Pretraining infrastructure underwriting

Own the cluster?
Price the uptime.

Turn fixed ownership, reserved capacity, and an on-demand parity point into exact utilization-sensitive economics.

Source-stated baseline
Modelfull-use on-demand = owned TCO ÷ parityscenario on-demand = full-use total × utilizationproductive hours = GPUs × years × 8,760 × utilization
Cheapest pathWaiting
Savings vs next best-
Owned TCO-$/productive GPU-hour
Reserved TCO-$/productive GPU-hour
On-demand at utilization-Derived from parity
Productive hours77% parity85% utilization

Run a scenario to inspect exact outcomes.

Decision boundaryModeled inputs only. Add power, facilities, financing, staff, networking, downtime, resale, contract terms, and delivery risk to owned TCO before procurement.

Utilization earns the economics. Operations earn the utilization.

Owning is a fixed-cost bet that pretraining demand stays high enough to keep the cluster productive.

The source's 77% figure is parity with on-demand, not automatically with reserved capacity.

InfiniBand, scheduling, monitoring, staffing, and supply timing remain outside this simple model unless priced into owned TCO.

Entered parity and utilization expose modeled cluster economics

Read the explanation

The source scenario owned total fifteen point five million and parity seventy seven percent imply full-utilization on-demand cost fifteen point five divided by point seven seven, about twenty point one two nine eight seven million. Bars share fifteen pixels per million dollars: two hundred thirty two point five owned pixels and about three hundred one point nine four eight on-demand pixels. These are modeled source anchors, not verified quotes, procurement options or actual cluster invoices. Reserved cost is a separate input. At entered utilization sixty percent, modeled on-demand is about twelve point zero seven seven nine two two million. At eighty five percent it is about seventeen point one one zero three nine million. Bars share fifteen pixels per million dollars, about one hundred eighty one point one six nine and two hundred fifty six point six five six. The fixed owned fifteen point five and reserved sixteen point eight remain unchanged. Thus the source chooses on-demand in the lower scenario and owned in the higher, under these assumptions only. Two hundred fifty six GPUs times three years times source eight thousand seven hundred sixty hours per year gives six million seven hundred twenty one thousand nine hundred twenty full-use GPU hours. At sixty percent this becomes four million thirty six thousand six hundred eight; at eighty five percent five million seven hundred eighteen thousand five hundred twenty eight. Bars share fifty millionths pixel per productive GPU hour, about two hundred one point eight three and two hundred eighty five point nine three. Cost per productive hour divides fixed millions converted to dollars by that denominator. Calendar variation, downtime and actual achieved throughput are not measured.

Super generates helpful tools and automates fact-checking across the internet proactively. If you enjoyed this tool, build your own with Super and share it with a friend.