A 100 H100 cluster delivers payback in roughly 2 to 5 years and a 3-year IRR of 7 to 33 percent. The range is that wide because the answer depends almost entirely on what cloud cost you'd otherwise be paying. Compared to hyperscaler on-demand at $3.50/GPU-hour, the math is decisively positive. Compared to reserved 3-year long-tail capacity at $1.80/GPU-hour, the math is marginal and depends on utilization staying high. This guide walks the three return metrics every CFO asks for on a GPU CapEx request, with formulas, current 2026 numbers, and a worked example you can plug your own utilization and cloud baseline into.
Why three return metrics and not just one
Most GPU CapEx requests get evaluated on a single number. That's a mistake. Payback period, IRR, and NPV each answer a different question, and a complete CFO memo includes all three.
Payback period tells you how long the cluster takes to recover its cost. Fastest number to communicate to a board, easiest to sanity-check, weakest because it ignores time value of money.
IRR tells you the annualized return the investment produces against the rent counterfactual. Right number for ranking GPU CapEx against other capital uses (product investment, debt paydown, equity buyback).
NPV tells you whether the investment is worth doing at your cost of capital. Most defensible single output because it explicitly prices the time value of money at a rate the business has already committed to elsewhere.
Most institutional buyers compute all three and require payback under 30 months, IRR above 20 percent, and positive NPV at the firm's WACC before approving GPU CapEx.
GPU payback period: how long until the cluster pays itself off
Payback is the simplest of the three. Cumulative cash savings (rent avoided minus owning costs) divided into the upfront CapEx.
The formula:
// text
Payback (months) = Net CapEx / (Annual savings / 12)
Net CapEx = Hardware CapEx − expected residual value
Annual savings = Annual rent cost avoided − annual owning operating costWorked example: 100 H100 cluster base case
// text
Hardware CapEx: $4.5M (100 H100 SXM5, 13 8-GPU HGX servers, 2026 OEM)
Residual at year 3: $1.1M (25% of CapEx, per H100 depreciation curve)
Net CapEx: $3.4M
Operating cost: $0.41/GPU-hr (power + colo + ops)
Utilization: 75% (657,000 useful GPU-hours/year)
Annual owning opex: $269,370The annual savings depend on what you're escaping. Three scenarios:
| Scenario | Rent rate | Annual rent equiv. | Annual savings | Payback |
|---|---|---|---|---|
| Hyperscaler on-demand | $3.50/hr | $2,299,500 | $2,030,130 | 20 months |
| Mixed cloud (60/40) | $2.50/hr | $1,642,500 | $1,373,130 | 30 months |
| Reserved long-tail | $1.80/hr | $1,182,600 | $913,230 | 45 months |
The payback verdict is straightforward. Owning beats hyperscaler on-demand decisively. Owning beats mixed cloud workloads inside a typical 3-year horizon. Owning does not beat optimized reserved cloud on payback grounds at this utilization. If the buyer is already on reserved long-tail and renewing, the CapEx case has to come from elsewhere (sovereignty, capacity monetization, workload-specific performance).
For the full cost breakdown that produces the $0.41/GPU-hour operating number, see colocation economics and the 100 H100 Cluster TCO.
GPU IRR: what return does the cluster deliver
IRR is the discount rate at which the cluster's net cash flows equal zero. Solved iteratively, not in closed form.
The setup. Treat GPU CapEx as a project with these cash flows:
// text
Year 0: −Hardware CapEx
Year 1: +Annual savings (rent avoided − owning operating cost)
Year 2: +Annual savings
Year 3: +Annual savings + Residual recoverySolve for r such that:
// text
−CapEx + Σ [Annual_savings_t / (1+r)^t] + Residual/(1+r)^N = 0Most teams use Excel's IRR or XIRR. Treasury teams running real models use full DCF.
IRR matrix for the 100 H100 base case (3-year horizon, 25% residual):
| Utilization | vs Hyperscaler ($3.50/hr) | vs Mixed cloud ($2.50/hr) | vs Reserved long-tail ($1.80/hr) |
|---|---|---|---|
| 75% | ~26% IRR | ~8% IRR | Fails (below cost of capital) |
| 85% | ~33% IRR | ~13% IRR | ~4% IRR |
| 95% | ~40% IRR | ~18% IRR | ~9% IRR |
The pattern matters. IRR moves more with cloud baseline than with utilization. A buyer escaping hyperscaler on-demand gets attractive returns even at moderate utilization. A buyer already on optimized reserved capacity needs near-perfect utilization just to clear a reasonable cost of capital. This is consistent with the buy-vs-rent breakeven framework, which puts the threshold at 75 to 80 percent utilization against reserved cloud.
Second sensitivity worth flagging. IRR is sensitive to residual value. A 10-percentage-point change in assumed residual moves the 3-year IRR by 3 to 5 points. The H100 depreciation curve uses 25 percent residual at year 3 as base case, but residual depends on the secondary market when the next generation hits volume. Stress-test it at 15 and 35 percent before signing.
GPU NPV: what is owning worth at your cost of capital
NPV converts the cluster's cash flows into a single present-value number using your firm's discount rate.
The formula:
// text
NPV = −CapEx
+ Σ [Annual_savings_t / (1+k)^t]
+ Residual / (1+k)^N
where k = your firm's discount rate (WACC or hurdle rate)The discount rate is the lever most teams underuse. Treasury has already picked the firm's weighted average cost of capital, and that's the rate the GPU project should use.
Typical 2026 cost-of-capital ranges by buyer type:
- Established enterprises, investment-grade balance sheets: 6 to 8 percent
- Late-stage AI companies with strong revenue: 10 to 14 percent
- Earlier-stage venture-backed AI buyers: 15 to 20 percent
NPV sensitivity for the 100 H100 base case at 75% utilization, $2.50/hr mixed cloud baseline:
| Discount rate | NPV |
|---|---|
| 6% | ~ +$560,000 |
| 10% | ~ +$90,000 |
| 12% | ~ −$130,000 |
| 15% | ~ −$440,000 |
| 18% | ~ −$700,000 |
Two takeaways.
First, NPV is positive at 6 to 10 percent cost of capital, negative above 12 percent. A venture-backed AI buyer with a 16 percent cost of capital should not own at 75 percent utilization against a mixed cloud baseline. An established enterprise at 7 percent cost of capital should. Same hardware, different answer.
Second, the gap between IRR and the firm's cost of capital is the most defensible NPV summary. If IRR comes in 5 to 10 points above cost of capital, the investment clears. If IRR is within 2 points, the investment is a wash and other factors (sovereignty, operational appetite, balance sheet treatment) decide.
Which return metric should you use
Payback for the board summary. IRR for the comparison against other capital uses. NPV for the final approval decision.
For most institutional buyers, the decision rule is:
- Approve if payback under 30 months, IRR at least 5 points above cost of capital, and NPV positive at WACC
- Two of three is a gray zone needing additional justification
- Zero or one of three is a no
A useful framing. The GPU CapEx case isn't an isolated investment, it's a refinancing of a cloud bill. The CFO is choosing between paying $1.6M to $2.3M annually to a cloud provider with no terminal value, or paying $4.5M upfront for $1.1M residual and a lower annual operating cost. Both paths produce the same compute. The question is which path has the better return profile.
How financing structure changes the ROI numbers
The metrics above assume cash purchase. Most institutional buyers finance the cluster. Financing structure affects payback, IRR, and NPV differently.
Equipment loan at 10% APR over 5 years. Reduces upfront cash outflow, adds interest expense over the loan term. Payback extends 4 to 6 months. IRR drops 3 to 6 points because the firm is leveraging at a rate above the project's unlevered return. NPV drops modestly because the interest deduction recovers part of the financing cost.
Capital lease. Similar mechanics to an equipment loan. Slight tax timing differences but no material IRR or NPV impact versus a loan at the same rate.
Operating lease. Highest cost financing path. Off-balance-sheet treatment is partial under current accounting standards. Payback extends 8 to 12 months. IRR drops 8 to 12 points. NPV often flips negative even when cash purchase NPV is comfortably positive. Operators that fall back to operating leases frequently end up worse off than if they had stayed on reserved cloud.
Full structure-by-structure comparison: financing AI compute.
When the ROI math fails
Owning is the wrong answer in several cases. The math points to four:
- Utilization will be variable or below 65 percent. Even against hyperscaler on-demand, payback extends past the 3-year horizon and IRR drops below cost of capital for most buyers.
- The buyer is already on reserved long-tail at $1.30 to $1.80/GPU-hour. Annual savings shrink to the point where IRR rarely clears 10 percent.
- Cost of capital exceeds 18 percent. Venture-backed buyers in this band almost never see positive NPV unless utilization is near 95 percent.
- Holding period is shorter than 24 months. CapEx amortization can't catch up to rent savings inside two years against any reasonable cloud baseline.
In all four cases, the right answer is to stay on cloud and revisit the question when utilization, cost of capital, or holding period improves.
For the broader decision framework: buy vs rent GPUs.
Frequently Asked Questions
How do you calculate ROI on a GPU cluster?
GPU cluster ROI is measured with three metrics: payback period (months until the cluster pays itself off versus renting), IRR (annualized return over the holding period), and NPV (present value of savings versus rent at your cost of capital). Payback is the fastest read, IRR captures time value, NPV picks a winner once you commit to a discount rate.
How long is the payback period on a 100 H100 cluster?
Payback ranges from 20 to 45 months on a $4.5M, 100 H100 cluster at 75 percent utilization, depending on the cloud cost being avoided. Against hyperscaler on-demand at $3.50/GPU-hour, payback is roughly 20 months. Against mixed cloud at $2.50/GPU-hour, payback is roughly 30 months. Against optimized reserved long-tail at $1.80/GPU-hour, payback extends past the 3-year useful life.
What IRR does a GPU cluster deliver in 2026?
3-year IRR on a 100 H100 cluster at 75 to 85 percent utilization lands between 8 and 33 percent depending on the cloud baseline. Against hyperscaler on-demand, IRR is 26 to 33 percent. Against mixed cloud, IRR is 8 to 13 percent. Against reserved long-tail, IRR falls below 5 percent and the case fails for most buyers.
How do I calculate NPV for a GPU investment?
NPV equals the discounted annual savings versus renting, plus the discounted residual value, minus the CapEx. Use your firm's WACC as the discount rate. At 6 percent WACC, a 100 H100 cluster at 75 percent utilization against mixed cloud produces NPV of approximately +$560,000. At 12 percent WACC, NPV turns negative. The breakeven WACC for the base case is roughly 11 percent.
Which return metric should I use to evaluate GPU CapEx?
Use payback for the board summary, IRR for ranking against other capital uses, NPV for the final go/no-go decision. NPV is the most defensible because it explicitly prices cost of capital. Most institutional buyers require payback under 30 months, IRR at least 5 points above cost of capital, and positive NPV before approving GPU CapEx.
How does financing structure affect GPU ROI?
Equipment loan financing extends payback by 4 to 6 months and reduces IRR by 3 to 6 points versus cash purchase. Operating leases extend payback by 8 to 12 months and reduce IRR by 8 to 12 points, frequently turning a positive NPV case negative.
Does cloud rental ever beat owning on ROI?
Yes. When utilization is variable or below 65 percent, fleet size is below 20 GPUs, the buyer is already on optimized reserved long-tail capacity, or cost of capital exceeds 18 percent. In those cases reserved or on-demand cloud produces a better return profile because the buyer avoids the CapEx drag and the technology obsolescence risk.
Methodology
Pricing in this article is derived from Mercatus GPU Index, which tracks H100 on-demand and reserved cloud pricing across 30+ providers globally, refreshed daily. Hardware CapEx uses published OEM pricing for H100 SXM5 servers. Operating costs assume U.S. wholesale colocation at $0.10/kWh, PUE 1.4, and standard ops loading. Residual values use Mercatus depreciation tracking, with 25 percent base case at year 3. Last verified: May 2026.
Stop modeling GPU ROI on last quarter's pricing. Mercatus GPU Index publishes real-time H100, H200, and B200 pricing across 30+ cloud providers, broken out by region and reservation term. Plug current reserved rates into the formulas above and rerun payback, IRR, and NPV before your next board review.