Mithril Launches Flexible Reservations, Recycling Idle GPU Time Into the Mithril Spot Service

New service recycles idle GPU time into Mithril's Spot service, earning AI teams recoverable compute credits they can reinvest in development, savings, or next-generation hardware

PALO ALTO, Calif., March 19, 2026 /PRNewswire/ -- Mithril, the omnicloud platform for AI compute, today launched Flexible Reservations—a new service that recycles idle GPU time into the Mithril Spot service. When reserved GPUs sit unused between projects, Flexible Reservations puts that idle capacity to work for other customers and returns recoverable compute credits to the original reservation holder—credits teams can reinvest in development, use to create savings, or apply toward next-generation hardware.

The Problem: Billions in Idle GPUs

GPU reservations are the backbone of AI infrastructure, but they come with a costly trade-off. Between training runs, teams routinely pause to prepare data, evaluate results, or rethink model architectures. During those gaps, reserved GPUs sit idle—and the meter keeps running. Industry estimates suggest that 20–60% of reserved GPU capacity goes unused at any given time, representing billions of dollars in wasted spend across the industry.

Until now, teams have had no choice but to absorb those costs. Flexible Reservations changes that.

"The AI industry has always treated idle compute as a sunk cost. Flexible Reservations recycles it. When teams pause or pivot, their unused GPU time flows back into our Spot service and earns credits they can reinvest," said Jared Quincy Davis, Founder and CEO of Mithril. "That means more experimentation, faster breakthroughs, and a smoother path to next-generation hardware."

How It Works: Pause, Relist, Resume

Flexible Reservations follows a simple three-step workflow:

• Pause. When a team pauses a training job, Mithril securely saves the workload state and frees the GPUs.
• Relist. The freed GPUs are moved to the Mithril Spot service, where other customers can use them. The original reservation holder earns recoverable credits for every hour their GPUs are in use by someone else.
• Resume. When the team is ready, the workload is restored within minutes—returned to the exact state where it left off.

The net effect is more usable compute without installing a single additional GPU—Mithril expands the industry's effective AI capacity by recycling idle time that would otherwise go to waste.

Built for the Way AI Teams Actually Work

Modern AI development is not a continuous, straight-line process. Teams stop and start constantly—evaluating checkpoints, swapping datasets, coordinating across research groups. Traditional GPU reservations penalize these natural pauses by charging full price for capacity that no one is using.

Flexible Reservations turns those pauses into a source of value. By earning credits during downtime, teams can shift from running experiments one at a time to running several in parallel—a contest-style approach that Mithril projects can help customers bring AI projects to market much more quickly. Credits can also be applied toward the cost of migrating to newer hardware, including NVIDIA's Blackwell B300s and future GPUs, making it easier for teams to upgrade on their own timeline without a large upfront investment.

"Burst compute is necessary for us to leverage our capital to train frontier models, by using it at times in our research process where it's most valuable and getting fast feedback loops for our largest runs," said Devansh Pandey, Co-Founder of Standard Intelligence. "Flexible Reservations gives us that freedom, and we're excited to use it at scale."

A New Economic Layer for AI Infrastructure

Flexible Reservations reflects Mithril's broader mission to eliminate waste in AI infrastructure. Today, GPU reservations work like a gym membership—you pay whether you show up or not. Mithril's model recycles every idle hour back into productive use, helping AI organizations get more out of every dollar they spend on compute.

The launch builds on recent Mithril milestones, including the introduction of partnerships with Nebius and Oracle, the availability of NVIDIA B200 GPUs and a continued focus on giving AI builders the tools to move faster and build better.

About Mithril

Mithril is the omnicloud platform for AI compute, built to give every AI practitioner the freedom to access and scale the resources they need, when they need them. By aggregating capacity from leading cloud providers and data centers into a single marketplace, Mithril maximizes price-performance while eliminating the complexity of managing infrastructure. Trusted by AI leaders including Cursor, LG AI Research and Arc Institute, Mithril enables teams to focus on advancing AI—not negotiating for compute. Learn more at mithril.ai.

SOURCE Mithril