Problem

Lilac addresses the AI infrastructure bottleneck created by expensive, scarce, and poorly utilized GPUs. Most GPU clusters run at only 30–50% utilization, while enterprise on-premise clusters can average as little as 10%; the hardware is already powered on, cooled, and paid for, yet its idle capacity cannot easily be sold or accessed. For AI companies, the alternative is to reserve enough dedicated capacity for peak demand and pay for machines that sit idle between training runs or product launches.

The killer use case is bursty model inference and serving. A model endpoint may run at low volume for days and then require substantially more GPUs immediately after a launch or customer rollout. Lilac lets providers monetize that idle window while giving startups and researchers cheaper compute without requiring them to procure or reserve a cluster for the peak.

Product / Service

Lilac is an AI cloud and GPU network that combines dedicated GPU infrastructure, serverless OpenAI-compatible inference, batch workloads, subscriptions, and a capacity exchange. On the supply side, its Kubernetes operator detects idle GPU capacity, deploys secured paid workloads, meters usage, and preempts those workloads within seconds whenever the provider’s own jobs need the GPUs back. Suppliers can define which node pools and availability windows to expose, while Lilac manages demand routing and revenue sharing.

On the demand side, customers can call supported models through a drop-in OpenAI-compatible API, pay per token with no contract or reserved-capacity commitment, and route requests to idle enterprise GPUs. Customers can also reserve dedicated GPU VMs, nodes, or clusters, monetize eligible unused reservation windows through Lilac Flex, or transfer part of an eligible commitment to another verified Lilac customer. The economic benefit is lower inference cost and faster access to capacity, while providers earn revenue from hardware that would otherwise remain idle.

Market

Lilac competes at the intersection of GPU cloud infrastructure, spot or marketplace-based GPU compute, and serverless model inference. Relevant comparables include Vast.ai, which operates a marketplace connecting users with GPU providers; CoreWeave, an AI-focused cloud competing in training and inference; Modal, which provides serverless GPU execution; and Lambda’s on-demand GPU cloud. Lilac’s distinctive positioning is its focus on idle enterprise and cloud-provider capacity, preemptible workloads, OpenAI-compatible inference, and mechanisms that make dedicated commitments more flexible.

The company was founded in 2025, is an active Y Combinator Summer 2025 company with a reported team size of four, and had moved from early design-partner outreach to self-serve inference by April 2026. It has also announced a partnership in which Saturn Cloud selected Lilac as a GPU capacity provider for Token Factory, with Lilac capacity available to Saturn customers for model serving and per-token inference. The evidence establishes commercial availability and early partnership traction, but does not disclose revenue, ARR, customer count, or profitability; therefore Lilac should be characterized as an early commercial company rather than definitively labeled pre-revenue.