For labs / Reserved capacity

Reserved capacity

LAST UPDATED IN SEPTEMBER 2026

Reserved capacity is how a lab locks in a price and a quantity for a year and knows exactly what a training run will cost before it starts. H100 only.

Minimum purchase: 8 GPUs. Term: 12 months. Fixed price: 1.89 USD per GPU-hour, for the whole term, unchanged by the on-demand price. The smallest unit we sell is a full cluster: 8 × H100-80GB, 480 GB total GPU memory. Available on cluster M1 (us-west).

The maths

Monthly price = GPUs in reservation × $1.89 per GPU-hour × 720 GPU-hours per GPU. Every month is billed as 720 GPU-hours per GPU — 30 days × 24 hours.

If a reservation uses fewer GPU-hours in a month, the unused hours are not refunded and do not roll over. Billed monthly, on the first of the month. A partial month is billed as a full month. If a reservation is cancelled, the remaining months are not refunded.

How it works

• Email reserved@melaps.system with your estimated GPUs and duration.
• We send you a quote for the number of GPUs you want, with the price locked in for 12 months.
• You sign the reservation agreement and pay the first month in advance. A reservation agreement is a plain text file, 3 pages. No deposit is required.
• We provision your reservation within 5 business days.

Changes, cancellation, transfer

You can increase the size of your reservation at any time. The price is locked in for the remainder of the term. You cannot reduce the size of your reservation.

You can cancel your reservation at any time. You will not be charged for the remaining months, but you will not be refunded for the months you have already paid for. If you cancel early, you pay a penalty of 25% of the remaining months’ price. Penalty = 25% × [monthly price × remaining months]. Example: cancel after 6 months on a full cluster — 25% × ($10,886.40 × 6) = $16,329.60.

You can also transfer your reservation to another lab, at no cost. Reservations automatically renew at the end of the term, with the same price and the same number of GPUs. You can cancel the renewal at any time before it starts.

Availability

A reservation guarantees access to the GPUs you reserved, at any time, for the full term. If a GPU fails, we will replace it within 4 hours. If the whole cluster fails, we will rebuild it within 36 hours. A reservation includes the right to run one job at a time on each GPU, and to use the reserved GPUs for any job, at any time.