Stop comparing GPU hourly rates: the new math of buying AI infrastructure

Total cost beats sticker price

AI infrastructure vendors are converging on the same pitch: stop shopping on hourly rates. In recent procurement-facing arguments — Crusoe on total cost of ownership, fal on verifiable serverless-GPU checklists, and Cerebras on why the agent harness decides performance more than the model — each makes its case directly. Each is published by the vendor itself, so read them as vendor-written buying guides with auditable parts. m:…7896

Crusoe's contribution is a framework built on an independent Surf Report study. The headline finding: the provider with the lowest quoted price was not the one that delivered the lowest total cost. Crusoe says the true cost comes down to three things — the quoted price, the staff time to get the system running and keep it running, and the amount of useful work actually produced. It introduces goodput — the share of paid compute time that advances a task toward completion, excluding errors and overhead — as the metric that matters more than raw throughput. m:…7896

The company then scores itself on that metric: Crusoe Cloud claims 56% better ROI than a hyperscaler over three years, required just under four staff to operate versus more than six on the hyperscaler, kept 99% of paid compute time productive in a failure-injection stress test versus roughly 94% and 96% for two competitors, recovered in about 9 minutes versus 13 to 14 minutes, and delivered the lowest cost per training token at $2.89 per million versus $3.32 and $4.39. The study is described as independent, but the findings are published on Crusoe's own blog, so treat the framework as the vendor's measurement proposal rather than a neutral benchmark. m:…7896

The serverless checklist

Fal's guide is more checklist than pitch: five verifiable criteria for choosing a serverless GPU platform — cold-start time for your model, which runner states are billed, GPU pricing, scaling control, and container compatibility. Two corrections bite hardest. First, headline cold-start numbers are typically measured on small models with spare cluster capacity, so fal tells teams to test with their own model weights at peak hours. Second, per-second billing is not always cheaper than hourly: fal puts the break-even for an H100 at about 55% utilization — above that line, a dedicated machine at a lower hourly rate wins. m:…6600

The guide also makes its own billing auditable: fal says it does not bill for image pulls, pending time, or failed 5xx requests, while idle runners stay up for a default 60-second keep_alive that is billed. It offers B300, GB200, B200, H200, H100, and RTX PRO 6000 GPUs for serverless inference. The author is fal's J Jim Cai, so read the checklist as a vendor's transparency move and verify the billing claims against your own invoice. m:…6600

The harness decides agent speed

Cerebras makes the case that infrastructure buying starts above the hardware layer. An agent, in its definition, is a model inside a harness operating in an environment, and the harness decides what the model can see, what actions it can take, how results feed back into the next inference, and when work is actually done. With fast inference, the company argues, the primary bottleneck has shifted from inference speed to the harness itself — more specifically, tool calls. The evidence it offers: the same Claude Opus 4.5 scored 45.9% on a standardized SEAL scaffold but 55.4% through Claude Code on SWE-bench Pro — a benchmark result that belongs to a model-harness pair, not a model alone — and in one controlled run, the Factory Droid harness finished a task 2.4x faster than Claude Code while using roughly half the tool calls. m:…3032

How to read them

As reasoning rather than measurement: put goodput, billing-state transparency, and harness efficiency into the evaluation before comparing models or hourly rates. Each vendor's numbers favor its own platform, which is exactly why the verifiable parts are the parts worth adopting — fal's billed-states list, Crusoe's goodput definition, Cerebras's model-harness score pairs. The open question is how these frameworks hold up under independent replication, since no neutral party has yet published the same comparisons. m:…6600