Mithril Aims to Recoup Billions in Wasted AI Compute Power

PALO ALTO, Calif. – March 19, 2026 – AI infrastructure firm Mithril today launched a new service designed to tackle one of the AI industry's most expensive and persistent problems: idle GPUs. The service, called Flexible Reservations, creates a marketplace for unused computing power, allowing companies to turn the sunk cost of reserved, inactive hardware into a recoverable asset.

At the heart of the announcement is a system that recycles the immense, untapped power of GPUs that sit dormant between AI training jobs. When a company with a reserved block of GPUs pauses its work, Mithril’s platform can now relist that capacity on its Spot service for other customers to use. In return, the original reservation holder earns compute credits, which can be used to offset future costs, fund more research, or even finance upgrades to next-generation hardware. This model introduces a new economic layer to the AI compute stack, promising to unlock value from what was previously considered unavoidable waste.

The Billion-Dollar Problem of Idle GPUs

The backbone of modern AI development is the Graphics Processing Unit (GPU), a specialized processor essential for training large models. To guarantee access to these high-demand resources, organizations often enter into long-term reservation contracts with cloud providers. However, this security comes at a steep price. Industry estimates, cited by Mithril and corroborated by independent market analysis, suggest that between 20% and 60% of this reserved GPU capacity goes unused at any given time.

This underutilization stems from the very nature of AI research and development. The process is not a continuous, straight-line sprint but a series of starts and stops. Teams frequently pause intensive training runs to prepare new datasets, evaluate model performance, debug code, or rethink their architectural approach. During these natural lulls, the reserved GPUs sit idle, but the billing meter keeps running. Some industry reports suggest the waste is even more severe, with certain AI clusters running at only 30-50% utilization and as much as 70% of GPU capacity in some organizations remaining idle or underutilized.

Across the industry, this inefficiency translates into billions of dollars in wasted expenditure annually. For individual companies, a single high-end GPU running at low utilization can waste tens of thousands of dollars per year. Until now, this has been largely accepted as a necessary cost of doing business in AI.

A New Economic Model for AI Compute

Mithril's Flexible Reservations aims to fundamentally change this dynamic. The system operates on a simple but technically sophisticated workflow: Pause, Relist, and Resume.

When an AI team pauses a training job, the platform securely checkpoints the workload, saving its exact state—including model weights, optimizer states, and data progress—to persistent storage. The now-free GPUs are then securely wiped and moved to the Mithril Spot service, a marketplace where other customers can purchase the capacity at a lower, variable rate. For every hour their GPUs are used by another party, the original reservation holder earns recoverable compute credits.

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

When the original team is ready to continue, the workload is restored from its saved state, resuming from the exact point where it left off. The company promises this restoration happens within minutes, a critical factor for maintaining development velocity. The net effect is a dramatic increase in the industry's effective AI capacity without installing a single additional GPU.

Beyond Spot Markets: A Hybrid Approach to Flexibility

The launch positions Mithril in a competitive landscape dominated by hyperscale cloud providers like AWS, Google Cloud, and Microsoft Azure, as well as specialized GPU cloud companies such as CoreWeave and Lambda Labs. While these providers offer solutions to manage costs—primarily through a combination of expensive On-Demand instances, discounted long-term Reserved Instances, and volatile, low-cost Spot Instances—none offer a direct mechanism for a reservation holder to monetize their own idle time.

Traditional spot markets allow users to access spare capacity at a deep discount, but they come with the significant drawback of potential interruptions, making them unsuitable for long, mission-critical training jobs. Reserved instances offer stability but create the idle capacity problem. Mithril's solution creates a hybrid model that aims to provide the best of both worlds: the security of a reservation with the economic efficiency of a dynamic spot market.

By allowing reservation holders to become suppliers to the spot market, the platform introduces a new form of liquidity into the AI compute ecosystem. It transforms a static, one-way financial commitment into a dynamic, two-way value exchange, a concept that could prove disruptive if widely adopted.

Fueling the Next Wave of AI Innovation

The most significant impact of this new economic model may be on the pace of innovation itself. By lowering the effective cost of compute, Flexible Reservations could empower teams to shift from running experiments serially to running multiple in parallel, a contest-style approach that can accelerate breakthroughs.

For startups and academic labs operating on tight budgets, the ability to earn back a portion of their largest infrastructure expense could be transformative. Early adopters cited by Mithril include AI startup Standard Intelligence, developer tool company Cursor, LG AI Research, and the Arc Institute, representing a cross-section of the AI ecosystem that stands to benefit.

"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," commented Devansh Pandey, Co-Founder of Standard Intelligence. "Flexible Reservations gives us that freedom, and we're excited to use it at scale."

Furthermore, Mithril is positioning the credits as a direct pathway to next-generation hardware. The company states that credits can be applied toward the cost of migrating to newer GPUs, including NVIDIA's powerful Blackwell B300 series. This could significantly ease the financial burden of hardware refresh cycles, allowing more organizations to stay at the cutting edge of AI capability. The launch builds on other recent company milestones, including new partnerships with cloud providers Nebius and Oracle and the integration of NVIDIA's B200 GPUs into its platform, reinforcing its mission to build a more efficient and accessible global marketplace for AI compute.