The TCO and ROI of GPUÂ infrastructure by Surf Report
The True Engineering Cost of Open Model Training: A TCO and ROI Analysis of GPU Infrastructure by Surf Report compares Crusoe Cloud, a competing neocloud, and a hyperscaler on cost, effort and reliability across the full cluster lifecycle.
Independent testing and analysis by William McKnight and Jake Dolezal of Surf Report. Sponsored by Crusoe.
The rate card tells you what an hour of compute costs. It doesn’t tell you how many engineers it takes to run the platform, or how much of that hour turns into finished training.
So the engineers ran a common test suite on all three clouds, from a cold-account deployment to a 48-hour, 24-GPU training run on Llama 3.1 70B, with failures injected along the way. Then they modeled what a lean team would spend over three years.
| Metric | Crusoe Cloud | Neocloud | Hyperscaler |
|---|---|---|---|
| Three-year TCO | $5.36M | $6.14M | $8.37M |
| Cost per 1M training tokens | $2.89 | $3.32 | $4.39 |
| Goodput under injected failures | 99.1% | 94.2% | 95.9% |
| Engineers to run the platform | 3.9 | 4.1 | 6.3 |
- How compute, staffing and reliability add up to three-year TCO
- Why the lowest hourly rate in the study didn’t deliver the lowest cost per token
- How the three clouds compare on setup effort, incident response and recovery before you sign
Results reflect lab testing published in September 2026. Production results will vary. The competing neocloud was tested on NVIDIA HGX™ B200 GPUs because NVIDIA HGX™ B300 GPUs weren’t available on demand during testing.