GPU cloud infrastructure model
GPU cohorts, contracts and backlog, power, facility debt. Build a GPU cloud from cohorts of capacity: GPUs and capital cost per MW, contracted and on-demand GPU-hours and revenue, a backlog roll-forward, power from PUE and the price per MWh, GPU and fit-out depreciation, a term loan per cohort sized on its contract, corporate notes and a revolver, liquidity, covenants and the equity return.
By Surojit Chakraverti, founder of L3VLUP and an investor running a long-short healthcare and technology equities strategy.
On this subject: gpu cloud infrastructure model.Updated 30 October 2026
Who builds it, and for whatThe model an infrastructure or private credit investor builds before lending to or investing in a GPU cloud, and the one a banker or research analyst builds to understand how such a company is financed. It answers what the contracts are worth and when they turn into revenue and cash, how much each cohort of GPUs can borrow against its contract, what the business earns once the contracts end and the GPUs are rented at whatever the market pays, and whether there is enough liquidity to meet the notes when they fall due.
Every sheet below, in your browser, with its outputs and checks
Starter workbook free with an account. Worked model: one a month with Plus, every one with Pro. Downloads
| A | B | C | D | E | F | G | H | I | J | K | L | M | |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 1 | Assumptions | ||||||||||||
| 2 | $m unless stated. GPU-hours in millions; prices in dollars per GPU-hour; power in dollars per MWh. | ||||||||||||
| 4 | Year | Unit | Y0 | Y1 | Y2 | Y3 | Y4 | Y5 | Y6 | Y7 | Y8 | Notes | |
| 5 | Year number | # | 0 | 1 | 2 | 3 | 4 | 5 | 6 | 7 | 8 | Y0 = financial close and the first cohort’s build. | |
| 7 | Scenario | ||||||||||||
| 8 | Scenario | # | 1 | 1 = base, 2 = upside, 3 = downside. Sets on-demand utilisation, the fall in the spot price and the power price. | |||||||||
| 9 | On-demand utilisation, base | % | 80.0% | Share of uncontracted GPU-hours sold, once a cohort has ramped. | |||||||||
| 10 | On-demand utilisation, upside | % | 88.0% | ||||||||||
| 11 | On-demand utilisation, downside | % | 65.0% | ||||||||||
| 12 | On-demand utilisation, selected scenario | % | 80.0% | ||||||||||
| 13 | Fall in the spot price, a year, base | % | 10.0% | ||||||||||
| 14 | Fall in the spot price, a year, upside | % | 6.0% | ||||||||||
| 15 | Fall in the spot price, a year, downside | % | 16.0% | ||||||||||
| 16 | Fall in the spot price, selected scenario | % | 10.0% | ||||||||||
| 17 | Power price, Y1, base | $ per MWh | 70.00 | ||||||||||
| 18 | Power price, Y1, upside | $ per MWh | 62.00 | ||||||||||
| 19 | Power price, Y1, downside | $ per MWh | 85.00 | ||||||||||
| 20 | Power price, Y1, selected scenario | $ per MWh | 70.00 | ||||||||||
| 22 | Cohorts | ||||||||||||
| 23 | Cohort 1: first year in service | year | 1 | Whole years, 1 to 8. Bought and fitted out the year before; in service from the start of this year. | |||||||||
| 24 | Cohort 1: capacity | MW | 20.0 | Megawatts of IT load: the power the GPUs and servers can draw. | |||||||||
| 25 | Cohort 1: GPUs per MW | GPUs | 800 | Falls with each generation as each GPU draws more power. | |||||||||
| 26 | Cohort 1: GPUs and servers, capital cost per MW | $m | 24.0 | ||||||||||
| 27 | Cohort 1: data-hall fit-out, capital cost per MW | $m | 6.0 | Racks, liquid cooling, networking and power distribution inside a leased hall. | |||||||||
| 28 | Cohort 2: first year in service | year | 2 | ||||||||||
| 29 | Cohort 2: capacity | MW | 30.0 | ||||||||||
| 30 | Cohort 2: GPUs per MW | GPUs | 750 | ||||||||||
| 31 | Cohort 2: GPUs and servers, capital cost per MW | $m | 25.0 | ||||||||||
| 32 | Cohort 2: data-hall fit-out, capital cost per MW | $m | 6.2 | ||||||||||
| 33 | Cohort 3: first year in service | year | 3 | ||||||||||
| 34 | Cohort 3: capacity | MW | 40.0 | ||||||||||
| 35 | Cohort 3: GPUs per MW | GPUs | 700 | ||||||||||
| 36 | Cohort 3: GPUs and servers, capital cost per MW | $m | 26.0 | ||||||||||
| 37 | Cohort 3: data-hall fit-out, capital cost per MW | $m | 6.4 | ||||||||||
| 38 | Hours in a year | hours | 8760 | ||||||||||
| 40 | Contracts and backlog | ||||||||||||
| 41 | Cohort 1: share of GPU-hours contracted | % | 85.0% | Take-or-pay: billed whether or not the customer runs them. | |||||||||
| 42 | Cohort 1: contract tenor | years | 3 | ||||||||||
| 43 | Cohort 1: contracted price | $ per GPU-hour | 2.60 | ||||||||||
| 44 | Cohort 2: share of GPU-hours contracted | % | 75.0% | ||||||||||
| 45 | Cohort 2: contract tenor | years | 4 | ||||||||||
| 46 | Cohort 2: contracted price | $ per GPU-hour | 2.40 | ||||||||||
| 47 | Cohort 3: share of GPU-hours contracted | % | 70.0% | ||||||||||
| 48 | Cohort 3: contract tenor | years | 3 | ||||||||||
| 49 | Cohort 3: contracted price | $ per GPU-hour | 2.20 | ||||||||||
| 50 | Contracted hours the customer actually runs | % | 90.0% | Drives power drawn, not revenue. | |||||||||
| 51 | Prepayment at signing, % of contract value | % | 10.0% | Released against revenue evenly over the tenor. | |||||||||
| 52 | Receivables, days of revenue | days | 30 | ||||||||||
| 53 | Days in the year | days | 365 | ||||||||||
| 55 | On-demand | ||||||||||||
| 56 | On-demand utilisation in a cohort’s first year, % of ramped | % | 60.0% | ||||||||||
| 57 | Spot price, Y1 | $ per GPU-hour | 2.70 | Falls each year as newer GPUs arrive. | |||||||||
| 59 | Operating costs | ||||||||||||
| 60 | Power usage effectiveness (PUE) | ratio | 1.25 | Facility energy / IT energy: cooling and losses on top of what the GPUs draw. | |||||||||
| 61 | Power price growth, a year | % | 2.5% | ||||||||||
| 62 | Co-location, per kW of IT load a month, Y1 | $ per kW | 150.00 | Lease of the hall, its power and cooling plant, by the kW. | |||||||||
| 63 | Site staff per MW | FTE per MW | 1.50 | ||||||||||
| 64 | Cost per site employee, Y1 | $k a year | 150.0 | ||||||||||
| 65 | Network, insurance and software per GPU, Y1 | $ a year | 1000 | ||||||||||
| 66 | Corporate costs, Y1 | $m | 30.0 | ||||||||||
| 67 | Growth in space, staff and other costs, a year | % | 3.0% | ||||||||||
| 69 | Depreciation | ||||||||||||
| 70 | GPU useful life | years | 5.0 | An accounting choice, typically 4 to 6 years. Straight line from the first year in service. | |||||||||
| 71 | Fit-out useful life | years | 10.0 | ||||||||||
| 72 | GPU value at the end of Y8, % of cost | % | 10.0% | For the equity return only. | |||||||||
| 74 | Facilities | ||||||||||||
| 75 | Term loan: maximum loan to cost | % | 70.0% | On the cohort’s GPUs and fit-out. | |||||||||
| 76 | Term loan: sizing DSCR on the first contract year | x | 1.3x | ||||||||||
| 77 | Term loan: minimum DSCR covenant | x | 1.2x | ||||||||||
| 78 | Term loan: interest rate | % | 8.5% | ||||||||||
| 79 | Upfront fees, % of debt raised | % | 2.0% | ||||||||||
| 80 | Corporate notes issued at close | $m | 300.0 | ||||||||||
| 81 | Corporate notes: coupon | % | 10.0% | ||||||||||
| 82 | Corporate notes: repaid at the end of year | year | 5 | ||||||||||
| 83 | Revolver commitment | $m | 150.0 | ||||||||||
| 84 | Revolver: interest rate | % | 8.0% | ||||||||||
| 85 | Revolver: commitment fee on the undrawn amount | % | 0.5% | ||||||||||
| 86 | Minimum cash | $m | 50.0 | ||||||||||
| 87 | Refinancing: interest rate on new money | % | 11.0% | ||||||||||
| 88 | Maximum net debt / EBITDA | x | 4.0x | ||||||||||
| 89 | Planned distributions to equity | $m | - | - | - | 100.0 | 300.0 | 150.0 | 350.0 | 250.0 | 200.0 | Paid as planned. Where cash cannot cover them the revolver, then refinancing, does: the gap a sponsor would have to close. | |
Select a cell to see its row, its schedule and the type of cell. Font colours: blue for inputs, black for formulas, green for links to other sheets. Scroll sideways to see every period.
Download
GPU cloud infrastructure model: the workbook
An Excel workbook with working formulas, no macros and no external links: the same model you can work with on this page.
- Browse without an accountPreview every model online, sheet by sheet, with its outputs and checks.
- Free accountWork with every model online: change the inputs, switch the case, see every formula. And every starter workbook, with its key schedules left blank to build.
- L3VLUP Plus$4.99/monthEverything in Free, and one worked model each billing month as an Excel file: every schedule built, formulas live, checks at zero, yours to keep.
- L3VLUP Pro$25/monthEverything in Plus, every worked model in the library, and the Pro applications.
- The worked model
- 10 sheets and 1,985 live formulas, all 9 schedules built and linked, every check reading zero. Change an input in Excel and the whole model moves.
- The starter
- The same inputs and layout with the facility debt: contract-backed term loans and notes left blank for you to build, each blank cell marked. The worked model is its answer.
What the base case says
- Revenue, Y3
- $1,242.5m
- Backlog at the end of Y2
- $2,507.3m
- EBITDA margin, Y3
- 71.3%
- EBIT margin, Y4
- 23.4%
- Cohort 1: GPU book value left when its contract ends
- $192.0m
- Equity at close
- $1,060.3m
- Cohort 2: minimum DSCR
- 1.3x
- Minimum liquidity before refinancing
- $439.1m
- Refinancing needed, Y1 to Y8
- $0.0m
- Equity IRR
- 14.3%
Read from the workbook as served, every input at its default. Periods: Y0, Y1, Y2, Y3, Y4, Y5, Y6, Y7, Y8. The figures are invented and move with whatever you type in.
What this model is
An eight-year model (Y1 to Y8, from financial close in Y0) of a fictional GPU cloud: three cohorts of data-centre capacity, 90 MW in all, commissioned a year apart, rented by the GPU-hour and financed largely with debt secured on customer contracts. Every figure is invented.
Each cohort is a block of GPUs with its own power draw, capital cost, contracts and debt. A share of its hours is sold on take-or-pay contracts at a fixed price for a fixed tenor; the rest is sold on demand at a spot price that falls as newer GPUs arrive. Power follows the hours actually run; the hall is leased by the kW.
Backlog is not revenue and not cash. A contract is booked as backlog when it is signed, becomes revenue only as the hours are made available, and becomes cash later still. The Revenue sheet rolls the backlog forward; the Cash Flow sheet keeps receipts apart from revenue.
The useful life of a GPU is an accounting choice; the tenor of the contract that pays for it is not. When the life runs past the tenor, the rest of the GPU’s cost must be earned at whatever the spot market pays by then. The Capacity sheet shows the book value left when each contract ends, and the margin grid moves the useful life.
Set Scenario on the Assumptions sheet to 2 (upside) or 3 (downside) to move on-demand utilisation, the fall in the spot price and the power price together.
Seats: Private equity, Investment banking, Equity research and hedge funds.
Careers that do this work: infrastructure & real assets, debt capital markets, private credit.
How the schedules connect
Every row in the workbook is tagged with the schedule it belongs to; the inspector above shows the tag when you click a row. These are the schedules this model is made of and where each one lives.
Capacity cohorts: MW, GPUs, utilisation and revenue
Cohorts of compute, GPU-hours, power and revenue
Assumptions, rows 23–38 · Assumptions, rows 56–57 · Assumptions, rows 60–67 · Capacity, rows 6–11 · Capacity, rows 14–19 · Capacity, rows 22–27 · Capacity, rows 30–36 · Revenue, rows 6–17 · Revenue, rows 20–35 · Operating Costs, rows 6–7 · Operating Costs, rows 10–21 · Operating Costs, rows 24–40
Backlog roll-forward: contracts to revenue to cash
Bookings, revenue recognised, backlog, prepayments
Assumptions, rows 41–53 · Revenue, rows 38–46 · Revenue, rows 49–54
Capex and depreciation (PP&E roll-forward)
Opening, plus capex, less depreciation, closing
Assumptions, rows 70–72 · Capacity, rows 39–47 · Capacity, rows 50–64
Facility debt: contract-backed term loans and notes
Delayed-draw loans by cohort, DSCR, notes
Assumptions, rows 75–82 · Financing, rows 19–29 · Financing, row 32 · Financing, rows 35–45 · Financing, row 48 · Financing, rows 51–61 · Financing, row 64 · Financing, rows 67–72
Covenant compliance and headroom
Leverage, interest cover, fixed charge cover, EBITDA cushion
Assumptions, row 88 · Financing, rows 30–31 · Financing, rows 46–47 · Financing, rows 62–63 · Cash Flow, rows 42–45
Liquidity forecast and covenant headroom
Cash, availability, headroom, shortfall
Assumptions, rows 83–86 · Assumptions, row 89 · Cash Flow, rows 6–9 · Cash Flow, rows 12–23 · Cash Flow, rows 26–29 · Cash Flow, rows 32–33 · Cash Flow, rows 36–38
Maturity profile and refinancing
What falls due when, and whether it can be refinanced
Assumptions, row 87 · Financing, rows 75–79 · Cash Flow, rows 30–31 · Cash Flow, row 39
Sources and uses
What the deal costs and who funds it
Financing, rows 6–16 · Cash Flow, rows 55–58
Sensitivity tables
Two assumptions at once, without a data table
Sensitivity, rows 5–8 · Sensitivity, rows 11–16 · Sensitivity, rows 19–26 · Sensitivity, rows 29–35 · Sensitivity, rows 38–66
What you should be able to explain
- Why backlog is not revenue and not cash: a contract is booked when signed, recognised as the hours are made available, and collected later still.
- Why GPU useful life against contract tenor is the central risk: depreciation that outlasts the contract must be earned at a spot price that keeps falling.
- How contract-backed debt works: a loan per cohort, drawn as it is built, sized on the contracted cash flow at a DSCR and repaid over the contract.
- Why power is modelled from GPU-hours actually run, kW per GPU and PUE, and space from the kW leased.
- Why liquidity, cash plus the undrawn revolver, is the number to watch when contracts end and corporate notes fall due in the same years.
What a reviewer looks for
- Adding backlog to revenue, or treating it as cash in a liquidity forecast.
- Depreciating GPUs over a life longer than the contract without asking what the hours will sell for after it.
- Contract-backed debt amortised over the GPU life rather than the contract, so a balance is still owed when the contracted cash stops.
- Power cost from MW of capacity rather than from the hours actually run, or without PUE.
- One revenue line at one price, so the shift from contracted to on-demand revenue as contracts end is invisible.
Conventions this workbook uses
Stated on the cover sheet too. A model is only as trustworthy as the decisions it tells you it made.
- Annual periods. Y0 is financial close and the first build year. Each cohort is bought and fitted out in the year before its first year in service and runs from the start of that year; it stays in service to the end of Y8.
- Contracted GPU-hours are available hours x the contracted share while the contract runs, billed at the contracted price whether or not they are used. On-demand hours are the rest, sold at the ramped utilisation (a share of it in a cohort’s first year) and the spot price.
- Power is GPU-hours run x kW per GPU x PUE x the price per MWh; contracted hours are run at the usage input. Co-location is MW x 1,000 kW x the monthly rate x 12. Space, staff and other costs grow at the cost index.
- GPUs and fit-out are depreciated straight line from the first year in service over their useful lives; a part-year life ends with a part-year charge. Tax is not modelled: depreciation and interest keep a company of this kind in tax losses, so profit is shown before tax.
- Each term loan is drawn in its cohort’s build year, sized to the lesser of loan to cost and the first contract year’s contracted cash flow at the sizing DSCR, and repaid in equal instalments over the contract tenor. The notes are repaid in full at maturity. Interest is charged on opening balances, so nothing is circular.
- Equity is committed at close as the balancing source of the whole build and paid in alongside each cohort’s capital spending. The revolver draws only to hold the minimum cash; anything it cannot supply is a refinancing need, assumed raised as new notes. Liquidity is cash plus the undrawn revolver; liquidity before refinancing deducts new money raised.
- The grids hold financing as sized: the margin grid moves contracted prices and the GPU life in Y4, and the liquidity grid moves on-demand utilisation and the power price in every year, keeping the revolver’s and the refinancing interest as in the selected case.
Build it yourself
The starter workbook
The term loans and the notes have been cleared from the Financing sheet: each cohort’s sizing, draws, interest, instalments, balances, DSCR and covenant headroom, and the corporate notes. Rebuild them from the Assumptions sheet and the contracted cash flow on the Operating Costs sheet. The Checks sheet shows when every facility rolls forward and each loan is repaid by the end of its contract.
Blanks: Facility debt: contract-backed term loans and notes. Download it beside the worked model.
The path around this model
Understand it, drill it, read the build, then apply it to a real company.
Build · Lab · ~12 min
DSCR Sculpting
Sculpt debt service to CFADS at a target DSCR, size the loan as its present value, split a payment into interest and principal, and explain why a level repayment profile fails a project lender.
Build · Lab · ~10 min
Covenant Headroom
Compute the three maintenance ratios from a certificate, find the binding test and its EBITDA cushion, and say whether a given downside breaches.
Build · Lab · ~10 min
Depreciation Waterfall
Forecast depreciation from a capex plan rather than a ratio, read a capex-to-depreciation multiple, and say which part of the spend is buying growth and which is only standing still.
Read · Guide · 12 min
Project Finance Modelling: Cash Flow Waterfalls, DSCR Sizing and the Debt Sculpt
Read · Guide · 11 min
Credit and Covenant Modelling: What a Lender Actually Tests
Read · Guide · 11 min
Financial Modelling Best Practices: The Conventions That Make a Model Auditable
Apply · Skill
Model Audit
Find the errors in a financial model before someone senior does.
Apply · Skill
Diligence Question Builder
Build a prioritised diligence list that attacks the thesis, not a generic checklist.
Vocabulary: DSCR (Debt Service Coverage Ratio), Useful Economic Life, Depreciation & Amortisation (D&A), CapEx, Term Loan (TLA / TLB), Revolving Credit Facility (Revolver), Covenant, Liquidity, Maturity Wall, Sources and Uses, Deferred Revenue, IRR.
Questions about this model
Why is backlog not revenue?
Because nothing has been delivered yet. Here the three cohorts’ contracts are signed in their build years, so by the end of Y2 the backlog is $2,507 million while revenue so far is $1,100 million. Backlog becomes revenue only as each year’s hours are made available, and cash later still. It says how much revenue is secured, not how much has been earned, and it shrinks to nothing as the contracts run off.
Why does the GPU useful life matter so much?
Because the contract that pays for the GPUs is shorter than the life over which they are depreciated. Cohort 1’s contract runs three years against a five-year life, so $192 million of book value is left when the contract ends, to be earned at spot prices that have fallen by then. That is why EBIT turns negative in Y6 while the business is still cash generative. Shortening the life to four years lowers the Y4 EBIT margin from 23.4% to 13.5% without changing a dollar of cash.
How is the debt sized?
Each cohort has its own delayed-draw term loan, drawn in its build year and sized to the lesser of 70% of its capital cost and the loan its first contracted year can carry at a 1.30x DSCR, then repaid in equal instalments over the contract. Cohort 1’s loan is set by cost, the other two by coverage. The contract is the security, so the loan is gone when the contract ends.
Where does the liquidity risk sit?
In the years the contracts end and the corporate notes mature. Liquidity, cash plus the undrawn revolver, is lowest in Y5 at $439 million, the year the $300 million of notes are repaid and the last contracts run off. In the downside, with on-demand utilisation at 65%, spot prices falling 16% a year and dearer power, the planned distributions cannot be paid without about $960 million of new money.
Related models
Infrastructure and project finance · Advanced
Project finance model
Sculpted debt, reserves, lock-up, coverage ratios
Infrastructure and project finance · Advanced
Renewables project finance model
Solar or wind: P50 yield, PPA, merchant tail, debt sizing
Infrastructure and project finance · Advanced
Data centre colocation model
MW by phase, lease-up, rent per kW, power, debt and exit
Credit and restructuring · Advanced
Credit model
Debt capacity, covenants, sweep, maturities, recovery
Questions about this model
Discussions open here soon.