I’ve consulted on a few GPU cluster deals now, and everyone quotes them in dollars per GPU-hour, as if that number were a fact about the hardware. It isn’t.
For a representative 576-GPU build, the same silicon supports a breakeven price anywhere from about $2.78 to about $6.24 per GPU-hour. That’s a more than 2x spread, and it comes from three assumptions almost nobody says out loud: how long you depreciate the asset, what it’s worth at the end, and what it costs to borrow the money to build it. The last one moves the number more than any price gap you’re likely to be negotiating. Below is the arithmetic.
The reference build
8 racks of Nvidia GB300 NVL72: 576 GPUs, roughly 135 kW per rack, about 1.08 MW of IT load. This is a generic reference configuration built from published specs, not a specific site.
| Line | Basis | Amount |
|---|---|---|
| 8x NVL72 racks | ~$5.3M/rack, includes in-rack NVLink switching | ~$42M |
| Scale-out InfiniBand | XDR leaf/spine plus optics, 576 endpoints at 800 Gbps | ~$2M |
| North-south Ethernet | non-blocking, 400 Gbps/node class | ~$0.6M |
| Storage | ~2.3 PB usable at 4 TB/GPU, VAST/WEKA class | ~$1M |
| Management nodes | 6x CPU-only control plane | ~$0.1M |
| Racks, PDU, CDU, cabling | ~$0.5M | |
| Total | ~$46M |
The GPU chassis line is 85 to 90 percent of the bill of materials. So the “everything else” line (networking, storage, power distribution) is still close to $4M ($46M minus $42M). That’s the part people wave away when they say the GPUs are “basically the whole cost.”
What it costs to run, per year
Roughly $4.55M/yr, in four lines:
- Power: ~$0.8M. At 1.3 MW facility load (1.08 MW IT load at PUE 1.2) and $0.07/kWh, that’s 1,300 kW x 8,760 hr x $0.07 = about $0.80M.
- Colo: ~$2.25M
- Managed ops: ~$1.2M
- Transit and spares: ~$0.3M
Sum: $0.8M + $2.25M + $1.2M + $0.3M = $4.55M/yr.
Billable hours. A cluster doesn’t sell all 8,760 hours a year. Maintenance, failed jobs and idle capacity eat into it. At 85 percent utilization: 576 GPUs x 8,760 hr/yr x 0.85 = 4,288,896, call it 4.29M GPU-hr/yr.
$4.55M / 4.29M GPU-hr = $1.06 per GPU-hour in opex alone. That’s before a single dollar of the $46M is paid back.
Four honest breakevens
This is where the number stops being one number. Depreciation life, residual value and cost of capital are all judgment calls, and each one moves the floor price a lot. You can run the same levers (depreciation life, residual value, financing rate) against your own rack count and financing terms with our GPU ROI calculator. Don’t take my reference config on faith.
| Scenario | Capex recovery | + opex | Breakeven $/GPU-hr |
|---|---|---|---|
| 5-yr life, 20% residual, unlevered | $1.72 | $1.06 | ~$2.78 |
| 3-yr full recovery, unlevered | $3.57 | $1.06 | ~$4.63 |
| 3-yr full recovery, 15% cost of capital | $3.57 + ~$1.61 | $1.06 | ~$6.24 |
| 5-yr, 20% residual, 7% cost of capital | $1.72 + ~$0.75 | $1.06 | ~$3.53 |
Sanity check. The capex-recovery lines are plain division: $46M x 80% (after a 20% residual) / 5 years / 4.29M GPU-hr/yr is about $1.72/hr, and $46M / 3 years / 4.29M GPU-hr/yr is about $3.57/hr. Both land within a couple cents of the table. The small gap is rounding carried through the model, not a different assumption.
The ~$2.80 number you keep hearing is a five-year number
The breakeven people quote near $2.80 is the first row of that table. It quietly needs two things: a five-year useful life for the hardware, and 20 percent of the original capex still recoverable (resold or redeployed) at the end of that life.
But the contracts being signed right now are three-year contracts. If the asset has to pay for itself inside the contract that’s paying for it, the floor moves to the second row: above $4.60, unlevered, with no financing cost at all.
That’s a $1.85/hr gap between “the number everyone quotes” and “the number that’s true for a three-year deal,” and all of it comes from a depreciation assumption. Nothing about the hardware changed.
Most arguments about price are really arguments about depreciation that nobody has named. If you’re signing a three-year contract, ask which life the number in front of you assumes. It’s rarely written down.
Cost of capital is the big one
This is the part that should change how you spend your negotiating time.
In this simplified model, the financing cost on top of straight-line capex recovery is the outstanding capex times the annual rate, divided by billable hours: $46M x rate / 4.29M GPU-hr/yr. (I’m costing the full capex balance at the stated rate every year. A real term loan pays down principal, so the gap shrinks a bit over time. The order of magnitude and the direction both hold.)
At 15 percent: $46M x 0.15 / 4.29M = $1.61/hr, matching the $1.61 in the table above. At 7 percent: $46M x 0.07 / 4.29M = $0.75/hr, matching the $0.75 in the table above.
The difference: $46M x (0.15 - 0.07) / 4.29M = about $0.85 to $0.87 per GPU-hour.
That’s bigger than most of the price gaps people fight over line by line in a term sheet. A vendor holding firm on $0.20/hr, or even $0.50/hr, is leaving less on the table than the financing terms behind the deal.
So the best move an operator has is not winning the rate argument. It’s landing a creditworthy anchor customer whose contract is strong enough to unlock cheaper financing.
A lender or lessor prices risk into the rate. A long-term, investment-grade counterparty de-risks the loan more than a better GPU price ever will. If you’re the principal in the room, don’t ask “can you get to $X.XX.” Ask “who is on the other side of this contract, and does that change what a lender will charge you.”
Reserved and on-demand serverless are different products
This is the most common mistake I see: comparing a serverless, pay-by-the-second on-demand rate to a reserved, multi-year bare-metal rate as if they were the same thing.
Modal’s public pricing page, checked in August 2026, lists on-demand serverless rates of $7.10/hr for B300 and $3.95/hr for H100 SXM5. Lambda’s public pricing page, also checked in August 2026, lists on-demand H100 SXM at $3.99 to $4.29/hr depending on cluster size. Those are pay-as-you-go, no-commitment prices, and they’re what shows up first in a Google search.
Reserved, multi-year bare-metal contracts price differently. Compute Exchange’s public reserved-GPU pricing page, published April 10, 2026, states that “H100 reserved pricing currently runs around $1.07 to $1.70 per hour versus on-demand rates of $2.50 to $5.00 per hour,” and separately that a one-year reserved H100 SXM contract runs near $1.89/hr and a three-year reserved H100 PCIe contract near $1.84/hr. The same page quotes on-demand H100 SXM at about $2.99/hr.
Put the two side by side for the same chip: $3.95/hr (Modal, serverless) against roughly $1.07 to $1.89/hr (Compute Exchange, one- to three-year reserved). That’s somewhere between about 2.1x and 3.7x, depending on which reserved tier you pick.
That’s the “2 to 3x apart” gap people quote past each other all the time. Some of it is commitment type, some is chip generation and vendor tier, and it all gets blended into one number with no label. If you ask “what’s the GPU rate” without saying which product, which commitment length and which chip generation, you’ll end up comparing two numbers that were never meant to be compared.
What to check, in order
- Ask what depreciation life and residual value are baked into the quoted breakeven, and whether that life is longer than the contract term you’re actually signing.
- Ask what cost of capital is in the rate, and whether a stronger anchor counterparty on your side of the deal would move it. This is worth more than the negotiated price itself.
- Ask whether you’re comparing a reserved, committed rate to an on-demand, serverless rate, or actually comparing like to like.
You don’t need to be an engineer for any of this. You need to ask which assumptions are hiding inside a single dollar figure, and have the arithmetic in hand before the meeting, not after. The GPU ROI calculator does the same thing interactively. Put in your own contract term, residual and financing rate instead of trusting mine, and the honest breakeven falls out.