Case studies
Workloads, written up end to end.
Each study follows one team from the first node to steady production: the setup, the numbers that mattered and what they would change next time.
The studies below are samples that show the format. Companies and figures are fictional.
- 64 × B300
- 18.4 TB HBM3e
- per-sec billing
How a video studio trains on 64 B300 GPUs
Training runs outgrew one machine. Now each new model starts on an 8-node cluster the day it is planned.
- B300 GPUs in one cluster
- 64B300 GPUs in one cluster
- nodes on a single fabric
- 8nodes on a single fabric
- EU region, EU-AMS-1
- 1EU region, EU-AMS-1
How a speech lab serves transcription on Serverless
Traffic follows European office hours. Their GPU bill now does too, with workers that scale to zero at 19:00.
Serverless
- fold-batch-01€16.00
- fold-batch-02€16.00
- fold-batch-03€16.00
How a protein lab runs batch jobs on reserved H100s
A shared university queue meant days of waiting. Four reserved servers now take every job the same morning.
Cloud Servers
- GPUs
- —GPUs
- nodes
- —nodes
- region
- —region
Your workload, written up next
Running something interesting on Nexpod? We write the study with your engineers, and you approve every line before it goes live.
Tell us about itState of GPU Compute 2026
Seven chapters on where GPU-hours go, how big clusters really are and how busy rented hardware is. Free as a PDF.
7
chapters
38
pages, free PDF
61%
of GPU-hours on inference
Your turn
Run it on whole servers. Write it up with us.
Pick a server, pick a region, and connect over SSH. Stop it whenever you are done; the meter stops with it.
- B300· 288 GB HBM3e€6.00/GPU·hr
- B200· 180 GB HBM3e€5.00/GPU·hr
- H200· 141 GB HBM3e€4.00/GPU·hr
- H100· 80 GB HBM3€3.00/GPU·hr