Economics

Buy vs rent a GPU cluster: 5-year TCO

A decision guide for a five-year GPU cluster: owning in colocation, a dedicated multi-year reservation, or short-term rental, compared per productive GPU-hour, with the conditions under which owning comes out ahead.

All notesPublished 13 min read
In this note
  1. The short answer, by situation
  2. Five-year cost per productive GPU-hour, side by side
  3. What owning a GPU cluster really costs
  4. When owning a GPU cluster wins
  5. Reserving a dedicated cluster for five years
  6. Renting on demand or short term for five years
  7. Utilization: the number that decides it
  8. Three buyers, three answers
  9. A worksheet for your own numbers
  10. Reserving with Lightyard

For most buyers planning five years of GPU capacity, a dedicated multi-year reservation is usually the better fit than owning, and on-demand rental costs the most per hour. Owning a 32-node HGX B300 cluster (256 GPUs) in leased colocation costs about US$4.84 per productive GPU-hour in our middle case, and US$3.13 to US$7.23 across our low and high cases, once the team, support, failures, insurance, residual value, cost of capital and the months before the cluster does useful work are counted. A 60-month reservation is billed per GPU-hour at a rate fixed for the term; the from-rates are on the pricing page. On-demand B300 had a median list price of US$9.50 per GPU-hour in October 2026, and large multi-node blocks are hard to hold that way for months, let alone years.

Owning comes out ahead for a narrower group than most build cases suggest: buyers that already run GPUs around the clock with their own team, have powered space that's already paid for, can borrow cheaply and will keep the cluster busy for all five years. For them, our model gives about US$2.79 to US$3.14. For everyone else, a new team, leased space and the months of waiting cost more than owning saves on the hardware.

The ownership figures come from our cost model for a 32-node HGX B300 cluster. This guide adds the other two ways to get the same GPUs and puts all three on one unit.

The short answer, by situation

Your situationWhat usually costs least over five years
Need is under a year, or the workload may not lastRent on demand or on a short term
Need is 3 to 5 years, no GPU operations team yetReserve dedicated capacity for the term
Need is 5 years, but could shrink after the first yearReserve; the contract has an exit option
Existing 24/7 GPU team, own powered hall, cheap capital, utilization above 80%Own
Hundreds of nodes, several sites, a decade of demandOwn, and probably build
Data can't leave your control, but no team or spaceReserve single-tenant capacity in your region

The rest of this post shows where those rows come from, with every assumption in a table so you can swap in your own numbers.

Five-year cost per productive GPU-hour, side by side

OptionUS$ per GPU-hourBasis
Own 32 × HGX B300 in colocation, all-in3.13 / 4.84 / 7.23Our estimate, low / mid / high, per productive hour
Own 32 × HGX B200 in colocation, all-in2.97 / 4.52 / 6.69Our estimate, same model with B200 server prices
Own 4 × GB300 NVL72 in colocation, all-in3.10 / 5.18 / 8.32Our GB300 NVL72 cost model
Own 8 × HGX B300, all-in4.51 / 6.76 / 9.79Our 8-node cost model
Own B300: hardware, power and colocation only1.96 / 2.49 / 3.15What most build cases count
Reserve for 60 monthsQuoted per deal; from-rates on the pricing pageEvery contracted hour
B300 on demand, list price9.50 median, 7.40 to 17.807 providers, October 2026
B200 on demand, list price8.19 median, 6.00 to 14.249 providers, October 2026
B200 on demand, executed transactions6.40Spot index, 1 September 2026

On-demand figures are from Uvik's October 2026 index (opens in a new tab) of public list prices and Ornn Data's (opens in a new tab) daily index of executed on-demand rentals. Reservation rates are "from" rates; the quoted rate depends on the platform, size, region and prepayment.

The units differ in one way that matters. Owned figures are divided by productive hours, the ones left after deployment time and failures. Reserved and rented figures are per billed hour. Utilization, the share of those hours you actually keep busy, applies to all of them and is covered further down.

What owning a GPU cluster really costs

The middle case for 32 HGX B300 nodes, from our cost model, spread over its 9.34 million productive GPU-hours:

LineUS$ per productive GPU-hourMiddle-case assumption
Capex: servers, fabric, storage, fit-out, spares2.07US$19.3M; servers at US$475K each
Colocation0.62500 kW at US$210 per kW-month, from mid-deployment
Energy0.21415 kW IT, PUE 1.4, 9 cents per kWh, once running
Support and maintenance0.685% of hardware a year in years 1 to 3, 10% in years 4 and 5
Team0.806.5 FTE at US$250K loaded, from mid-deployment
Cluster management, insurance0.14US$6K per node a year; 0.4% of hardware a year
Residual value at year 5-0.2415% of server cost
Cost of capital0.568% a year, five-year amortizing
All-in4.849-month deployment, 2% of hours lost to failures

Two things in that table surprise most first-time buyers. The servers are only about a third of the five-year cost. And electricity, the line people negotiate hardest, is the second smallest: dropping from 9 to 5 cents per kWh saves about US$0.09 per productive GPU-hour.

For the inputs, GPUsmith (opens in a new tab) reports HGX B300 servers at US$430K to US$550K and a vendor guide (opens in a new tab) DGX B300 at US$400K to US$500K, both mid-2026. Wholesale colocation for 250 to 500 kW in primary North American markets averaged about US$196 per kW-month at the end of 2025, with vacancy at a record 1.4% (CBRE (opens in a new tab)).

For B200 we swapped in new 8-GPU HGX B200 node prices of US$360K to US$470K from American Compute's October 2026 report (opens in a new tab), with US$415K in the middle, and kept every other line the same. That slightly overstates B200, whose 400 Gb/s fabric costs less than B300's 800 Gb/s. The result is US$4.52 per productive GPU-hour in the middle case, only 7% below B300, because most of the cost isn't the server.

The months before productive use

Leased colocation takes 6 to 18 months from decision to first production job, and greenfield two to four years (deployment timeline). The model takes 6, 9 and 12 months out of a five-year life, because capital is committed at the order and the next GPU generation arrives on schedule regardless. Holding everything else at the middle case, 6 months gives US$4.62, 9 months US$4.84 and 12 months US$5.09. Each quarter of delay adds about 5%.

Financing

The model charges 6%, 8% and 10% a year on the capex. That's in line with what GPU-backed lending priced at in 2026. Wing VC (opens in a new tab) lists an investment-grade-rated facility at SOFR plus 2.25%, facilities backed by non-investment-grade customers at SOFR plus 4.5% to 5.5%, and quotes of SOFR plus 8% to 9% for small borrowers without investment-grade customers. With SOFR (opens in a new tab) at 3.97% on 3 November 2026, that's roughly 6.2%, 8.5% to 9.5% and 12% to 13%. Lenders size these loans to amortize before the hardware ages out, so expect to repay principal fast.

Moving the middle case from 8% to 5% takes US$0.21 off the cost per productive GPU-hour. Moving it to 12% adds US$0.29.

Residual value

We credit 25%, 15% and 10% of the server cost at year five. For B200, American Compute forecasts retention of 16% to 46% of the new price after 36 months and 6% to 24% after 48. That argues for planning on the low end. Going from 15% to 5% adds US$0.14 per productive GPU-hour. GPU depreciation and useful life covers the book-life side.

Tax and accounting

Owning has a genuine advantage in the US. IRS Publication 946 (opens in a new tab) says the 100% special depreciation allowance was reinstated for qualified property acquired and placed in service after 19 January 2025. A buyer with taxable income to shelter can deduct the hardware in year one, which our pre-tax model leaves out. Ask your tax adviser what it's worth to you; for a loss-making company it may be worth little until later.

Owned servers are capex. A multi-year dedicated-capacity contract is usually opex, though your auditors should check whether it contains a lease under ASC 842 or IFRS 16.

When owning a GPU cluster wins

Owning wins when the lines that don't depend on the hardware are already paid for. Starting from the US$4.84 middle case, one lever at a time:

LeverChange per productive GPU-hour
An existing 24/7 GPU team absorbs the cluster with 2 extra people-0.56
Your own hall with spare power and cooling, already paid for (no colocation line)-0.62
Servers bought at US$400K instead of US$475K-0.40
Deployment in 4 months instead of 9-0.35
Cost of capital 5% instead of 8%-0.21
Residual value 25% instead of 15%-0.14
Electricity at 5 cents instead of 9-0.09

Put the first, second, fourth and seventh together with 6% capital, and the all-in cost falls to about US$3.14 per productive GPU-hour. Add the cheaper servers and it's US$2.79. Those are our estimates, but the order of the levers is the useful part: team and facility first, then purchase price and time, power last.

Take away either the team or the hall and the same owner lands at US$3.67 to US$3.73, or about US$3.32 to US$3.38 with the cheaper servers. That's close to a five-year reserved rate, so the choice then comes down to the actual quote, how much the exit option is worth to you, and whether you want to run a cluster at all.

Scale helps too, mostly through the team. On the same assumptions, 64 nodes with one more person come to US$4.50, and 128 nodes with 9 people to about US$4.31. Past a few hundred nodes, with your own sites, owning becomes the normal choice, which is what the largest operators do.

The honest summary: owning a GPU cluster beats reserving one when you're already a GPU operator. If you'd be becoming one for this cluster, the first five years carry the cost of learning.

Reserving a dedicated cluster for five years

A reservation gives you single-tenant capacity that the provider operates, at a rate per GPU-hour fixed for the term. At Lightyard the service includes monitoring, replacement of failed parts and running the cluster to an SLA, on the RoCE v2 network. Anything outside the rate is listed in the order form, so add it to the reservation side when you compare; the owned model carries US$0.5M to US$2.0M of storage capex.

Four things change the per-hour comparison with owning.

Dates count from signing. You don't spend months finding space and waiting on hardware. HGX B300 is about 10 to 14 weeks from signing, and larger clusters are handed over in monthly waves of 32 nodes.

A prepayment is required, and it lowers the rate. It is paid into escrow with an independent escrow agent, which releases it in stages as supply, installation and acceptance milestones are met. From the first full month, it is credited in equal parts against each monthly invoice over the term. There is no security deposit. Owning puts the whole outlay in before the first productive hour. Here the prepayment stays with the escrow agent until milestones are met, then comes back as invoice credits.

Utilization is still your risk. Every contracted hour is billed whether the nodes are busy or not, as with an owned cluster. Failures are the provider's to fix under the SLA; read how it treats the hours a failed node is out.

There's an exit option. The contract includes an exit option, on terms set out in your proposal. Part of the commitment can also roll into a newer GPU generation when it ships. An owner who wants out at year three sells servers that American Compute's forecast puts at 16% to 46% of their new price for B200, and still has the colocation lease and the team.

Renting on demand or short term for five years

Short-term rental costs the most per hour and commits you to nothing. For a five-year need it carries two risks that the hourly price hides.

The first is availability. In April 2026 SemiAnalysis (opens in a new tab) reported on-demand capacity sold out across GPU types, half the providers it asked fully sold out when it looked for 8 nodes (64 GPUs) of H100 or H200, and all capacity coming online through August to September 2026 already booked. A 256-GPU cluster on one fabric is rarely something you can hold by the hour.

The second is renewal price. The same report has the one-year H100 contract index rising from US$1.70 per GPU-hour in October 2025 to US$2.35 in March 2026, almost 40% in five months. Ornn's spot index for H100, H200 and B200 rose 70% to 111% between March and September 2026. Renting a year at a time for five years means renegotiating four times in a market that can move that fast in either direction.

The market prices term the way you'd expect. Ornn's analyst marks, as a share of each GPU's one-month term price:

TenorB200B300
1 year97.9%81.3%
3 years69.1%60.9%
5 years53.8%53.8%

The marks are indicators assembled from active deals, not executable quotes, and they're relative to a one-month term price rather than to on-demand list prices. The shape is what matters: a five-year commitment is priced at roughly half of a one-month term, on both B200 and B300.

Utilization: the number that decides it

Owned and reserved capacity cost the same whether it's busy or not, so divide by the share of hours you'll actually use:

US$ per GPU-hour actually used100%80%60%40%
Own 32 × B300, middle case4.846.058.0612.09
Own 32 × B300, low case3.133.915.217.82
Own 32 × B200, middle case4.525.657.5311.29
Any committed rate RRR ÷ 0.8R ÷ 0.6R ÷ 0.4

Against on-demand, the breakeven is the committed rate divided by the on-demand rate. At the US$9.50 median B300 list price, the owned middle case needs 51% utilization and the low case 33%. At the US$7.40 low end of the on-demand range, those become 65% and 42%. For B200 at the US$8.19 median, the owned middle case needs 55%. Divide any reservation quote by the same on-demand rate to get its breakeven.

What utilization to expect depends on the work. A team pretraining and fine-tuning its own small language models usually keeps a training queue full and plans on 80% or more. Inference for one region's office hours lands well below half unless batch work fills the nights. Be honest about which one you are, because utilization moves the answer more than any line in the cost model.

Three buyers, three answers

A model company training small language models. It needs 32 to 64 nodes for pretraining runs in the 1B to 8B range and continuous fine-tuning, has ML engineers but no data center staff, and expects demand for at least three years. Owning means hiring a five-person on-call rotation and waiting 6 to 12 months for the first run. A reservation with an exit option fits.

An inference-heavy software company. Traffic is growing and diurnal, so utilization sits around 50% to 60% with batch work in the troughs. It should commit to the base load for five years and rent the peaks short term. Owning the base load only makes sense if it already runs its own data center footprint.

A regulated enterprise building private AI. It needs 8 to 16 nodes in one region with its data under its own control. At 8 nodes owned, the team alone costs US$1.90 to US$3.72 per productive GPU-hour, which is most of the case against owning at that size. Single-tenant reserved capacity keeps the hardware dedicated to it without the team.

A worksheet for your own numbers

  1. Price the owned cluster all-in: capex plus colocation, energy, support, team, management, insurance and cost of capital, minus a conservative residual. The tables above have every line.
  2. Divide by productive hours: GPUs × 8,760 × 5, minus the deployment months and 1% to 3% for failures.
  3. For each reservation quote, add storage and anything else outside the rate, then divide by the contracted hours.
  4. Divide both by your expected utilization, and test 80% and 60%.
  5. Compare against on-demand at the same utilization, and check whether the block size you need can actually be rented that way.
  6. Price the exit scenario: resale and stranded colocation and team if you own, the exit terms if you reserve.

If your owned figure lands below the reserved rate after step 4, and you already have the team and the hall, own it. If not, reserve the base load and rent the rest.

Reserving with Lightyard

Lightyard reserves single-tenant clusters of 8 HGX nodes (64 GPUs) or one NVL72 rack (72 GPUs) and up for 60 months. Most clusters are 32 to 128 nodes; larger ones are scoped case by case. Rates are per GPU-hour, so they drop straight into the worksheet above; the from-rates are on the pricing page. The prepayment is held in escrow and the contract has exit and rollover options. See how delivery works, contract terms, or send us your numbers and we'll quote against them.

Questions this note answers

Sources

  1. Uvik: GPU price index, on-demand list prices captured 7 October 2026 (opens in a new tab)
  2. Ornn Data: The Economics of Open-Weight Inference (September 2026), term-price marks and spot index (opens in a new tab)
  3. SemiAnalysis: The great GPU shortage, rental capacity (April 2026) (opens in a new tab)
  4. American Compute: HGX B200 server values and residual forecast (October 2026) (opens in a new tab)
  5. GPUsmith: NVIDIA HGX B300 system page (mid-2026 pricing) (opens in a new tab)
  6. Vendor guide: NVIDIA B300 Blackwell Ultra (prices checked July 2026) (opens in a new tab)
  7. CBRE: North America Data Center Trends H2 2025 (February 2026) (opens in a new tab)
  8. Wing VC: How the GPU became collateral (September 2026) (opens in a new tab)
  9. Federal Reserve Bank of New York: Secured Overnight Financing Rate (opens in a new tab)
  10. IRS Publication 946: How to Depreciate Property (opens in a new tab)
Get started

Start with one rack.

Tell us the size and timing you have in mind. We usually reply within a few hours and come back with a written proposal.

sales@lightyard.ai