The company formerly known as Foundry has a new name—and a bold new ambition. Now Mithril, the rebranded platform is on a mission to make AI compute as accessible and fluid as electricity, starting with a serious upgrade to how inference workloads are run.
Alongside its rebrand, Mithril today launched a batch inference API that promises to cut inference costs by 2–5x, while eliminating much of the traditional pain around provisioning scarce GPU resources. And if early traction is any indicator—the platform’s usage has already surged 550% since its limited preview last year—this isn’t just another infrastructure play. It’s a fundamental shift in how AI-native companies access and scale compute.
What’s in a Name?
The new name, Mithril, isn’t just about branding polish. It reflects the company’s deeper mission: abstracting away the rigidity of traditional AI infrastructure in favor of something more dynamic, flexible, and fair.
“Traditional AI cloud procurement models force companies to overbuy or constantly hunt for better GPU pricing,” said Jared Quincy Davis, founder and CEO of Mithril. “We’ve built a system where they don’t have to. Whether you need compute for a few hours or a few months, we give teams access to global capacity through a single, transparent platform.”
That platform—called omnicloud—is Mithril’s not-so-secret weapon. It virtualizes GPU resources across multiple providers, creating a fluid, market-based marketplace that aggregates supply and simplifies access. The result: no long-term contracts, no bespoke negotiations, no bidding wars.
Enter Batch Inference: Cheaper, Faster, Easier
Mithril’s new batch inference service is the company’s most important product launch yet—and one that could change the calculus for AI startups and enterprises alike. It’s built for workloads like:
- Document summarization
- Multimodal annotation
- Video/audio transcription
- Synthetic data generation
- Large-scale content generation
Instead of building and managing custom infrastructure or battling for GPU allocation, customers simply submit their inference jobs via API, pay with tokens, and let Mithril handle the orchestration across its distributed compute network. It supports both off-the-shelf and custom models, with automatic scaling and fault-tolerant infrastructure baked in.
The value prop is straightforward: high-performance inference at a fraction of the cost—and with none of the usual operational drag.
“We trained the first textless conversational audio base model for under $200,000—something that would’ve cost over $10 million with traditional infrastructure,” said Devansh Pandey, co-founder of Standard Intelligence. “Mithril’s burst compute made that possible.”
Meeting the Moment in a GPU Gold Rush
As demand for GPU compute skyrockets—driven by everything from LLMs to generative media—the global GPU market is projected to hit $230 billion by 2030. But paradoxically, a lot of GPU capacity sits idle due to fragmented provisioning, inefficient allocation, and outdated pricing models.
Enter Mithril’s market-driven approach, which lets idle capacity find jobs and jobs find capacity—without the cloud lock-in or high upfront costs.
Paul Milgrom, Nobel laureate and Mithril advisor, put it bluntly:
“Tremendous demand and a large fraction of idle time make sharing a perfect solution. Mithril’s market is the right approach to making that happen.”
The Bigger Picture
The rebrand to Mithril isn’t just cosmetic. It signals a strategic leap from niche compute provider to core infrastructure for distributed AI workloads—aiming to become the connective tissue between global GPU supply and the exploding demand for inference and training.
Whether you’re a research lab needing GPUs for a week, or an enterprise running nightly inference across terabytes of data, Mithril’s omnicloud platform offers a compelling alternative to the slow-moving, locked-in world of legacy cloud compute.
“Our goal is to make compute as universally available and accessible as electricity,” Davis said. Judging by their momentum—and the caliber of early customers like Cursor, LG AI Research, and Arc Institute—they might just pull it off.
Power Tomorrow’s Intelligence — Build It with TechEdgeAI
