Introducing Lilac: Turn Idle GPU Capacity into Revenue

The problem with GPU utilization

If you're running GPU workloads on Kubernetes, you already know: GPUs are expensive. An 8x A100 node costs upwards of $25/hour, and most clusters sit idle for significant chunks of the day. Training jobs finish, inference traffic dips overnight, and those GPUs just... wait.

That idle time is money left on the table.

What Lilac does

Lilac is a Kubernetes operator for GPU suppliers. It detects idle GPU capacity in your cluster and automatically deploys paid workloads on it. When your own workloads need the GPUs back, Lilac preempts immediately — your jobs always come first.

Here's what makes it different:

• One command install — kubectl apply and you're live with minimal setup.
• Policy-driven — You choose which node pools to expose, set availability windows, and define preemption thresholds.
• Revenue sharing — We handle demand routing. You earn from capacity that would otherwise sit idle.
• Transparent reporting — Usage and earnings are metered for clear payout accounting.
• Zero risk — Your workloads always take priority. Lilac workloads are preemptible by design.

How it works under the hood

When Lilac detects an idle GPU node, it deploys secured inference workloads and meters usage for payouts. Workloads are isolated from your cluster jobs. When your scheduler needs capacity back, Lilac's preemption controller gracefully terminates monetized workloads and returns GPUs within seconds.

No downtime. No disruption. No config changes.

Who it's for

Lilac is built for teams running GPU-heavy Kubernetes clusters — ML platforms, inference services, research labs, and GPU cloud providers. If you have nodes that aren't busy 24/7, you're a fit.

Get started

We're currently onboarding early design partners. If you're running GPU workloads on Kubernetes and want to turn idle time into revenue — talk to us.

Want to understand the economics of GPU idle capacity? Read our deep-dive: The GPU Scarcity Paradox.

If you're on the demand side and want to run inference on idle GPUs, check out Lilac's inference API.