Crusoe’s AI cloud study says the hourly GPU rate misses much of the bill

Crusoe says a three-provider study found its cloud had the lowest three-year cost for the workloads tested, with staffing and usable compute helping explain the gap. Its figures make a useful case for looking beyond GPU prices, though the study was sponsored by Crusoe, so the comparisons deserve a careful read.

Crusoe Watch analysis

What happened

The study, conducted by Surf Report and sponsored by Crusoe, modelled comparable workloads across Crusoe Cloud, another neocloud and a hyperscaler. Crusoe’s account puts the three-year totals at $5.36 million, $6.14 million and $8.37 million respectively. It says just under four people were needed to operate its platform, compared with more than six on the hyperscaler, accounting for about $1.27 million of the difference.

The study and its findings also compare training-token costs and performance during injected failures. Crusoe reports 99.1% of paid compute time remained productive on its platform, against 94.2% and 95.9% for the other providers. The competing neocloud was tested on NVIDIA B200 GPUs because B300s were not available on demand, a detail that matters when comparing the results.

Why it matters

An hourly GPU rate is easy to compare, but it leaves out the staff needed to run a platform, the time spent getting started and the work lost to downtime or retries. For teams running long training jobs, cost per completed task or productive compute time may say more about the bill than the neatest number on a rate card.

These figures come from a study sponsored by Crusoe and presented on its own site. They are a useful prompt for what buyers should measure, not a universal verdict on which provider is cheapest for every workload.

Our read

The strongest takeaway is the accounting, not the winner’s podium: include labour, setup and reliability in infrastructure comparisons. Crusoe’s results are striking, but a sponsored comparison should be treated as a starting point for due diligence, not a substitute for testing a buyer’s own workload. The invoice, as ever, enjoys the final say.

What to watch

• Whether Surf Report publishes the full methodology and underlying data.
• How results change across other workloads, hardware configurations and operating teams.
• Whether buyers can reproduce the reported productivity and cost differences in their own environments.

Discussion spark: When choosing an AI cloud, should teams prioritise the lowest headline GPU rate or the lowest cost per completed job, even when that means relying on a provider’s own sponsored comparison?

not affiliated with or endorsed by Crusoe