How much a GPU server and GPU rental make in 2026: our $2,165 month, card by card

Over 31 days, from 15 September to 15 October 2026, our six machines with ten GPUs brought in $2,165. The share of paid hours is about 95%. We hold to one definition for the whole article: each machine's income divided by its own card-hours and its own asking rate, then summed across the fleet: at full load the six machines would have made $2,285 over the window. Not "average machine utilization as reported by the marketplace", not uptime. There is something to check it against: the income of every machine is open on the "Fleet" page of our site, vast.sale — our live fleet page. That page's window is rolling and includes a client's machine, so its numbers will differ; below are our six machines only, for 15 September — 15 October. The previous breakdown (29 August — 28 September) showed $1,820 and 82% — the rows show what changed.

A word about the title. We stopped mining long ago — the cards are rented out for other people's workloads on the Vast.ai marketplace. But the hardware is the same, and the question we keep getting is the same: how much does a GPU server make today. Here are the numbers, what the sum is made of — and what is not in it.

In order: the facts per machine, two card comparisons (5090 against 4090, 3080 Ti against 5080), the income formula, catalogue rates, disk and traffic, downtime and failures with dates, costs in roubles, the payback method, and who this suits and who it does not.

31 days, 15 September — 15 October 2026: what each machine earned

The last column is not our asking price but one card's income over the window divided by all 744 hours, idle time included. The real rate is higher, because part of the time the machine sits empty. The split for this window: card rental $1,919.47, storage $238.31, traffic $6.82 — disk is 11.0% of revenue, traffic 0.32%. The share of paid hours below is computed on the full sum, storage included, so it runs on the high side: on card rental alone it would be 84%, not 95%.

The all-time total on the marketplace keeps growing — for the live figure see "Fleet": our live fleet page. In the article we keep only the measurement for the window.

What the table shows at a glance

The average of $216 per card explains nothing here. The spread runs from $94 to $324 per card — a factor of 3.4. The four-4090 machine made $1,148: 53% of the fleet's income from 40% of the cards. The three smallest machines by total — the 3080, the 3080 Ti and the single 5080 — made $359 between them, 3.2 times less.

RTX 5090 vs RTX 4090: why the gap is not "several times"

$324 against $287 per card — 1.13×. Not twice, not three times — and the gap has narrowed: a month earlier it was 1.39. Both machines went through the window almost without empty hours: 98% of paid time on the 5090 and 94% on the four-4090 machine.

A client pays for an hour of rental, not for raw performance. From there, everyone plugs in their own purchase price: if the new card cost noticeably more than the previous generation's flagship, it does not win that difference back in return per dollar invested. What it does give you is more money per slot in the chassis — and slots and kilowatts are finite.

A separate note on the 4090 machine — it is what explains the narrower gap. Until 20 September it ran without virtual-machine mode: after a card died on the way out of a virtual machine, we blamed "the incurable RTX 4090 reset bug", and we were wrong — that bug belongs to Blackwell cards. The real cause is the card dropping into deep sleep when detached from the guest, and device rules fix it. On 20 September the hardening was applied, a ladder up to four cards in virtual machines ran without a single reset, and VM went back on: in the first twenty-four hours the machine took 19 rentals, nine of them in VM mode. In this window only the first five of 31 days ran without VM — and income per card grew from $228 to $287. The step-by-step procedure: VM mode on a multi-GPU host

RTX 3080 Ti vs RTX 5080: the newer card wins, but not by much

The 3080 Ti made $101.00. The single 5080 made $164.20. The newer card earned 63% more. A month earlier the gap was only 17.5% — a good example of how much a single window depends on utilization.

The breakdown sits entirely in two factors. The 5080's rate is 1.45 times higher: $0.210 against $0.145 per hour. And its paid time is no longer lower: 105% against 94% on the 3080 Ti — over a hundred percent appears because the sum includes storage income. Check it by multiplication: 0.210 × 1.05 = $0.221 per hour per card, 0.145 × 0.94 = $0.136 — exactly the last column of the table.

In the previous window the two factors pulled in opposite directions: the 3080 Ti had more paid time, and utilization ate more than half of the 5080's rate advantage. In this window both factors favour the 5080, and the income gap nearly matched the rate gap. Income is the rate multiplied by paid hours, and that second factor is governed by price, bandwidth and reliability, not by the card model: the same two cards gave a gap of 17.5% and of 63% in two neighbouring months.

Within one generation the pair of 5080s lost to the single card in the previous window: $122 against $136 per card, 78% of paid time against 87%. In this window they drew level: $166.41 per card for the pair and $164.20 for the single one. Two things helped. The pair ran with virtual-machine mode for almost the whole window (the VM flag wiped on 11 September was restored on the 21st), and we raised its rate inside the window from $0.21 to $0.25 — it rents at $0.25 now against a $0.23 median. A two-card machine is rented whole, and such orders come less often — but with VM mode working and an honest price that is not a verdict. We still deliberately take no month-long contracts on it: we intend to rebuild it.

A caveat for the whole table: a row's number is not the card's ceiling, but the result of our rate and our plans for that machine over one month. The takeaway from the generation comparison is not "buy 5080s": the income gap against the same 3080 Ti was 17.5% and 63% in two neighbouring windows. Run the numbers on your own purchase and over several months, not one.

Why we left mining for renting

We got into Ethereum mining in 2017. After Ethereum moved to PoS in September 2022 came the hopping from coin to coin: you find the one that pays more, reconfigure, then find another. That took the autumn of 2022 and all of 2023. We were about to sell the cards, but prices for used GPUs had just collapsed.

That left a third option — rent the card out instead of computing something on it. In mining you sell the result of computation, whose price you do not control. In renting you sell the card's time. Same hardware, different business.

Our first attempt was Clore, and the main lesson from it is about the settlement currency. Earnings were paid only in the platform's token or in bitcoin, with a high fee: we were escaping volatility and ran into it through another door. If the settlement currency jumps around, you have no stable income, whatever the hourly rate. That was how it worked when we joined — check the payout terms yourself.

Vast.ai: our review after two years

The next marketplace we took on was Vast.ai: in 2023 the project looked more serious and clearly had more clients — a judgement from the time, we have no numbers for it. Almost a year went into studying it: how the price is formed, and what decides whether a client comes to you or walks past. Our first machine went live in June 2024.

If you were looking for Vast.ai reviews from the host's side rather than the renter's — this is ours. There were difficulties; there is a whole section on them below. But after more than two years: for us this was the right decision. We have seen coins that paid more in the moment — stability matters more. The fleet's numbers since June 2024 are open at our live fleet page.

What income is made of: the formula and the money mechanics

First, who pays for the card. Demand is carried by neural-network work: training and fine-tuning models, inference, image and video generation, less often password recovery and rendering. What the renter cares about is not the card class but the amount of video memory: either the model fits or it does not, and a 24 GB card sees orders a 10 GB card will never see. We have not measured the exact split by task type — this is an observation from two years of our own rentals.

There is one formula with four factors: monthly income = number of cards × $/hr rate × 730 hours × the share of paid time. 730 is an average month; the window we are breaking down had 744 hours. Everything else is derived from those four numbers.

The host receives 100% of their asking price: the marketplace markup is added on top for the client, not deducted from your money. There is no "minus 20% commission" to plan for.

# rough estimate of a host's monthly income
utilization = paid_hours / total_hours
revenue = utilization * rate_per_hour * gpu_count * 730   # 730 h = average month
# the host keeps the full asking rate: the markup is charged to the client, not to the host

Three of the four factors you set yourself: the number of cards, the hours in a month and your own rate. All the risk sits in the fourth one, the share of paid hours. That share in turn depends on four things:

• The rate — fully controllable, within the hour. The only lever of its kind.
• Bandwidth — controlled with money and configuration: speed, stability, a tunnel around your ISP's restrictions.
• The reliability score — only indirectly: it drops from downtime and IP changes and recovers slowly.
• The card class and its video memory — not controllable at all once bought. Which is why the first three are what to discuss.

The share of paid hours is that fourth number, and it is the decisive one. Our 95% is a very good month and should not be treated as normal for a first machine: a month earlier it was 82%. And it is an average in two senses at once: an average across the month rather than 95% every day, and an average across the fleet rather than per machine.

Catalogue rates: what to check against

Rates move, so look at the median for your model rather than the best offer in the listings. Below is a snapshot of the Vast.ai catalogue from 25 September 2026: single cards available to rent. We take this snapshot for the "Fleet" page on vast.sale, where our own rates sit next to the market. On the left the client price, on the right what the host receives.

Checking against the first table shows how a rate turns into money. The four-4090 machine made $1,148 over 744 hours — $287.04 per card, $0.3858 per hour with idle time included. Against our rate of $0.41 that is 94% of paid time. Every machine in the fleet checks out the same way:

The 105% on the single 5080 is not a record: the sum includes storage income. The same thing makes every other percentage in the column an upper estimate. The pair of 5080s had its rate raised inside the window from $0.21 to $0.25; we divide by the new one, so its 89% is a lower estimate. Multiply by your own share of paid hours, not by 100%.

Prices float with demand, and spot is cheaper than on-demand. Re-check any table of rates, including this one, in the catalogue before you buy: a median moves in a month. And an honest gap: we did not record medians for the 3080 and the 3080 Ti in this snapshot, so their rows have nothing to be checked against.

Disk and traffic: two components not everyone counts

Beyond GPU hours a host sells two more things. Storage — $/GB per month; it accrues even while the instance is stopped, and it is the only line that trickles in without a client. Over the window storage brought in $153.52 — 8.4% of the fleet's revenue; on the 3080 Ti it is what pushed paid time past a hundred percent. The second is traffic, $/TB in both directions: $7.63 over the window, 0.42% of revenue.

There is one caveat about traffic that we verified on ourselves. It is billed on Docker rentals only: the marketplace does not bill traffic for virtual-machine rentals at all. With $3 per terabyte set, a VM client pushed 8.32 TB in 37 hours, while the report for that machine reads Upload $0.00 / Download $0.00. We have written to support and have had no answer yet. For now, assume traffic on VM rentals brings you nothing, and keep that line out of the income when you model a multi-GPU VM machine.

Downtime and failures: the section usually left out

Income calculated without subtracting downtime is not an estimate, it is advertising. The window holds 744 hours, that is 7,440 card-hours across ten cards: at 95% paid, roughly 390 card-hours stay empty — about 39 hours per card — and that is an exceptionally good month: in the previous window 1,340 stayed empty. They are spread unevenly: from 98% of paid time on the 5090 machine to 89–94% on the others.

Downtime comes in two kinds. The first: no client at your price — treated with price and bandwidth, and you lose exactly that time. The second: the machine cannot take a client at all — a failure, a misconfiguration, disappearing from the listings. The second is more expensive: it hits the reliability score, and reliability decides how often you are seen in search.

Below is the log of failures we have caught across seven machines — six of our own and one client's. Where we kept the date, it is there; where we did not, we say so. This is not a list of risks borrowed from somebody else's article, it is what happened to us.

• 6 August: a dead memory module on the 4090 machine. The marketplace raised a "GPU PCIe issue" flag and sent the search towards the GPUs, memtest stayed silent; the real cause was in the BMC hardware event log — a module failure, the system took it out of service, and the machine carried on with 224 GB instead of 256. Several days of searching.
• The marketplace controller getting stuck: 25 August — four machines. Fixed by resetting the connection state (/var/lib/vastai_kaalia/controller_connection), but you have to spot it yourself: heartbeat = 0 with the hardware alive.
• The VM support flag is wiped by the marketplace's own hourly job when every card goes into a virtual machine. The two-5080 machine stood without VM mode for ten days — from 19:22 on 11 September to 17:04 on 21 September. No notification arrives.
• A tunnel loop at boot: xray came up before the machine got a DHCP address, so packets went into its own interface. 28,000 sockets piled up and filled the home router's connection table — and on 20 September four machines falsely failed over to the backup VPS.
• ISPs throttle download speed by the hosting provider's autonomous system — we tested them one at a time; the numbers are in the section on costs. No exact date, and we ran into it more than once.
• A stale CDI specification after a driver change: the machine vanishes from the listings with the status Not Listed. Everything looks fine from the outside, except nobody can see it. No dates kept — we ran into it twice.
• Changing the IP address penalizes the reliability score, and with it the flow of clients. We did not record dates: the address changed on every move to the backup VPS, including the false failover on 20 September.
• The marketplace agent hanging on a call to the marketplace's own web API: a blocking HTTP request with no timeout. The process is alive, the heartbeat does not go out, the machine is offline — this was the main cause of our September 2026 failures: the call went around the tunnel and was filtered; after sending it back through the tunnel — 37 calls out of 37 without a hang.
• Traffic from a VM rental that was never paid for: the numbers and our exchange with support are in the section on disk and traffic.

The money side is simple arithmetic. On the four-4090 machine, at our $0.386 per hour per card, a day offline is about $37, roughly 3,200 ₽. The ten days without VM mode on the two-5080 machine did not cost the full rate — they cut off a slice of demand: the machine never saw clients who need a virtual machine. Six of them fell into this window, all ten into the previous one; the difference shows in its row: $122 per card then, $166 now.

None of these problems solved itself. Each one is an evening at a console, and that evening is in no profitability table, including ours.

Paid hours and consumed electricity are not the same thing

Those 95% of paid hours are the share of time the machine is rented and paid for, not the share of time the GPU is computing. A client can book a server a month ahead, pay up front and never launch anything: for you that is full income, for the meter it is close to idle. We measured it — fifteen samples at 20-second intervals: an RTX 5090 with a 600 W limit averaged 43 W with 30 of 32 GB of memory occupied and its fans stopped, while an RTX 5080 in the same run sat at exactly 360 W under full load. Between an idle card and a working one the gap is roughly tenfold. Three of the ten cards were passed through to virtual machines — their draw is not fully visible from inside the host.

For a card owner the conclusion is the opposite of the usual one: the frightening electricity bill arrives not when the card sits idle but when a client runs heavy training and keeps it pinned for weeks. And that is good news — high-consumption months are also high-income months. The cost grows together with the revenue, not instead of it.

Costs: what goes out every month

$2,165 is revenue, not profit. We can name the fixed part of our costs precisely, from our own bills:

The tunnel line is not a whim. This is the same throttling by the hosting provider's autonomous system from the log above: of the three providers we tried a tunnel through, two turned out to be throttled and one was clean (later that one faded too — the list changes). Without a decent connection the card sits idle at any price, so the 3,750 ₽ a month for two VPS (1 core is about 750 ₽, 2 cores about 3,000 ₽, depending on the provider) is a condition of operating, not an overhead we could trim.

Then come the things that do not fit into a monthly column: spare parts, wear on the cards, fans, disks, replacing a dead memory module — and our own time spent on the list from the previous section. We pay for the home internet connection regardless of the servers, so it is not in the table. We deliberately do not smear depreciation across the months as one figure: it depends on the purchase price and on how long you plan to keep the hardware.

So we do not name a "net" figure: we have not yet tallied costs as carefully as income, and we do not want to pass revenue off as profit. Only one thing can be said honestly: fixed costs take about five percent of revenue, and the rest is eaten by depreciation and time.

What this is in roubles

Payouts arrive in dollars while electricity, VPS and spare parts are paid in roubles, so the conversion is a mandatory part of the arithmetic. At an exchange rate of about 87 ₽ to the dollar, the 31-day window looks like this:

• The whole fleet, ten cards: $2,165 ≈ 188,000 ₽ of revenue over 31 days.
• The average card: $216 ≈ 18,800 ₽ — except there is no average card, see the spread in the table.
• One 4090 in the four-card machine: $287 ≈ 25,000 ₽.
• The single 3080, the weakest row: $94 ≈ 8,200 ₽.
• Fixed costs: ≈ 9,750 ₽ a month.

We take the exchange rate as a reference at the time of measurement, not a forecast: it is not in our hands, and over a payback horizon measured in years it is a separate risk. If a calculation falls apart when the exchange rate moves ten percent either way, it is not a calculation but a bet.

Payback: it is a method, not a number of months

We deliberately do not name a payback period as a single number — it is made of things we do not know about you: what you paid for the hardware, what a kilowatt-hour costs you, what your bandwidth costs.

The method: take a card's monthly income, subtract its electricity and its share of the bandwidth costs, divide the purchase price by what is left — that gives the period in months. Then repeat it with utilization ten percent lower: that is your realistic horizon.

We do not name our own purchase price, and this is not secrecy. The fleet was assembled over several years, machine by machine, and hardware has got noticeably more expensive since. Any number we quote would be a lie either about today's market or about our costs. Run the numbers off your own receipts.

$1,148 a month from four 4090s is 94% of paid time at our rate of $0.41 with VM mode working — not the class of the card: a month earlier the same cards made $913. Payback is decided by bandwidth, reliability and price — that is, by the work after the purchase. The hardware sets the ceiling; the work decides how close you get to it.

And the main thing: our $287 a month from one 4090 is the 31 days from 15 September to 15 October 2026 at our share of paid hours and our rates — and a very good month at that. It is a reference for the order of magnitude, not a promise to you.

What we do every month to keep utilization up

• We check our rate against the median for the model: once a month we open the catalogue and adjust the price. The rate is the only factor that changes within the hour.
• We react to downtime the same day: a silent machine loses not only hours but its place in the listings, and reliability then takes weeks to climb back.
• We do not touch the IP without need: moving to the backup VPS is a last resort, and every move changes the address.
• We decide per machine whether to go spot: it is cheaper than on-demand, but it fills empty hours.
• Sometimes we deliberately hold the price above the market: we take no month-long contracts on the two-5080 machine because we intend to rebuild it.
• Once a month we check that the machine is listed and with the right flags. That is how you catch Not Listed after a driver change and a wiped VM flag.

This is the work after the purchase that makes the difference between half the hours paid for and our 82–95%. If there is nobody to do it, we do it for hosts as a service: our services

If you want to rent out your own card: what to check first

• Bandwidth. It affects utilization no less than price, and what you check is not the tariff but the specific hosting provider: the link to the marketplace is throttled by its autonomous system.
• Willingness to administer: drivers, containers, networking. The log above shows what kind of tasks come up and how often.
• A stable IP. Changing the address penalizes the reliability score, and that affects the flow of clients.
• One machine to start. Watch it for a month: a week shows weather, not climate.
• Calculate after electricity and bandwidth. Counting before subtracting those two lines is pointless.
• Headroom on utilization. If the plan only works at 70% paid hours, you do not have a plan.
• The material to hand: installing the OS — installing Ubuntu for Vast.ai, reaching the server without a public IP — server access without a public IP, virtual-machine mode on a multi-GPU machine — VM mode on a multi-GPU host, our fleet's live income — our live fleet page

Renting out a GPU is technically possible with almost any card. The economics are not: see the $94 and $101 rows.

Who this is not for

• Those who want to set it up and forget it: the log above is our evenings.
• Those not willing to deal with drivers, containers and networking: half the failures on that list are fixed by hand at a console.
• Those with one card in a gaming PC they need themselves: $94–101 a month from our weakest rows is unlikely to pay for the inconvenience.
• Those who need a fixed payment: utilization floats both from month to month and from machine to machine.
• Those with poor or expensive bandwidth: the card will sit idle at any price.
• Those who need the maximum right now: coins that pay more at this very moment appear regularly, and we deliberately do not go there.

If two or more items on this list apply to you — better not to start. We say this as people who count these numbers every month.

Who this is for

• Those with a stable connection — and the willingness to run a tunnel if their ISP throttles the link to the marketplace.
• Those ready to spend one evening a month at a console: update drivers, check the flags and the listing, work through one failure from the log above.
• Those with two or three cards in a dedicated machine rather than one card in a gaming PC, and a horizon measured in years, not weeks.

If that is you, we can do the sums with you: send your cards, your electricity tariff and your bandwidth via our services. There is also a turnkey option — from OS install to the first order, with a launch guarantee.

What next

First do the maths, then decide. The formula above has four factors, and it is those, not the card model, that decide the outcome: plug in your cards, your rate, your share of paid hours and your electricity tariff. Use 744 hours if you are checking against our table, or 730 for an average month. If you do not like the number, that is a result too, and a cheaper one than bought hardware.

There is something to check against: the figures for each of our machines are open on the "Fleet" page — our live fleet page — and refresh every hour; its window is rolling and includes a client's machine, so they will not match the 29 August — 28 September table. If you want a review of your own hardware, send a request at our services or https://t.me/flux_vast: name your cards, your electricity tariff and your bandwidth. We will tell you which of the pitfalls above apply to you. The first consultation is free. Nearby: installing Ubuntu 22.04 — installing Ubuntu for Vast.ai, reaching the server without a public IP — server access without a public IP, VM mode on a multi-GPU machine — VM mode on a multi-GPU host, password cracking on a 4090 — what password cracking costs