Lilac Builds a Spot Marketplace to Monetize Idle GPUs and Cut Inference Costs

Lilac, a Y Combinator Summer 2025 company, is on a mission to turn idle GPU capacity into a market — automatically. The same data centers that everyone is racing to fill with new GPUs are sitting on a quieter problem: a meaningful share of the silicon they already have is idle at any given moment. Founded in 2025 by brothers Ryan and Lucas Ewing, the four-person San Francisco startup is building a spot GPU marketplace that taps idle compute from cloud providers and enterprises and routes it to startups and researchers that need cheaper inference.

Lilac’s pitch is elegant in its symmetry. On one side of the marketplace are organizations — cloud platforms, enterprises with on-premise clusters, and operators of mixed fleets — that have GPUs sitting underutilized but still drawing power, depreciation, and opportunity cost. On the other side are AI builders for whom inference costs have become the dominant line item in their infrastructure bill. By matching the two through automated, Kubernetes-integrated infrastructure, Lilac aims to give buyers significantly cheaper compute while letting providers monetize capacity that would otherwise simply sit there. The mechanics resemble the spot markets that have long underpinned other commoditized cloud resources, but tuned for the unique scheduling and reliability requirements of GPU workloads.

The founders bring complementary backgrounds to a problem that sits squarely between cloud infrastructure and machine learning. Ryan Ewing previously built cloud and networking services at AWS, where he gained firsthand experience with the systems that abstract away physical compute and turn it into a billable resource. Lucas Ewing, a Harvey Mudd College graduate (Class of 2023) who was born and raised in Stockholm before moving to the United States for high school and college, brings a builder’s perspective and a stated mission of democratizing GPU access. Together, the brothers are betting that the right software layer can transform a fragmented, opaque hardware market into something closer to a transparent commodity exchange.

The timing is hard to argue with. Demand for GPU inference is climbing as more applications graduate from prototype to production, while the unit economics of running models at scale remain a constant source of anxiety inside AI startups. At the same time, the buildout of new data center capacity has run ahead of fully optimized scheduling, leaving pockets of unused compute scattered across providers and on-premise environments. A marketplace that can surface that latent supply, abstract it behind a single interface, and present it to buyers at spot prices addresses both sides of an inefficiency that has so far mostly been talked about in private.

For Lilac, the early focus is on automation and developer experience. By handling Kubernetes integration out of the box, the platform aims to remove the operational friction that has historically made multi-source GPU strategies the preserve of large engineering organizations. If the company can deliver on that promise, the result could be lower inference bills for the next generation of AI products and a new revenue stream for the operators of the world’s growing GPU footprint.