Capacity cohorts: MW, GPUs, utilisation and revenue
Cohorts of compute, GPU-hours, power and revenue.
By Surojit Chakraverti, founder of L3VLUP and an investor running a long-short healthcare and technology equities strategy.
On this subject: capacity cohorts: mw, gpus, utilisation and revenue.Updated 30 October 2026
Part of project finance and infrastructure models, with the models, labs and guides around it.
What it does
The operating base of a GPU cloud. Capacity arrives in cohorts, each commissioned in its own year with its own MW of IT load, GPUs per MW and capital cost. Available GPU-hours are GPUs times the hours in a year; a share is sold on take-or-pay contracts at a fixed price while the contract runs, and the rest on demand at a utilisation and a spot price that falls as newer GPUs arrive. Power follows the hours actually run, at kW per GPU times PUE times the price per MWh; space, staff and other costs follow the capacity in service.
Where it lives
The same schedule in each model that carries it, with the rows to open. Open a model in the browser, go to the sheet, and click the lines.
- Assumptions, rows 23–38: Cohort 1: first year in service, Cohort 1: capacity, Cohort 1: GPUs per MW, Cohort 1: GPUs and servers, capital cost per MW…
- Assumptions, rows 56–57: On-demand utilisation in a cohort’s first year, % of ramped, Spot price, Y1
- Assumptions, rows 60–67: Power usage effectiveness (PUE), Power price growth, a year, Co-location, per kW of IT load a month, Y1, Site staff per MW…
- Capacity, rows 6–11: Cohort 1: in service, Cohort 1: year of service, Cohort 1: build year, Cohort 1: under contract…
- Capacity, rows 14–19: Cohort 2: in service, Cohort 2: year of service, Cohort 2: build year, Cohort 2: under contract…
- Capacity, rows 22–27: Cohort 3: in service, Cohort 3: year of service, Cohort 3: build year, Cohort 3: under contract…
- Capacity, rows 30–36: MW, opening, MW commissioned, MW in service, GPUs commissioned…
- Revenue, rows 6–17: Cohort 1: available GPU-hours, Cohort 1: contracted GPU-hours, Cohort 1: on-demand utilisation, Cohort 1: on-demand GPU-hours sold…
- Revenue, rows 20–35: Spot price, Cohort 1: contracted revenue, Cohort 2: contracted revenue, Cohort 3: contracted revenue…
- Operating Costs, rows 6–7: Power price, Cost index
- Operating Costs, rows 10–21: Cohort 1: power for contracted hours, Cohort 1: power for on-demand hours, Cohort 1: space, site staff and other, Cohort 1: contracted cash flow for its lender…
- Operating Costs, rows 24–40: Revenue, Power, contracted hours, Power, on-demand hours, Co-location…
- Reported, rows 89–95: Active power at 31 December 2022, approximately, Active power at 31 December, as stated, Total contracted power at 31 December, approximately, Data centres…
- Derived, rows 49–60: Average active power in the year (mean of the year ends), Revenue per average active MW, a year, Lease cost per average active MW, a year, Active power added in the year…
- Forecast, rows 11–14: Active power at the year end, base (editorial), Active power, downside (editorial), Active power, upside (editorial), Active power at the year end, selected
- Capacity, rows 6–10: Active power at the year end, Average active power in the year, Average active power in the second half of 2026, Active power added…
What a reviewer looks for
- All capacity sold at one price, so the end of a contract changes nothing.
- Power from MW of capacity rather than the hours run, or without PUE.
- Every cohort commissioned at once, which hides the build-out.
Learn it, then build it
Vocabulary: CapEx, Useful Economic Life.