On-prem H100 cluster pays back inside 18 months
“Payback period outran the price list.”
Cause
Two AU-region hosted price cuts in Q1 and Q2 moved the breakeven to over four years before we finished the model; the standing answer now carries the calculation and its refresh date.
What was tried, what was hit, what it means
The proposal was a spreadsheet, not a run: eight H100s in a Sydney colocation, amortised over three years, against hosted per-token pricing at November 2025 rates for a modelled 400M tokens a month across two banking patterns. At those prices it paid back in 16 months and the argument for sovereignty came free.
The spreadsheet was assessed by the lab and a sector owner in January and looked sound. Then hosted AU-region pricing dropped in February and again in April, the second time by roughly 40% on the mid-tier models that do most of the modelled volume. Recomputed, the breakeven moved past four years, which is longer than the useful life of the hardware. We never got to a run because the input assumption did not hold still long enough to preregister against.
The question is still real and now lives in sa-onprem-when, which carries the breakeven calculation with a stated refresh cadence rather than a fixed answer. The cases where on-prem still wins are the ones where the volume is above roughly 2B tokens a month or the data cannot leave the building at all; both are named in the standing answer.
Lessons
- 01A breakeven against a price that moves quarterly is a standing answer, not a position.
- 02Sovereignty is a separate argument from cost; when they are bundled, one of them is doing the work and it is usually not the one on the slide.
- 03Name the volume threshold at which the answer flips; that number ages better than the answer.
Record
Field
On-Prem Inference Lineage intact. The reasoning that killed this stays attached to everything it touched.Superseded by
When does on-prem inference make sense?Try again?
Proposals that touch this entry get its returned notes attached automatically.