The 2-Year Math on Renting Out GPU Compute in the US
A reproducible two-year ROI model for US GPU compute rental: the H100/B200 rate ladder, $0.49 per GPU-hour of real operating cost, utilization sensitivity, and the residual-value debate that decides the return.
Every week someone arrives at the same idea with the same spreadsheet: buy a few hundred GPUs, put them in a colocation hall, sell the compute, and pocket the spread between the rental rate and the cost of running the machines. In 2023 the spread was so wide the spreadsheet wrote itself — an H100 rented for $8 an hour and cost a fifth of that to operate. Two years later the same GPU rents for less than a third of that number, the purchase price of the card has not fallen, and the entire question of whether a two-year deployment returns its capital has become a genuine modeling exercise rather than a formality[1].
This article works that exercise honestly. It prices the market you would be selling into, itemizes what it actually costs to put one GPU to work in a US colocation facility, builds the full two-year P&L for both an H100 fleet and a B200 fleet, runs the utilization sensitivity that decides everything, and deals seriously with the single most contested variable in the whole model — what the hardware is worth on the day the two-year term ends. All of the arithmetic is reproducible from the cited market data.
The market you would be selling into
The first thing to understand about 2026 GPU rental pricing is that it is at least three markets wearing one label. An H100 can be had from a GPU marketplace for $1.49 an hour, from a specialist neocloud for roughly $2.31 on median, and from a hyperscaler for a median of $6.11 — an eightfold spread for functionally identical silicon[1]. A small operator with a colocation footprint is competing in the first two tiers, never the third.
The second thing to understand is the direction of each tier. Spot and on-demand pricing spent two years falling — the tracked H100 median fell 59% from $8.00 in mid-2023 to $3.29 by late 2026, with an October 2025 trough near $1.70 on the one-year contract index[2]. But committed capacity moved the other way: one-year contract pricing rose roughly 40% from that $1.70 low to $2.35 by March 2026, because on-demand capacity sold out across nearly every GPU generation and holders of locked-up compute refused to release it[1][3]. The Silicon Data index sat near $2.53 per GPU-hour by mid-2026[1]. If you are planning to rent compute out, that split market is the central fact: your realistic realized rate is a contract rate in the low-$2 range for H100-class capacity, not the headline on-demand number and certainly not the hyperscaler number.
| Channel (H100, per GPU-hour) | Rate | What it means for an operator |
|---|---|---|
| Marketplace floor (Vast.ai-type) | $1.49–2.01 | The price a pure commodity seller is pushed toward |
| 1-yr contract, Mar 2026 | $2.35 | The realistic committed rate for new capacity |
| Silicon Data index, mid-2026 | ~$2.53 | Benchmark realized rate, volatile |
| Neocloud on-demand median | ~$2.31–2.99 | Where uncontracted inventory clears |
| Hyperscaler on-demand | $6.11–6.98 | Not a market a colo operator sells into |
Table 1 — The H100 rental ladder as of mid-2026, compiled from market indices and provider listings [1][3]. The 3.3–4.6× spread between tiers is margin structure, not product difference.
Blackwell pricing sits one tier up and one generation tighter. B200 on-demand has ranged from $3.70 on neoclouds to $14.24 on AWS, with a full published span of $2.69–16.11 per GPU-hour depending on provider and commitment[4][5]. Silicon Data's realized B200 tracking ran just under $5.00 in January 2026 and rose to $5.50–5.80 by August — a 50–80% appreciation inside eight months, driven by inference demand and constrained supply[6]. CoreWeave's reserved rate for committed B200 capacity works out near $3.44 per GPU-hour, which is the number an operator should use when modeling a contracted two-year book[4].
What it costs to put one GPU to work
The capex side is well documented. A new H100 80GB trades around $31,000, and a complete 8-GPU HGX server — chassis, CPUs, NICs included — runs $250,000–320,000[7]. B200 ships only soldered to an 8-GPU HGX baseboard, and a complete server runs $400,000–500,000, with NVIDIA's own DGX B200 appliance listing near $515,000[8][6]. On top of the servers sit the things nobody's spreadsheet includes on the first pass: InfiniBand fabric for multi-node work at $50,000–150,000 per 8-node pod, racks, PDUs, and for Blackwell, liquid-cooling-ready infrastructure at the rack level[9].
The opex side is where colocation pricing earns its keep. GPU-dense colocation — 20–100 kW per rack, liquid cooling in the B200 case — quotes at $150–250 per committed kW per month in US markets, with Dallas-type metros at $140–240 and Northern Virginia at $180–300[10][11]. Metered power adds $0.04–0.09 per kWh depending on market. And the headline rate understates the real bill by 25–30% once cross-connects, power billing model, and remote hands are added — one Colorado example worked a $78/kW headline into a materially higher all-in number[11].
Put together, the per-GPU-hour operating cost of an H100 in a mid-market US colo works out like this:
That last line is the article's first hard conclusion: at the March 2026 contract rate and 85% utilization, the operating margin alone takes about 2.7 years to repay the capital — so a two-year window does not close on operating cash flow for an H100 fleet. Everything hinges on what happens next, and on one contested number.
The two-year P&L for both fleets
The table below is the full model, run per 8-GPU node with identical assumptions on both sides: Dallas GPU colocation at $190/kW-month with the +27.5% real-bill adjustment, ERCOT power at $0.05/kWh, shared engineering staffing amortized at $3,000 per node per year, and an average power draw of 55% at idle plus 45% of full draw scaled by utilization. Revenue assumes the fleet sells every busy hour at the stated rate. Residual value is credited at exit.
| Scenario (2-yr hold) | Rate / util | CapEx | Revenue | OpEx | Operating ROI | ROI with resale |
|---|---|---|---|---|---|---|
| H100 base | $2.35 / 85% | $296,000 | $279,970 | $74,961 | −30.7% | +24.2% |
| H100 bull | $2.53 / 92% | $296,000 | $326,236 | $75,248 | −15.2% | +39.7% |
| H100 conservative | $2.00 / 70% | $296,000 | $196,224 | $74,346 | −58.8% | −3.9% |
| H100, no resale credit | $2.35 / 85% | $296,000 | $279,970 | $74,961 | −30.7% | −30.7% |
| B200 base | $5.61 / 85% | $475,000 | $668,353 | $92,812 | +21.2% | +140.8% |
| B200 bull | $5.80 / 92% | $475,000 | $747,894 | $93,129 | +37.9% | +157.5% |
| B200 conservative | $4.50 / 75% | $475,000 | $473,040 | $92,359 | −19.9% | +39.1% |
| B200, no resale credit | $5.61 / 85% | $475,000 | $668,353 | $92,812 | +21.2% | +21.2% |
Table 2 — Two-year node P&L. Resale credits use Silicon Data's September 2026 going-concern residual values ($20,308 per H100, $71,057 per B200) [12] for the base/bull rows and purchase parity for the B200 conservative row. The H100 conservative row also uses those residuals — and still loses money.
Three conclusions fall out of that table, and they are starker than most pitch decks admit.
First, the H100 is a residual-value trade, not a rental business. At contract rates, two years of selling H100 compute returns only about 69% of the capital even before opex is counted against it — operating ROI is −30.7%. The fleet returns its money only because Silicon Data's model values a four-year-old H100 at $20,308, roughly 58% of new price, far above what a 3-to-5-year straight-line depreciation schedule would predict[12]. If you believe that residual, the two-year ROI is +24.2%. If you don't — if the used-H100 market repeats the $40,000-to-$6,000 collapse that one market participant documented for early buyers[13] — the same deployment loses 31% of its capital. The operating business is identical in both worlds; the entire return lives in the exit assumption.
Second, B200 is the only fleet where the operating business stands on its own. At the realized August 2026 rate of $5.61 and 85% utilization, a B200 node repays its capex in 19.8 months and exits the two-year term +21.2% on operating cash flow alone — before any resale credit. The premium rental rate more than covers the premium hardware and the liquid-cooling-ready colo tier. That is the structural difference between selling last generation's commodity FLOPS and selling this generation's scarce FLOPS.
Third, utilization below roughly 75% takes the operating business negative for both fleets at any rate a small operator can realistically command. The sensitivity grid makes the point without mercy:
| 2-yr operating ROI | $1.70/GPU-hr | $2.00 | $2.35 | $2.53 | $3.00 |
|---|---|---|---|---|---|
| 60% utilization | −76.7% | −68.2% | −58.2% | −53.1% | −39.8% |
| 70% utilization | −68.8% | −58.8% | −47.2% | −41.3% | −25.7% |
| 85% utilization | −56.9% | −44.8% | −30.7% | −23.5% | −4.6% |
| 95% utilization | −49.0% | −35.5% | −19.8% | −11.7% | +9.5% |
Table 3 — H100 fleet, two-year operating ROI excluding resale, by utilization and realized rate. Exactly one cell in the grid is positive. This is what "the spread collapsed" looks like in a table.
Those numbers reconcile with the wider literature once you adjust horizons. An independent payback model run at 100% utilization and $2.35 — with zero opex, which is flatteringly wrong — puts H100 payback at 2.33 years[13]. A hosting-focused analysis using a $0.35 per GPU-hour all-in hosting fee and a $2.50 rental rate puts the break-even at roughly 14,000 GPU-hours: about 19 months at 100% utilization, four years at 40%[14]. A build-versus-buy comparison puts the colocation crossover near 20 months at full utilization and concludes that below about 65% utilization, renting wins on pure economics[9]. Our model is more conservative than all three because it carries the real-bill colo premium and staffing — which is the point. The optimistic versions of this math are the ones that skip line items.
Why the "breakeven utilization" you were quoted is wrong. A widely shared rule of thumb says an owned H100 pays for itself above ~85% utilization on two-year amortization at $2.50 rental[7]. That arithmetic treats the rental rate as pure margin and ignores colo, power, and staffing. Carry the $0.49 per GPU-hour of real opex and the two-year, at-any-utilization operating breakeven for an H100 disappears entirely — even 100% utilization only shortens the 19,891-hour repayment to about 29.5 months. Anyone showing you a two-year H100 breakeven below 85% utilization is not counting the building.
The residual value debate is the whole ballgame
Because operating cash flow does not close a two-year H100 deal, the exit assumption deserves its own scrutiny rather than a single line in a sensitivity table.
The bull case is Silicon Data's residual value model, which values accelerators as revenue-producing infrastructure rather than depreciating hardware: B200 at approximately $71,057 — 158% of its original cost baseline, up 14.4% during 2026 — H100 at $20,308, and even the 2020-era A100 near $5,000[12]. The model is a discounted cash flow on remaining rental earning power, adjusted for utilization expectations, operating costs, interest rates, and new architecture arrivals. The same logic shows up in behavior: CoreWeave has contracted A100 capacity through 2029, keeping 2020 hardware commercially productive for nearly a decade, and AWS still rents V100s from 2017[12].
The bear case is equally empirical. The same used market that now quotes H100 residuals above straight-line schedules watched early-generation cards fall from near $40,000 in late 2023 to as little as $6,000 by mid-2026 — an 85% drawdown that any residual-dependent model has to survive[13]. One infrastructure economics guide tells owners to model GPUs as losing 50–70% of value within two years as new generations land, and to treat anything better as upside[9]. And the two cases are not actually contradictory: Silicon Data's number is explicitly a going-concern valuation — what the card is worth while it keeps earning — not a clearing price for a pallet of used cards in a soft market[12].
The disciplined way through is to underwrite the deal so it survives the bear case, and to note what the structure of the market does to the odds. The same supply constraints that pushed B200 residuals above par — HBM shortage, 3.6-million-unit Blackwell order backlog, allocation queues stretching to 30+ weeks[5][12] — are also what keep rental rates elevated. Residual value and rental rate are the same trade viewed from two dates. A two-year model that assumes today's rental rate persists and assumes the residual collapses is internally inconsistent, and so is the reverse. Pick a consistent world and price both years in it.
Scaled to a hall: the 1 MW picture
Most readers of this blog do not buy single nodes, so here is the base case scaled to a 1 MW IT-load hall: 89 HGX nodes (712 H100s) at 10.4 kW each with an 8% allowance for fabric and spares.
| 1 MW H100 hall, 2-year term | Value |
|---|---|
| Nodes / GPUs | 89 / 712 |
| Total CapEx (nodes + fabric + racks) | $26.3M |
| 2-year revenue at $2.35 / 85% | $24.9M |
| 2-year colo + power + staffing | $6.7M |
| Operating cash position at exit | −$8.1M |
| Residual value at $20,308/GPU | $14.5M |
| Total 2-year position (with resale) | +$6.4M (+24.2% on capital) |
Table 4 — The same trade at hall scale. The −$8.1M operating trough at month 24 is the financing reality that has lenders asking AI tenants for 30–40% down and three-year compute reserves instead of one[12].
Notice what the hall-scale view exposes that the node view hides: the model assumes every busy hour is sold at a contract rate for 24 consecutive months. Producing that revenue requires a sales operation that signs committed buyers before the metal arrives — which is why the operators who make this work tend to be the ones who secured offtake first and shopped for hardware second, not the reverse. It also assumes the facility delivers: committed power on schedule, liquid-cooling-ready rows where the roadmap needs them, and the environmental disciplines — power quality, battery backup, burn-in — that keep 712 GPUs earning through their full contracted life. Those disciplines are the hardware half of this blog: the UPS sizing that rides through the events, the burn-in testing that removes the early failures before a tenant's SLA clock starts, and the distribution architecture — increasingly the 800 VDC model — that decides how many of those $37,000-per-GPU assets the building can feed.
GPU-ready racks staged and awaiting hardware. The racks, PDUs, and containment are a rounding error in the model above — under 2% of capex — but they are the part that determines whether the other 98% ever earns.
What breaks the model, ranked
- Rate decay you don't control. H100 rental fell 59% in about three years; AWS cut H100 on-demand 44% in a single June 2025 revision[2][3]. A two-year model at a flat $2.35 is a bet that contract rates hold, which the March 2026 rebound supports and nothing guarantees. Contract with escalation floors, or model a declining curve.
- Utilization you assume instead of sign. Production telemetry across 23,000 clusters measured average GPU utilization at 5% — at which duty cycle a committed $3.44 B200 rate bills out at $68.80 per useful GPU-hour[4]. The model's 85% is a sales outcome, not an operating default.
- The 25–30% real-bill gap. Committed-kW billing means idle racks still invoice at full rate; cross-connects and remote hands stack on top[11]. Underwriting the headline $/kW is the most common single spreadsheet error.
- Generation transition mid-term. Migrating a training pipeline to a new architecture costs $18,000–27,000 in engineering alone before retuning[5] — and the migration pressure arrives on the seller's side of the market exactly when the new generation's rental premium tempts buyers away.
- Term-structure mismatch. Colo pricing favors 36-month terms — 12-month commitments run roughly a 12% premium[11] — while GPU rental contracts are drifting from one-year toward three-year commitments with heavy upfronts[12]. Matching a 24-month hardware payback to 12- and 36-month counterparty terms leaves exposure on both ends.
- Residual collapse in a soft exit window. The bear case above. Sell into strength or don't sell; model both.
The bottom line
Run honestly, the two-year arithmetic for renting out GPU compute in the US says three things. First, at 2026 contract rates the business is a thin-margin infrastructure trade, not a 2023-style gold rush: a mid-market operator's all-in cost to deliver an H100-hour is about $0.49 against a $2.35 realization, and after capex amortization the two-year operating return on H100 capital is negative at every utilization a newcomer is likely to achieve. Second, the current-generation premium is doing all the work — B200 economics clear their capex in under 20 months at 85% utilization and stand positive on operating cash alone, which is precisely why Blackwell supply is allocated three-million units deep and why the rental rate is 2.4× the H100's. Third, the exit assumption is not a footnote; it is the return. Between Silicon Data's going-concern residuals (+24% two-year ROI on an H100 fleet) and the documented 85% used-card drawdown (−31%), the identical operating business spans a 55-point outcome range on one line of the spreadsheet.
The practical playbook that survives every scenario: contract the offtake before buying the metal, underwrite utilization at what is signed rather than hoped, carry the real colo bill, prefer current-generation capacity where the rate spread funds the capex, hold to a residual assumption you can defend with comps, and put the facility disciplines in place before the SLA clock starts. Do all of that and the two-year return is modest but real — and it is earned in the sales contract and the building, not in the spreadsheet.
DC Build X supplies the racks, power distribution, containment, and cooling infrastructure that GPU rental economics run on, and we work with operators from first power study through first contracted megawatt. Request a quote with your target fleet and market — we will help you pressure-test the facility side of the model before you sign the GPU PO.
For related reading: the pre-deployment burn-in disciplines that protect the SLA behind your rental contract, the UPS battery sizing that keeps a GPU pod alive through transients, and the 800 VDC distribution architecture that determines how much of your hall the next generation of racks can occupy.
References
- Coloprice, "H100 Rental Prices in 2026: History, Benchmarks, Outlook" — Silicon Data H100 Rental Index, SemiAnalysis 1-year rental index, Compute Exchange auction data. coloprice.com
- Coloprice, "Price History" — GPU rental and colocation series since 2023, seeded from Silicon Data, SemiAnalysis, Introl, CBRE and JLL. coloprice.com
- GPUSmith, "Nvidia H100 Rental Price History and H200 Cost Trends 2026" — provider spread, AWS P5 44% cut, AIMultiple index, SemiAnalysis contract index. gpusmith.com
- Beri, "GPU Hour Pricing: On-Demand, Reserved, Committed, Spot" — 64-GPU B200 purchasing table, break-even duty cycles, Cast AI utilization measurement, September 2026. beri.net
- GPUSmith, "H200 vs B200 vs B300 vs MI325X: Full 2026 GPU Comparison" — B200 pricing spread, backlog, TDP, migration engineering cost. gpusmith.com
- Silicon Data B200 rental tracking via Duck IT Tech News and PChome, September 2026 — realized rates January vs August 2026. pchome.megatime.com.tw
- Bex, "H100 Rentals Crashed 70%. Should You Still Buy the GPU?" — purchase prices, utilization breakevens, hyperscaler spread, September 2026. bex.co
- Tech Insider, "NVIDIA Blackwell GPU Pricing: B200, B300 and DGX Cost Breakdown" — HGX/DGX system prices, street quotes. tech-insider.org
- Rax.ae, "GPU-as-a-Service Economics: Pricing, Build vs Buy & ROI for AI Infrastructure" — build-vs-buy table, IB fabric costs, staffing, depreciation guidance. rax.ae
- GPU Lease Index, "What Is the Average Colocation Price per kW? (2026 Calculator)" — retail/wholesale/GPU-dense rates by market, worked 100 kW example. gpuleaseindex.com
- QuoteColo, "GPU Colocation (AI & HPC Hosting) — Pricing, Providers, Power Availability" — market table, real-bill gap, term premiums. quotecolo.com
- Silicon Data GPU Residual Value model, September 2026, via Duck IT Tech News — B200 at 158%, H100 at $20,308, A100 near $5,000, going-concern DCF methodology, financing terms. duckittech.com
- AI Tech Connect, "Reserved, On-Demand or Spot: A GPU Commitment Model" — owned-payback model, 60% utilization rule, used-H100 drawdown warning. aitechconnect.in
- MillionMiner, "Cloud GPU Pricing in 2026: Rent vs Buy, and the Break-Even Math Nobody Shows You" — hosting fee model, 14,000 GPU-hour breakeven, 40% utilization rule. millionminer.com