Log in Become a host
RTX 3090 · 133 servers rented · at the latest count

The 3090 is a
renter favourite.

Earn up to ~$330/month per card with RTX 3090s on Clore.ai. That is the bare-metal contract figure, open to blocks from two GPUs on 30-day terms; a single 3090 listed per minute on the marketplace can earn up to ~$145/month. Demand is this card's strength: 133 RTX 3090 servers had a renter at the latest count, more than any other GeForce card on Clore.ai.

BTC USDT USDC CLORE
●You set the asking price ●Paid for the minutes it runs ●Withdraw in BTC, USDT, USDC or CLORE
RTX 3090 · TWO WAYS TO EARN monthly ceilings, per card
Bare-metal contract, per cardup to ~$330
Per-minute listing, per cardup to ~$145
Contract entry2 GPUs · 30-day terms
Servers rented, latest count133
Up to, per card / month $330USD
Single listing, up to $145USD
VRAM
24 GB
NVLink
Pairs
Payouts
Per minute
$330
Up to per card, per month, on a 3090 contract block from two GPUs
133
RTX 3090 servers rented at the latest count, the most of any GeForce card
$145
Up to per card, per month, when one 3090 is listed by the minute
24GB
GDDR6X per card, the capacity 4-bit fine-tuning renters shop for

Renters pay for capacity,
and a 3090 has 24 GB.

A large share of the work people rent GPUs for is limited by how much fits in memory rather than by how fast the tensor cores are. A 3090 holds as much as a 4090, which is why renters keep so many of them busy.

The 24 GB card renters reach for

A renter who needs 24 GB has two consumer options here, and for a job bound by capacity rather than tensor throughput the 3090 gets it done at a lower hourly rate than a 4090. That is why 3090 listings get picked, and why 133 of them were rented at the latest count, more than any other GeForce card.

Servers rented, latest count133
Memory per card24 GB GDDR6X

A pair opens a second market

Two boards on one server can hold a 40 GB quantised 70B between them, which a single 24 GB card cannot. If your cards are bridged with NVLink, the link runs at about 112.5 GB/s and makes tensor-parallel decoding noticeably less painful than going over PCIe. We do not expose an NVLink field to renters, so put it in the server description yourself.

NVLink 3, pairs only≈112.5 GB/s
70B at 4-bit needs≈40 GB

Built for long rentals

Rented cards run at load for hours at a time, and stability is what renters notice. A board that holds its clocks through a long job finishes on schedule, and renters come back to listings that do. Clore recommends running with overclocking off for exactly that reason.

Typical rentalHours at load
OverclockingOff, as recommended
★ OPTIONAL · LOCK CLORE, EARN CLORE

MFP Lock pays
up to the rental price, per tier.

MFP is a quality score the network assigns each server from its hardware, currently capped at 200. Locking CLORE against that score does two things. Tier 1 and Tier 2 pay an emission in CLORE, and a server can receive up to its own daily rental price from each of them. Every tier, Tier 3 included, raises what you keep on BTC, USDT and USDC rentals. The emission you actually receive depends on how much other capacity shares it, so this page shows the ceiling.

CAPACITY

What a full lock takes

MFP × 9,000CLORE across all three tiers

Tier 1 holds MFP × 2,000, Tier 2 × 4,000, Tier 3 × 3,000. Partial fills give proportional benefit.

CEILING

What each earning tier can pay

≤ 1 dayof your rental price, per tier, per day

Tiers 1 and 2 earn emission and are capped this way, while Tier 3 adds to what you keep on BTC, USDT and USDC rentals.

Optional, and nothing is withheld if you skip it · Your daily rental price is the per-tier ceiling · The full mechanics →

Spot bids or a fixed rate,
paid out minute by minute.

A 3090 listing can take spot bids, on-demand orders, or both. Every rented minute settles into your balance in BTC, USDT, USDC or CLORE, whichever order type brought the renter in.

Spot orders
Bids

Renters bid, the best bid runs

Spot suits checkpointed training runs. A higher bid or an on-demand order can take the server over, so busy periods pull the going rate up.

On-demand orders
Fixed

Your rate, held for the whole order

On-demand renters pay your published rate and cannot be preempted while their balance covers it. On a paired server they can also rent a single card, leaving the other free for the next order.

Withdrawals and payout currencies are covered in the host docs. Payouts and fees →

From one card
to a contract block.

A single 3090 earns on the per-minute marketplace, and two or more can also go on a bare-metal contract. Each figure below is a ceiling for one card over one month.

Per-minute marketplace
$145

Up to ~$145/month per card

One 3090 on the marketplace around the clock, priced at the top of what RTX 3090 renters currently pay on Clore.ai. The rate you publish decides where your listing sits.

Bare-metal contracts
$330

Up to ~$330/month per card

Priced at the top bare-metal contract rate for an RTX 3090 and counted around the clock, for blocks of 2 or more GPUs on 30-day terms. Clore agrees each contract with you directly instead of through a console switch, so talk to us about supplying bare-metal capacity.

Demand
133

Busy 3090 servers at the latest count

More than any other GeForce card on Clore.ai. Renters come for 24 GB: 4-bit adapter training, 70B models split across a pair, and image work that fills the whole card.

Ceilings are per card, per month, and MFP rewards can come on top of per-minute rentals. Host installation guide →

One install command
on a machine you already run.

There is no CLORE operating system and nothing to boot from a USB stick. The hosting agent installs onto an existing Ubuntu or HiveOS machine with working NVIDIA drivers.

01 / PREPARE

Ubuntu 22.04 or 24.04, or HiveOS

An x86_64 machine with NVIDIA GPUs; AMD is not supported. The installer does not install drivers, so nvidia-smi has to work before you begin. Ubuntu 18.04 and 20.04 are no longer supported.

02 / INSTALL

Run the installer as root

The dashboard hands you the command under My Servers, Add Server. Stable, beta and legacy channels exist; take stable unless you have a reason not to.

bash <(curl -s .../install.sh)
03 / REGISTER

The dashboard command pairs it for you

Copying the command from the dashboard installs and registers the machine to your account in one step. If you installed first, register afterwards by passing an init token to the agent's own script.

04 / PRICE IT

Set spot and on-demand per GPU

List on spot, on-demand or both, and change either at any time. A price that sits well with renters plus steady uptime is what fills a 3090 listing.

Six questions about listing
a well-occupied card.

Why are 3090s rented more often than 4090s on CLORE.AI?

Capacity at a lower hourly rate. Both boards hold 24 GB, and a great many workloads that need 24 GB are bound by how much fits rather than by how fast it computes, so a renter whose job is memory-bound gets the same room for less on a 3090. The latest count shows the result: 133 RTX 3090 servers had a renter on them, against 112 RTX 4090 servers.

Does my 3090's NVLink bridge make the listing more valuable, and how do renters see it?

We do not publish an NVLink field in the marketplace data, so a renter cannot filter on it and we will not claim they can. What they can see is that your server carries two cards, which is the prerequisite for splitting a 40 GB quantised 70B across a pair. If your two boards are bridged, say so in the server description; it is the only place that information reaches the person deciding. Treat the bridge as something that makes your two-card server work better for the renter who wants it, not as a field the platform surfaces on your behalf.

Can I supply RTX 3090s on a bare-metal contract?

Yes, from two cards. RTX 3090 bare-metal blocks start at 2 GPUs on 30-day terms, and contracts are what lift a 3090 to up to ~$330/month per card. Getting hardware into one is a conversation with the Clore team through support rather than a console setting. Until then, any 3090 can be listed by the minute for up to ~$145/month per card.

Do renters pay more for a 3090 pair than for two singles?

Not automatically, and it is worth understanding why. Listings are priced per whole server per day, so a two-card server is one listing at one price, and the per-GPU figures quoted across this site are that price divided by the card count and by 24. Nothing in the pricing model adds a premium for pairing. What a pair gives you is access to a different set of renters, the ones serving a model that does not fit on one board, and you set the asking price yourself. Whether that clears higher per GPU than two separate listings is your pricing decision, not a platform rule.

How do MFP rewards work on a 3090 server?

Lock CLORE against the server and it works for you in two ways. Tiers 1 and 2 pay an emission in CLORE while the server is rented on the marketplace, each capped at the server's own daily rental price, and every tier raises what you keep on BTC, USDT and USDC rentals. Full capacity is MFP times 9,000 CLORE across the three tiers, unlocking takes 10, 20 or 30 days depending on the tier, and the MFP calculator helps size a lock.

How should I price a 3090 listing?

Publish a spot rate, an on-demand rate, or both, and adjust whenever you like. Renters compare price alongside reliability, location and card count before they order, so a competitive rate on a server that stays online is what keeps a 3090 busy. A paired server can charge for the two-card demand it captures, since the asking price is yours to set.

What busy 3090 servers are running.

We do not inspect anyone's containers, so this is an argument from capacity rather than telemetry: these are the jobs whose memory footprint lands on 24 GB and nowhere smaller, which is the reason a renter reaches past a 10 or 12 GB board to yours.

Adapter training that needs headroom
a 4-bit 8B base, long sequences
≈4 GB of weights, 20 GB of room

The base is small once quantised. What the renter is paying your card for is everything that does not fit alongside it on a 10 GB board.

Host setup guide →
A 70B model across your pair
4-bit weights split over two boards
≈40 GB, so two cards or nothing

This is the demand a multi-card 3090 server captures and a single-card listing cannot. Mention the NVLink bridge in your description if you have one.

Payouts and fees →
Image models at the ceiling
Flux.1 dev at fp16, with offload
≈24 GB, the whole card

Jobs that use every gigabyte run for hours rather than minutes, which is the shape of rental that keeps a listing occupied rather than briefly busy.

MFP mechanics →

The large-VRAM cards, by how busy they are.

For a host, demand is the column to read first: servers with a renter at the latest count. The earning columns are ceilings per card, per month, worked out the same way for each GPU.

GPU
VRAM
Servers rented
Earn up to
Bare-metal route
Per-minute route
Contract entry
RTX 3090 / this page
24 GB
133
$330
$330
$145
2 GPUs, 30 days
RTX 4090
24 GB
112
$620
$620
$490
5 GPUs, 14 days
RTX 5090
32 GB
112
$700
$700
$690
8 GPUs, 14 days

the 3090 leads on servers rented · the 4090 and 5090 lead on the per-card ceiling

The docs behind every 3090 listing.

Installation, the MFP tier capacities and payouts are all covered in these pages. If anything here disagrees with them, they are right and this page is stale.

Host docs
Installing the hosting agent
Host docs
MFP lock mechanics
Host docs
Payouts and fees
Host docs
All host documentation

More cards you can put to work.

RTX 4090
24 GB Ada at 1,008 GB/s, contracts from 5 GPUs
Host page →
RTX 5090
32 GB Blackwell, the highest GeForce ceiling
Host page →
RTX 3080
10 GB Ampere, contracts from 2 GPUs
Host page →

The 24 GB card
renters keep busy.

Install the agent on a machine you already run, set an asking price, and add an MFP lock to the listing whenever you choose. Up to ~$330/month per card is waiting on the contract route.