resources across multiple cloud providers and exposes them through a single orchestration interface. The platform decouples compute provisioning from cloud vendor selection, eliminating the operational burden of managing accounts, pricing negotiations, and availability across providers for ML training and inference workloads.
The core trade-off is availability versus cost. Reserved capacity - available in hourly, daily, or monthly windows - provides predictability and lower latency for time-sensitive jobs. Spot-style provisioning on excess capacity offers significantly lower unit costs for asynchronous, interruptible work, particularly for batch inference and training jobs with flexible completion windows. Transparent, market-based pricing replaces bespoke vendor negotiations, reducing procurement friction but exposing users to capacity competition.
Access is programmatic: a batch SDK and batch inference API enable teams to define and submit workloads without managing cloud credentials or scheduling logic directly. The abstraction shifts complexity from per-cloud configuration to unified job specification, though users remain responsible for failure handling and retry strategies when using lower-cost spot capacity.