Foundry

Foundry Technologies, the a16z-backed AI compute startup founded by former DeepMind researcher Jared Quincy Davis, rebranded to Mithril in August 2025 and repositioned as an "AI omnicloud." Unlike neoclouds such as CoreWeave, Lambda, or Nebius that own or lease their own data centers, Mithril does not operate hardware of its own. It runs a multi-cloud GPU marketplace that orchestrates virtual machines and Kubernetes workloads across other providers, with algorithmically set spot prices recalculated weekly from the prior seven days' supply and demand.

The platform offers Nvidia H100, H200, and A100 capacity, added B200 spot instances in December 2025, and rolled out Flexible Reservations (capacity guarantees with credits for unused hours) in March 2026. A June 2026 SkyPilot partnership deepened its portability story, and it maintains SOC 2 Type II compliance. Storage is priced separately at $0.08 per GB per month for NVMe. The company also publishes Project Ember, an open framework for compound AI systems, reflecting founder Davis's academic work on compound AI at Stanford and DeepMind.

Disclosed customers skew toward research and AI-native teams rather than hyperscale training labs: the Arc Institute (developers of the EVO genomic foundation model), MIT, and Harvard's Broad Institute are named references, with the company describing users running 10,000+ GPU-hours per day. Foundry has not publicly disclosed an owned GPU fleet size, MW footprint, or total funding raised, consistent with an orchestration-layer business model where inventory sits on partner clouds rather than on its own balance sheet.