Mithril

Mithril describes itself as an "AI omnicloud": a platform that aggregates GPU capacity from multiple cloud providers into a single marketplace and exposes it through spot pricing and flexible reservations. Customers can place spot bids (a low bid means the workload waits until the market price drops; a high bid acts as a safety limit and starts the workload immediately at the clearing price) or take instant, hourly reservations that guarantee access while crediting back or allowing pause/return of capacity that goes unused. Commonly used GPU types cited on the site include H200, H100 and A100, with prices moving with supply and demand.

Beyond raw instances, the company has introduced abstractions on top of the aggregated market: a Batch Inference API with standard token-based pricing, and Flow, a batch job SDK for general-purpose ML workloads; both were announced as available in beta at the time of the August 2025 rebrand, while omnicloud compute was described as generally available for AI teams and research labs. Documentation covers instance types, reservations, spot bids, startup scripts, quotas, ephemeral and persistent storage, compute clusters, a CLI, and a compute API, and notes that the service is generally available for work purposes while individual users are waitlisted.

The company positions the platform as complementary to existing clouds rather than a replacement: it hosts a Kubernetes control plane so reserved and spot instances join a single cluster usable with kubectl, its CLI is built on open-source SkyPilot so existing task YAMLs run unchanged and remain portable, and cloud object storage buckets (S3, GCS, R2) can be mounted into tasks. The site also lists SOC 2 Type II, SSO, a status page and on-call engineers.

Founding story

The company states it was founded by alumni from Google DeepMind's core Deep Learning team and Stanford's computer science PhD program, and describes the team as researchers, engineers, economic experts and go-to-market leaders spanning ML theory, distributed systems and financial modeling. Individual founder names and a founding date are not given in the sources.

Business model

Mithril sells access to GPU virtual machines sourced from multiple cloud providers through a marketplace, with two consumption modes: spot capacity priced at a live market clearing rate set by user bids, and flexible hourly reservations that guarantee access and credit back unused capacity. A Batch Inference API is offered on standard token-based pricing. The company states it operates without long-term contracts.

Usage-based: spot GPU instances billed at the market clearing price, hourly/flexible reservations with credit-back for unreturned-to-use capacity, and token-based pricing for the Batch Inference API. Public pricing and live price pages are referenced on the site.

Traction

The website states teams run more than 10,000 GPU-hours per day on Mithril and lists customers and partners across AI-native startups, research labs, public companies and academic institutions. The rebrand post names Arc Institute, Cursor and LG AI Research as practitioners who used the platform under the Foundry name, and includes an endorsement from Arc Institute co-founder and CEO Patrick Hsu, whose team developed the EVO genomic foundation model. The Batch Inference API is stated to process 500M+ tokens per day by default.

Latest developments

August 4, 2025: rebrand from Foundry to Mithril, with the launch of the Batch Inference API and Flow batch-job SDK in beta and general availability of omnicloud compute for AI teams and research labs. Documentation was last updated approximately five months before retrieval and notes individual users remain waitlisted.