The economics of paying per GPU-minute

Here's an uncomfortable number: a typical on-prem CFD cluster sits idle for more than 90% of the calendar year. You pay for it on the quiet Sunday nights and the week everyone's at a conference, the same as you do during a crunch.

Usage-based pricing inverts that. Let's do the math.

The on-prem baseline

A modest 8×A100 node, fully loaded — hardware, power, cooling, rack space, and the fractional sysadmin who keeps it alive — runs into six figures a year before you've solved a single case. Depreciate it over three years and you're paying for capacity whether or not you use it.

The hidden line item is utilization. If your team genuinely keeps that node busy 60% of the time, the per-hour cost is reasonable. If it's busy 8% of the time — which is far more common — your effective cost per useful GPU-hour is roughly 7× the sticker price.

What pay-per-minute looks like

With Aviato Studio you pay for GPU-minutes while a case is actively solving. A representative external-aero run:

• Mesh: ~12M elements
• Hardware: 1×A100
• Wall-clock to convergence: ~18 minutes

At an illustrative $0.09 / GPU-minute (A100) that's about $1.60 for the run — and nothing for the 23 hours and 42 minutes of that day when you weren't solving.

The point isn't that cloud GPUs are cheaper per hour. They're often more expensive per hour. The point is you only buy the hours you use.

When on-prem still wins

To be fair: if you run a solver flat-out, 24/7, all year, owning the hardware can pencil out. A few labs genuinely do. For everyone else — bursty workloads, parameter sweeps, the occasional big study — paying per minute turns a capital project into a line item that tracks the work.

And because every Aviato Studio plan ships with hard budget caps, "pay for what you use" never becomes "pay more than you meant to."