In September 2026 a new RTX 5090 has a street median around $4,800, far above the $1,999 launch MSRP. The same card rents for about $0.69–$0.99 per hour. The sticker is not the decision. The decision is how many hours you will actually light the GPU. The tables below price three patterns—experiments, six hours on workdays, and 24/7—over 3 months, 1 year, and 3 years, with payback that includes electricity.
Fix the workload before you buy
A consumer card and a data-center GPU are not substitutes. An RTX 5090 has 32GB. It fits quantized mid-size inference, image generation, and small LoRA runs. An H100 80GB adds ECC and MIG; PCIe and SXM are different products. The extra memory is what 70B-class weights and larger inference batches actually use. Comparing a 5090 purchase price with an H100 hourly rate treats “cheaper” as “the same job.”
The other split is easier to miss. If inference goes through a GPT, Claude, or Gemini API, the GPU line can be $0. How subscriptions and usage stack is covered in the real monthly AI programming bill and OpenRouter model prices. This article only prices the case where weights must sit on a GPU you control.
Three utilization bands (one definition throughout)
- Experiments: about 10 hours a week, 40 hours a month. 120 hours in 3 months, 480 in 1 year, 1,440 in 3 years.
- Daily development: 6 hours × 22 workdays = 132 hours a month. 396 / 1,584 / 4,752 hours.
- 24/7: 730 hours a month (365×24÷12). 2,190 / 8,760 / 26,280 hours.
September 2026 prices and power
These are planning prices, not a quote, and they exclude tax. Rental anchors are RunPod’s public rates: RTX 5090 Community Cloud $0.69/hr (stable since early September 2026), Secure Cloud $0.99/hr, and H100 PCIe Secure Cloud $2.89/hr (rate card checked 2026-09-07). Community floors such as Vast.ai list a 5090 near $0.45/hr. Those instances can be interrupted, so they appear only in the sensitivity note, not in the main tables.
On the buy side, a new 5090 is $4,800 (public-listing median about $4,750–$4,900 in mid-September 2026). The base case assumes you already have a host and only add the card. The H100 case is a $28,000 GPU (midpoint of a roughly $25,000–$31,000 new range) plus an $8,000 single-GPU server (chassis, CPU, RAM, disk, PSU, NIC), $36,000 in total. That is not an 8-GPU HGX node.
Electricity is full load at $0.12/kWh. A 5090 system at about 0.70 kW is $0.084/hr. An H100 PCIe box at about 0.55 kW is $0.066/hr. Buy budget = purchase + hours × power. Rent budget = hours × hourly rate. Residual value is $0 in the main tables and treated separately below.
RTX 5090: 3-month, 1-year, 3-year totals
The buy column includes electricity. The rent column is GPU time only. Disk is called out in the next section.
| Plan | 3 months | 1 year | 3 years |
|---|---|---|---|
| Buy · ~10 hours/week | $4,810 | $4,840 | $4,921 |
| Rent $0.69 · ~10 hours/week | $83 | $331 | $994 |
| Buy · 6 hours each workday | $4,833 | $4,933 | $5,199 |
| Rent $0.69 · 6 hours each workday | $273 | $1,093 | $3,279 |
| Buy · 24/7 | $4,984 | $5,536 | $7,008 |
| Rent $0.69 · 24/7 | $1,511 | $6,044 | $18,133 |
| Rent $0.99 · 24/7 (stable instance) | $2,168 | $8,672 | $26,017 |
Three rows carry the result. Experiments over 3 years are about $994 to rent and $4,921 to buy. Daily development still favors rent at 3 years: $3,279 versus $5,199. Only 24/7 crosses the $0.69 rental near the end of year one: buy $5,536, rent $6,044. By year three that gap is $7,008 versus $18,133. A stable $0.99 instance pushes 24/7 toward buying sooner, but six hours a day still has not paid the card back at the end of year three.
One H100: when buying pays back
A $36,000 entry price rules out “buy one and see.” Occasional fine-tunes and a single shift of inference do not cover that capital inside three years.
| Plan | 3 months | 1 year | 3 years |
|---|---|---|---|
| Buy · ~10 hours/week | $36,008 | $36,032 | $36,095 |
| Rent $2.89 · ~10 hours/week | $347 | $1,387 | $4,162 |
| Buy · 6 hours each workday | $36,026 | $36,105 | $36,314 |
| Rent $2.89 · 6 hours each workday | $1,144 | $4,578 | $13,733 |
| Buy · 24/7 | $36,145 | $36,578 | $37,734 |
| Rent $2.89 · 24/7 | $6,329 | $25,316 | $75,949 |
After one year of 24/7, rent is about $25,316 and buy is about $36,578. Rent is still ahead by roughly $11,000. After three years the order flips: buy $37,734, rent $75,949. At six hours a day, three years of rent is $13,733, under half the purchase.
How the payback hours are calculated
Payback hours
Purchase cost ÷ (rent rate − electricity rate)
5090 vs $0.69: 4,800 ÷ (0.69 − 0.084) = 7,921 hours. At 24/7 that is about 330 days (~11 months). At 6 hours each workday it is about 60 months.
H100 vs $2.89: 36,000 ÷ (2.89 − 0.066) = 12,748 hours. At 24/7 that is about 531 days (~17.5 months). At 6 hours each workday it is about 8 years.
Three sensitivities. None of them change “rent for the first 3 months”:
- A new host: add about $2,000 of PSU and chassis to the 5090 and 24/7 payback vs $0.69 moves from ~11 months to ~15.5 months.
- Resale: the main tables assume you cannot sell the card. If a 5090 still fetches 40% of the purchase price after 3 years, 24/7 payback shrinks to about 198 days. The 2026 shortage inflated residuals. Do not use that as the default.
- Interruptible $0.45: 24/7 payback on a 5090 is about 546 days (~18 months). At the one-year mark, rent is still cheaper.
Costs left out of the main tables
- Cloud disk and egress. A few hundred gigabytes of persistent disk is often a few dollars to a few dozen dollars a month. It does not flip buy vs rent unless you store multiple terabytes.
- Calendar time is not GPU time. Payback uses hours the card is actually drawing load. Idle hardware has already spent the capital.
- Interruption. Cheap community instances can be preempted, and a long training run starts over. Price the stable $0.99 or $2.89 rate, not only the floor.
- Capability. A 32GB consumer card has no MIG and no data-center NVLink. One 5090 does not replace one H100 for 70B-class work.
- Tax and tariffed power. Import VAT and commercial electricity raise the buy side. The direction stays the same: power is the small term, hours are the large one.
Choose by horizon
| Use | 3 months | 1 year | 3 years |
|---|---|---|---|
| 5090, ~10 hours/week | Rent | Rent | Rent |
| 5090, 6 hours each workday | Rent | Rent | Rent |
| 5090, 24/7 vs $0.69 | Rent | Buy, slightly | Buy |
| 5090, 24/7 vs interruptible $0.45 | Rent | Rent | Buy |
| One H100, not 24/7 | Rent | Rent | Rent |
| One H100, 24/7 vs $2.89 | Rent | Rent | Buy |
Rent for three months and divide the GPU hours on the invoice by the payback hours above. A forecast that says “we will run 24/7” is usually the optimistic version of the workload.
Should a 3-month pilot buy hardware?
No. Even at 24/7, three months of a 5090 is about $1,511 to rent and $4,984 to buy including power. The H100 gap is wider: $6,329 versus $36,145. The pilot should rent, and it should record the hours.
Does 6 hours a day pay back in 3 years?
Not at $0.69 or $2.89. Three years of a 5090 is about $3,279 to rent and $5,199 to buy, and payback is about 60 months. The H100 case is about 8 years. Buying pays inside one to two years only when the card barely powers down.
Does electricity change the answer?
At a residential rate around $0.08–$0.12/kWh, power on a 0.7 kW 5090 system is roughly a tenth of a $0.69 rental hour. Moving the planning rate from $0.12 to $0.30/kWh stretches 24/7 payback vs $0.69 from 330 days to about 417 days. You still buy only when the card is nearly always on.
I already have a Mac. Do I need an NVIDIA card?
A coding agent that calls a hosted model does not require a local NVIDIA GPU. Budget a GPU when you must run private weights, fine-tune, or keep inference off the public network. Signing, simulators, and packaging on macOS are a separate bill: Mac for development, GPU somewhere else, and which server an agent loop actually needs. A cloud Mac covers the Apple toolchain. It does not replace the GPU hours above.
Keep the GPU bill separate from the Mac toolchain
Weights can sit on a rented GPU. Signing and xcodebuild still need macOS. Zilmac cloud Macs run the Apple toolchain. They are not sold as GPU hours.
Record GPU hours on the cloud invoice, and rent the Mac node for the build cadence. — View cloud Mac plans