VistanaContracted.ai cuts compute bill by 20%.
Contracted.ai is a San Francisco robot-learning institute. Its training and teleoperation workloads run in bursts, so a cluster sized for its busiest week sat idle, billed at full price, most of the time.
H200s in a dedicated cluster
Lower effective cost
Of idle hours resold as supply
The problem
Contracted.ai sized its cluster for peak demand, then paid full price for it while the GPUs sat idle most of the time.
Its workload isn't one steady training run. It's egocentric video ingestion, reasoning-trace annotation, live teleoperation against an in-house arm farm, and evaluation sweeps, each on its own schedule, so a cluster sized for the heaviest of those passes sat mostly idle the rest of the week.
The solution
We gave Contracted.ai a cluster of its own, then bought back their idle time, cutting total costs by 20%.
- 01
Dedicated cluster
32×H200s, reserved for Contracted.ai alone and managed end to end: hardware, drivers, networking, cluster ops.
- 02
Bought back
Every hour the cluster sat idle, Vistana bought it back instead of billing it at the reserved rate. That cut Contracted.ai's bill automatically, with no change to how they work.
- 03
Resold
Those reclaimed hours went back onto the Vistana marketplace as discounted capacity, growing the supply available to every other team training on Vistana.
The result
Contracted.ai kept its cluster sized for peak, paid 20% less , and turned its idle hours into discounted capacity.
Same cluster, same headroom for the workload that needed it. Neither side paid for GPU-hours nobody was running on.
Got a run coming up? We'll size the cluster and buy back what you don't use.
Get Compute