Earn up to ~$700/month per card by supplying RTX 5090s through Clore.ai. That top figure is the bare-metal contract route; list a single 5090 on the per-minute marketplace instead and it can bring in up to ~$690/month. You set the price, renters pay for every minute on your card, and payouts arrive in BTC, CLORE, USDT or USDC.
A 5090 carries more memory and bandwidth than any other GeForce card on Clore.ai, and the jobs that need them keep coming back to it. Three reasons a listing earns well.
Renters pay more per hour for a 5090 than for any 24 GB or 16 GB card, because 32 GB at 1,792 GB/s keeps Flux.1 dev, 720p video diffusion and 32B models at 4 bits resident on one GPU. That demand is why a single listed card can reach up to ~$690/month per card.
Earnings accrue for every minute a rental runs, in BTC, CLORE, USDT or USDC, and land in your Clore.ai balance as the order goes. Locking CLORE in MFP raises what you keep on the BTC, USDT and USDC rentals.
There is no fleet requirement. The same installer runs on a single desktop and on a rack, provided the box is x86_64 Ubuntu 22.04 or 24.04, or a HiveOS rig, with an NVIDIA card and a driver that already works.
MFP is a quality score Clore assigns each server, up to 200, from its GPUs, PCIe layout, CPU and network. One point of score permits 2,000 CLORE in Tier 1 and up to MFP times 9,000 across all three tiers. Locking there raises what you keep on BTC, USDT and USDC rentals, and hosts in Tiers 1 and 2 also earn MFP rewards in CLORE while the server is rented on the marketplace.
Unlock takes 10 days for Tier 1 and 20 for Tier 2. These are the two tiers that earn MFP rewards in CLORE on per-minute rentals.
Fill all three tiers and you keep the most on every BTC, USDT and USDC rental your 5090 serves on the marketplace.
Every listing can carry a spot price, an on-demand price, or both. Either way you publish the rate, and earnings accrue for each minute a renter is on the card.
Spot renters compete on price, and a higher bid can take the card over mid-rental. It suits restartable batch work, and competition among bidders works in your favour.
On-demand renters pay exactly the rate you set, and the order cannot be outbid while their balance covers it. Model serving and long video jobs usually arrive this way.
Payouts accrue per rented minute in BTC, CLORE, USDT or USDC. Payouts and fees →
List one 5090 by the minute, or bring a block of them to a bare-metal contract. Both figures below are per card, per month, and the higher one is the headline on this page.
A card rented around the clock at the top of what renters currently pay for an RTX 5090 on Clore.ai. You set your own rate, so the listing decides where in that range you land.
The top bare-metal contract rate for an RTX 5090, earned around the clock, on blocks from 8 GPUs and 14-day terms. Contracts are arranged with Clore rather than switched on in the console, so talk to us about supplying bare-metal capacity.
112 RTX 5090 servers on Clore.ai had a renter at the latest count, running everything from 32B models at 4 bits to Wan2.1 and HunyuanVideo at 720p.
Both ceilings are per card, per month, and MFP rewards can add to the per-minute route. Full host guide →
There is no Clore operating system to install and nothing to boot from. The agent goes onto an existing Ubuntu or HiveOS box, and the steps below follow Clore's own installation guide.
Ubuntu 22.04 or 24.04 LTS on x86_64, or HiveOS. NVIDIA only, with a driver that already answers nvidia-smi. At least 32 GB of disk and 8 GB of RAM, and overclocking off, which Clore strongly recommends.
One command from the stable channel, after sudo -i. It does not install or upgrade your GPU driver, so fix that first if it is missing.
Append --init-token <token> to that command, or register afterwards with /opt/clore-hosting/clore.sh --init-token <token>.
You publish a spot price, an on-demand price, or both, and can change either whenever you like. The marketplace shows what other 5090 hosts are asking.
Bare-metal RTX 5090 capacity is sold in blocks that start at 8 GPUs on 14-day terms, and that contract route is where the up to ~$700/month per card figure on this page comes from. Supplying it is arranged directly with Clore rather than switched on in the console, so the first step is a conversation with the team through support. Any 5090 you own can earn on the per-minute marketplace in the meantime, up to ~$690/month per card.
The responsibility is yours and it begins before the agent does. Clore’s installation guide requires a working driver that answers nvidia-smi and states that the installer will neither put one there nor upgrade it. From that point the compatibility question belongs to you and the renter jointly: they supply the container and the CUDA runtime inside it, and your host driver has to be recent enough to serve that runtime. That guide names no minimum version, though your server’s page in the dashboard flags any driver below the 535 branch as outdated. Beyond that, the useful diagnostic is simple: when images that run elsewhere fail on your machine, the driver is the first thing to look at.
Plenty, and it is broad. At the latest count 112 RTX 5090 servers on Clore.ai had a renter on them, and the work they carry spans 32B language models at 4 bits, Flux.1 dev with room for larger batches, and video diffusion at 720p. Renters compare price, reliability and location before they commit, so a well-priced listing that stays online is the one that fills.
For a specific set of jobs, yes. Thirty-two gigabytes is what keeps Flux.1 dev at full precision, and Wan2.1 or HunyuanVideo at 720p, resident on the card rather than paging across PCIe. It is also the line between a 32B model at 4 bits fitting with room for context and not fitting at all. What it does not reach is 70B, which needs roughly 40 GB at the same quantization, so renters chasing that size go to two cards or to datacenter parts no matter what you list.
To you the single card is simply less work: one slot, one listing, one set of drivers to keep current. To a renter it depends on the job. Two 4090s offer 48 GB but split across PCIe with no NVLink between them, which suits large-batch serving and penalises latency-sensitive decoding. One 5090 offers 32 GB behind 1,792 GB/s with nothing to split, so for anything that fits inside 32 GB the single card is usually the simpler and better-value rental for them, and that is what lets a 5090 listing ask more per hour.
Ubuntu 22.04 LTS or 24.04 LTS on x86_64, or a HiveOS rig. The installation guide states plainly that 18.04 and 20.04 are not supported and that the hardware must be NVIDIA, with AMD cards excluded. Beyond that you need at least 32 GB of disk and 8 GB of RAM, a working driver that nvidia-smi can query, and outbound HTTPS reachability to Docker Hub, the NVIDIA repositories, github.com, gitlab.com and api.clore.ai. Turning overclocking off and returning the cards to factory settings is strongly recommended rather than required.
Three jobs that run best with 32 GB on one card. They are the reason renters pay more per hour for a 5090 than for a 24 GB board.
The largest model class a renter can keep on one GPU here. A 24 GB card runs it too but with much less room for context.
Read the guide →Fits a 24 GB card exactly and a 32 GB one comfortably. That headroom is what lets a renter raise batch size and stay on the card.
Read the guide →These are long-running rentals, which is the kind a host wants: hours of paid time on a single order.
Read the guide →Per card, per month, worked out the same way for each GPU: the per-minute ceiling, the top bare-metal contract rate, and the higher of the two. The last column is servers with a renter at the latest count.
Installation requirements, the MFP lock mechanics, and how payouts work. Where this page and the documentation disagree, the documentation is right.
Rent one by the minute on Clore.ai, spot or on-demand, with root access inside your own container.
Up to ~$700/month per card on a bare-metal contract, up to ~$690/month per card listed by the minute. The install is one command on the Ubuntu or HiveOS box you already run.