RTX 3090 inference economics: the account-rate correction

An RTX 3090 rents for 0.041 per hour on our account at the community cloud batch tier. The published list floor for the same card read 0.090 per hour on 26 October 2026. That gap is the rate correction this page documents, with every input named and dated. The used 3090 is still the price-performance king of its class on the speed axis: 24 GB at 936 GB/s for hundreds, not thousands. But at the account rate, renting an hour costs less than the electricity one burns under your desk. Buy the card for control, privacy, and fit. Do not buy it to save rental money.

All amounts are United States dollars written as bare numerals. Every rate carries its named party and its read date. Every table is static and every input is dated. You can check the arithmetic with a pencil. There is no calculator widget on this page.

Last verified: 2026-10-26

The rate correction: what a 3090 hour actually costs

Search for the 3090 cloud price and you meet a scatter, not a rate. Aggregators list the card from 0.12 to 0.16 per hour. One comparator shows a floor of 0.50. A pod with a fat memory config lists at 0.5. None of these numbers is wrong. Each prices a different thing: a different party, a different tier, a different market snapshot. That market moved 32 percent in a month. Any page that quotes one 3090 cloud price without naming the party and the date is quoting a mood, not a rate.

Here is the full ladder as read on 26 October 2026, lowest rates first.

The account rate is the number our own cost records carry for 3090-class capacity at the batch class of work. That class of work is long-running, restartable, and tolerant of reallocation. The rate sits under every published floor on the ladder. That is not a typo and not a trick. Three things hold it down.

First, published tiers are on-demand ceilings, not floors. The community cloud's own pricing page states that volume deployments, long-running jobs, and committed spend get priced under list. Second, the batch tier is interruptible by design. The lowest published tier is described on that same page as ideal for retryable batch work. It costs less because higher tiers can preempt it. Third, community capacity is idle consumer hardware in homes and small offices. In the operator's own words, it is priced near the cost of power, not the cost of a data center. The 3060 class rents on the same market from 0.04 per hour on our account, the basis of the 3060 budget guide.

What does the correction mean for you? If you rent at list, you are paying a convenience premium of roughly two times the account class. If you run retryable work that survives restarts, the interruptible tier is the honest match. And if you quote the market rate to justify buying a card, quote the floor, not the pod price, or the math that follows will flatter the purchase.

Why a 2020 card got more expensive in 2026

The 3090 is a GA102 Ampere part released in September 2020. It ships 24 GB of GDDR6X at 936 GB/s, on a 350 W board, from a 1,499 launch price. It is out of production. Every card sold today is used stock, and the used price went up, not down, through 2026.

The dated reads tell the story. In spring, guides put the card at 600 to 800. In March, a measured comparison paid 750 to 850. On 13 September, a buying guide called 820 to 900 the healthy range and anything over 1,000 overpaying. On 15 October, an eBay sold-listing tracker put the same card at 1,400 to 1,550. List-price guides lag sold prices in a rising market. The rental floor rose with it. The on-demand floor for a single 3090 sat at 0.11 on 26 October, up 32 percent in a month. The six month range runs 0.08 to 0.16. Vast.ai's own dedicated 3090 page showed no current offers at all on the day we read it.

The cause is documented. On 14 June 2026, Tom's Hardware reported on a research preprint about Pearl, a blockchain that markets its mining as useful AI computation. The study put the network at roughly 320,000 RTX 3090-class GPUs drawing an estimated 112 megawatts. The machines grind random matrix math that looks like AI work but trains nothing and serves no one. The same reporting tied a 38 percent jump in budget rental prices on vast.ai to that mining wave. Utilization climbed from 57 to 94 percent after the mining software went public in May.

There is a bitter footnote in the driver license. The GeForce driver EULA carries the clause "No Datacenter Deployment". It carves out one exception: blockchain processing in a datacenter is permitted, per the driver EULA record on the NVIDIA developer forums. The one use the license expressly allows in datacenters is the use that ate the supply. Everyone else sits on the restricted side of that same sentence. Know the gray zone exists before you bet a business on rented GeForce cards.

Used-market discipline still matters under a rising price. Mining-worn cards and dead memory cells are the known faults, and dead cells often show only under a sustained full-memory load, so stress the card before the return window closes. Two 3090s draw about 700 W together; budget the power supply before the second card. And a desktop session on the card eats about 4.4 GB, the difference between a model that loads and one that fails.

The memory math: what 24 GB costs per gigabyte today

A September 2026 hardware rescore carried the 3090 as a physics reference: 29 per GB, 936 GB/s, 2.7 GB per second per watt. It stayed off that menu's buy list on license grounds. That reference is already stale on the first number. The per-gigabyte price moves with the card price, and the card price moved. Work it in the open.

cost per GB  =  card price / 24 GB

card at 900:    900 / 24    =  37.5 per GB
card at 1,200:  1,200 / 24  =  50 per GB
card at 1,550:  1,550 / 24  =  64.6 per GB

For contrast, a new Intel Arc Pro B70 lists at 949 with 32 GB, which is 29.7 per GB, new, with a warranty. Street has since run above list. That narrows the gap per gigabyte and widens the 3090's lead per gigabyte per second. Even at list, on the capacity axis, the used 3090 no longer wins. A September reference of 29 per GB has become 50 at the mid read.

Flip to the speed axis and the 3090 keeps its crown. LLM token generation is bandwidth-bound. The card reads weights from memory for every token. Bandwidth is the speed. Memory is merely the room.

bandwidth per dollar  =  GB per second / price

3090 used at 1,200:   936 / 1,200  =  0.78 GB/s per dollar
B70 new, list 949:    608 / 949     =  0.64 GB/s per dollar

Per gigabyte, the new budget card wins. Per gigabyte per second, the used 3090 wins, and it holds a mature CUDA software stack and the last consumer bridge link for pairing two cards. One sensitivity line: at the 1,550 sold-tracker top, the 3090's bandwidth per dollar falls to 0.60 and the new card takes both axes. Know which axis your workload bills before you shop.

What 24 GB holds that 12 GB does not

Memory is a filter, not a preference. A model fits, or it streams. Streaming from system memory costs the same fixed penalty per layer, about 0.53 milliseconds per token, forever. The fit table, at the standard 4-bit class of copies.

The 24 GB ceiling is a measured fact, not a spec sheet line. A full workload run on this platform peaked at 23,807 of 24,576 MiB with 461 MiB to spare. Another stack allocates 23.22 of 24.00 GB with zero free, by design. When the card is full, it is full. That is why the disciplined buyers run headless and why the desktop tax above matters.

Speed receipts on the same card, all named and dated. On 8B models at Q4: 87 to 101 tokens per second, community-measured, October 2026. On a 30B sparse model: about 114. A tuned single 3090 reached 157.66 tokens per second on the A3B class in an owner-run benchmark updated 16 October 2026. Peak memory: 21,652 MiB. On a dense 34B at Q4: 12 to 18 tokens per second, enough for interactive chat. Against a Mac Mini with the same 24 GB of unified memory, the 3090 made 33 to 38 tokens per second where the Mac made 5 to 9. The memory holds the model. The bandwidth sets the speed. That 936 GB/s is why a 2020 card still outruns newer, dearer, slower memory.

The 3060 12 GB lane shares this market and has its own guide, linked above. Its 360 GB/s bandwidth is a little over a third of the 3090's. Its memory ceiling pushes every model in the table's lower half into streaming. For small models it remains the value pick. For the classes in the lower half, the 24 GB card is the entry ticket.

Break-even one: renting against owning

When does buying beat renting? Three named variables decide it, no others.

h*  =  A / (R - E)

h*: hours a month above which owning wins
A:  card price over amortization months, 1,200 over 24 months = 50 a month
R:  rental rate per hour
E:  electricity per running hour

The electricity basis is measured, not assumed. A full system draws 300 to 380 W at the wall under sustained inference. Call it 330 W. At the 2026 United States residential average of about 18 cents per kilowatt-hour, per the EIA Short-Term Energy Outlook, E is about 0.06 per running hour. The table assumes the box is switched off between runs.

Two rows deserve plain words. At the published floor, the card never pays for itself inside a month. Renting always-on for 730 hours at 0.090 costs 65.70. Owning the same month costs 50 of payback plus about 44 of power. Lower duty only makes owning worse. At the account rate the crossover goes negative: an hour of rented card costs less than an hour of power for the card on your desk. A purchase sold on rental savings cannot clear that bar at any duty. One sensitivity: stretch the payback window to 48 months and A halves to 25. The floor row still needs 833 hours a month, past the calendar. The account row never crosses.

The honest buy case is everything else the card sells. Your data stays in your building. Your latency is yours. Your model versions do not change under you. And the card resells. A card that still sells for hundreds in two years is partly a deposit, not a cost. Those reasons survive the math above. The savings argument does not.

Break-even two: the deposit math and the serving side

Run the same account rate from the other side. A card price divided by the account-rate month tells you how many always-on months the purchase could have funded.

funded months  =  card price / 30

card at 900:    900 / 30    =  30 months
card at 1,200:  1,200 / 30  =  40 months
card at 1,550:  1,550 / 30  =  52 months

At the October sold-tracker top, one used card funds over four years of always-on rented capacity at the account class. Read that as a price signal, not as advice to rent. The same scarcity is what carries the card's resale value.

If you serve models to others on one card, the same 30 a month is the heat cost your revenue must cover. The rates below come from the platform's model data. The rendered table lists what the catalog lists. The models index is the live catalog.

With the rate side anchored to the fence, the revenue bars at the 30 a month basis: 600 images a month on the depth lane at the fenced image rate, or 20 a day. 750 images on the large segmentation lane at the fenced segmentation rate, or 25 a day. 5,000 calls a month on the openvla-oft action lane at the fenced per-call rate, about 167 a day. 10,000 calls on the pi05 lane at the fenced per-call rate, about 333 a day. The molmo2-er vision-reasoning lane at the fenced per-million token rates, so the bar is about 6 million output tokens a month. These are worked bars from the fenced rates, not promises about demand. The packing math that turns one card into many served robots is its own guide, the robot-hour cost derivation.

Decision rules

Frequently asked questions

Is a used RTX 3090 still worth buying in 2026?

For what it uniquely provides, yes: 24 GB with 936 GB/s, at a used price no new card matches on the speed axis. Pay the disciplined price, not the panic price. Sold listings ran 1,400 to 1,550 on 15 October 2026 while guides still called 820 to 900 healthy. Check a sold tracker on the day you buy. Treat any listing near 2,000 as a new-card comparison, not a deal.

What is the lowest 3090 cloud rate?

The published floors on 26 October 2026: 0.090 per hour at the community cloud batch tier. The open marketplaces show 0.08 to 0.09 in interruptible spot bids. Our account pays 0.041 at the same class of capacity for long-running work. Anyone quoting one number without party and date is quoting a single point of a market that moved 32 percent last month.

How fast is a 3090 for LLM inference?

Named receipts: 87 to 101 tokens per second on 8B models at Q4, about 52 on a dense 14B, about 114 on a 30B sparse model. Tuned, the A3B class reaches 157.66. A dense 34B runs 12 to 18. Tuning moves the small numbers the most: quantized caches and flash attention roughly double default-install speeds. Bandwidth sets the ceiling. The 936 GB/s figure is the number to compare across cards.

Why not a 3060 12 GB instead?

The 3060 shares the market at half the memory and a little over a third of the bandwidth. That bandwidth is 360 GB/s. Models above about 16 GB stream from system memory on it, at a measured 38.9 tokens per second for the A3B class with experts in RAM. For 7B and 13B work it is the value pick and has its own guide. For everything in the lower half of the fit table, the 24 GB card is the entry ticket.

Does electricity decide the rent-or-buy call?

At home rates, no, and that is the surprise. A running hour costs about 0.06 in power at the measured 330 W average. Renting at the 0.090 published floor already costs more than that, and renting at the 0.041 account class costs less. Your power bill cannot beat the account rate, so the card never pays for itself on energy savings. The decision rides on control, privacy, and fit, priced honestly above.

The correction in one line: the 3090 market has two price worlds. The published list starts at 0.090 per hour. The account class runs 0.041 for work that tolerates interruption. Every honest number on this page names its world. Rates for every hosted lane sit on the pricing page. Sign up, create a key, and rent or buy with the math in hand. The budget lane and the sparse-model lane each have their own guide: the 3060 budget guide and the A3B lane.

Last verified: 2026-10-26