We built SkyPilot to stop GPU budgets from disappearing.
A small team of distributed-systems engineers tackling the cost of idle reserved GPUs.
The GPU cost problem.
ML teams often pay a 40-60% premium for idle reserved capacity because spot interruptions feel too risky. That concern is valid: a 72-hour training run stopped at hour 68 without a checkpoint is a catastrophe. Teams respond by reserving and over-provisioning, then pay for GPUs idle between runs.
The numbers do not add up. Capacity reserved for peak load sits idle 60-70% of the time. Even with a realistic interruption rate, the supposedly safe option costs more per effective GPU-hour than spot.
With automatic migration and reliable checkpointing, spot is always the right choice. Reserved instances reflect the cost of interruptions that are not handled. Handle that risk programmatically, and the economics change entirely.
SkyPilot sits between your training job and cloud billing. It manages migration, cost routing, and attribution, letting engineers focus on model quality rather than cloud operations.
The team.
Zongheng Yang
CEO and Founder
Zongheng Yang leads product and engineering at SkyPilot. He brings a background in distributed systems and compute scheduling at ML scale. He founded SkyPilot to address GPU costs programmatically.
Marcus Webb
CTO
Marcus Webb is SkyPilot's CTO. He builds job orchestration systems, with a background at a major cloud provider and a high-growth ML infrastructure startup. He is the principal architect of SkyPilot's checkpoint-and-resume engine.
Priya Narasimhan
Head of Platform
Priya Narasimhan, Head of Platform, owns SkyPilot's cost intelligence layer and provider integrations. Her background is in cloud economics and capacity planning. She managed multi-cloud procurement for teams with significant GPU spend.
Build leaner ML infrastructure with us.
Want to join the team or work together? Reach out.