Renting your GPU to AI workloads in 2026 pays real money — about $30 a month for an RTX 5090 at realistic utilization and average electricity prices, roughly break-even for a used 3090, and a few dollars for anything smaller. The marketplace listings imply four times that. The gap is made of three things most people discover only after the card is installed: hosts earn about 25% less than the listed rate, the card sits unrented most hours, and it draws power all 730 hours of the month regardless. This guide runs the honest math on turning a graphics card into an AI landlord’s income — and on whether it’s worth buying hardware to do it.
Where GPU Rental Income Comes From
The main open marketplace is Vast.ai: you install Linux and NVIDIA drivers, open a few ports, list your card, and set your own price. It publishes a live public pricing feed (the numbers below come from the September 2026 snapshot), pays weekly with a $20 minimum via PayPal, Wise or Stripe, and takes no markup on host-set prices — though its own article notes hosts typically realize about 25% less than live listing prices once marketplace dynamics play out.
The alternatives are narrower. Salad runs on ordinary Windows consumer PCs and pays in app balance redeemable for PayPal or gift cards, at lower effective rates. RunPod’s Community Cloud and io.net connect home GPUs into AI clusters, with vetting and eligibility requirements that exclude many home setups. For a first host machine, Vast is where the observable market is; everything in this guide prices against it.
What the Market Pays Right Now
Vast.ai’s public pricing feed, September 2026, with the listing-to-reality haircut applied:
| Card | Median listed rate | Realistic host rate | Net/month at 30% util, 18.3¢ power |
|---|---|---|---|
| RTX 3060 12GB | $0.055/hr | ~$0.04/hr | −$2 (loss) |
| RTX 3090 24GB (used) | $0.147/hr | ~$0.11/hr | −$3 (break-even) |
| RTX 4090 24GB | $0.360/hr | ~$0.27/hr | +$30 |
| RTX 5090 32GB | $0.440/hr | ~$0.33/hr | +$35 |
The assumptions behind that last column: 30% of hours actually rented (a realistic open-marketplace figure), 25% below the listed rate realized, and an average power draw between idle and load (80 to 280 W by card) charged 24/7 at the EIA’s US residential average of 18.31¢ per kWh from July 2026.
Read the table for its shape, not its precision. High-end cards with 24 GB of memory (the 4090 and 5090) are what AI renters actually want, and they clear the power bill with room to spare. Lower-VRAM cards don’t: at the national average electricity price, a 3060 hosting for AI is a machine that costs more to plug in than it earns. And electricity is the swing variable everywhere — rerun the 5090 at a 10¢ rate and the net jumps to about $48 a month; rerun the 3090 in Hawaii at 48¢ and it’s a $60-a-month hobby.
The Costs Nobody Puts on the Listing
Power, around the clock. The card draws electricity whether rented or idle — that’s the column that decides everything. A two-3090 host rig pulling roughly 850 watts costs about $114 a month at the US average rate before earning a cent.
Storage and bandwidth. Vast bills storage while an instance exists — even when it’s stopped — and bandwidth per byte in both directions. Trivial on a quiet card, real on a busy one.
Wear. AI jobs load a card as hard as mining does. Fans and thermal pads age first; a year of hosting typically earns a card a $30-to-$60 service, and a used card’s resale price discounts hosted history.
Payout friction. The $20 minimum and weekly cadence mean small hosts wait for their money. Not a cost in dollars — a cost in cash flow.
Your time. Hosts are their own data center: reboots, driver updates, and a renter’s broken job are your tickets. The hourly rate does not include you.
Utilization Is the Whole Game
The single biggest gap between the listing page and your payout is that nobody rents your card most hours. At 100% utilization the listed rates look like a business; at the 30% a home host on an open marketplace realistically sees, they look like the table above. Interruptible capacity — renters who can be preempted — is what home hosts actually sell, and it’s both cheaper for renters and spottier for you.
Scale changes the picture, which is exactly why vendor marketing uses it. Vast’s own example — a four-5090 rig at 80% utilization earning $700 to $1,400 a month gross — is real math on $8,000-plus of cards, operating somewhere with cheap power and excellent uptime. It is not a one-GPU outcome, and utilization that high on a home connection is the exception, not the plan.
Three Hosts, Three Outcomes
The card you already own. An RTX 4060 sitting in a gaming PC: about $17 a month gross at realistic rates, $11 of power at the national average — call it $7 a month of found money. No capital at risk, fans age a little faster, and you learn exactly how the market works before spending anything.
The used 3090 buyer. A $1,200 used 3090 at average power breaks even monthly — and even at a cheap 8¢ rate nets about $11 a month, a nine-year payback. The verdict is blunt: a 3090 is a host-it-because-you-own-it card, not a buy-to-host card. Its 24 GB of memory is exactly what renters want, but not at prices that clear the power bill.
The 5090 owner. At a 10¢ rate a 5090 nets about $48 a month — real income, on a card that costs $2,000 to $3,600 depending on the memory shortage’s mood that week. As a payback story that’s four years; as “my gaming rig earns its electricity bill,” it’s genuinely reasonable. Buy-to-rent still doesn’t pencil. Monetize-what-you-own does.
Mining vs. Renting the Same Card
The alternative use for the same silicon is mining, and in late September 2026 the two are unusually close: WhatToMine showed an RTX 4090 grossing about $8 a day and a 5090 about $7.70 (coin-led, spike-inflated, and wildly volatile) versus the $1 to $2.40 a day the rental market implies for the same cards. Mining currently out-earns median rental for high-end cards when the coins cooperate; renting pays in dollars instead of coins, with no price risk and no pool cut. Neither changes the core rule: the card you already own is the only entry price that makes either one easy money.
The Mistakes That Eat the Income
Believing the listing rate. Hosts realize roughly 25% below it, per Vast’s own article. Budget from the realized rate, always.
Ignoring idle draw. The silent killer. A rig pulling 200 W around the clock costs $27 a month at the national average before earning a cent — which is most of a 3090’s gross.
Assuming 100% utilization. Plan at 30% and treat anything better as a bonus.
Hosting on a bad power circuit or cheap PSU. Same fire-risk math as any always-on machine, with renters’ jobs hammering the card on top.
Skipping the fine print. Storage billed while stopped, per-byte bandwidth, $20 payout minimums — all published, all discoverable before you buy hardware, all somehow discovered after.
Bottom Line
GPU rental income in 2026 is a monetize-what-you-own play: about $7 a month for a modest card, $30 to $50 for a 4090 or 5090 with cheap power, and roughly break-even for a used 3090 at the national average electricity rate. It is almost never a buy-to-rent play — the payback math doesn’t clear even on flagship cards. Run your own card and rate through the GPU rental income calculator, compare against the same card’s mining math in the GPU mining guide, and if you were about to spend four figures on a rig, read the Antminer numbers first — the industrial version of this trade has the same lesson at higher stakes.
FAQ
How much money can you make renting out your GPU?
With a high-end card: roughly $30 to $50 a month net at realistic utilization and average-to-cheap power. With a used 3090: about break-even at the 18.3¢ US average electricity rate. With anything smaller: a few dollars, sometimes negative. One card is pocket money — Vast’s own $700-plus monthly examples require four flagship GPUs at 80% utilization.
Is renting out a GPU profitable?
For a card you already own, usually yes — it’s found money against a power bill you pay anyway. For a card bought to rent out, rarely: even a 5090 netting $35 a month takes years to repay a $2,000-plus card, and a 3090 never quite does at average power. Profitability lives in cheap electricity (under about 10¢) and cards with 24 GB of memory or more.
Which platforms pay you to share your GPU?
Vast.ai is the main open marketplace — hosts set prices, payouts are weekly with a $20 minimum via PayPal, Wise or Stripe, and it publishes live per-GPU pricing. Salad runs Windows consumer PCs and pays in redeemable balance. RunPod’s Community Cloud and io.net connect home GPUs with stricter eligibility. Terms change often; read current host docs before buying hardware.
Why does my rig earn less than the rate times 24 hours?
Three haircuts: utilization (most hours nobody rents the card — 30% is realistic), the listing-to-earnings gap (about 25%, per Vast’s own article), and 24/7 electricity on all 730 hours of the month whether rented or not. Multiply the listed hourly rate by 730 and you’ve found the marketing number; apply the three haircuts and you’ve found your payout.