An A100 80GB hosted on Clore.ai can earn up to ~$1,500 a month per card. That ceiling comes from the bare-metal contract route: one card earning around the clock at the highest A100 contract rate, in blocks from 8 GPUs on 30-day terms that are set up directly with the Clore team. The same machine can also go on the per-minute marketplace, where you name the price and are paid by the minute in Bitcoin, CLORE, USDT or USDC.
Renters come to this board for its memory, contract buyers take it in whole nodes, and every per-minute listing carries the price you chose.
Renters looking for this card are usually looking for a capacity threshold rather than a brand. A 70B model quantised to INT8 is about 70 GB of weights, which loads on an 80 GB board and does not load on a 40 GB one. Anyone with that requirement filters on memory first, and an A100 80GB listing is what the filter returns.
Per-minute listings carry the spot and on-demand prices you choose, and you can change them whenever demand shifts. Renters are billed by the minute, so your balance grows from the moment an order starts.
This board partitions the same seven ways as the 40 GB part, but each instance is roughly twice the size. That is the difference between selling a slice that fits a 1B model and one that fits an 8B at FP16, and it is worth saying in your listing.
MFP, the Maximum Fair Price Clore assigns to each server, sets how much CLORE you can lock against that machine. Locking is optional, and it pays a host in two ways: the locked CLORE earns MFP rewards while the server is rented through the marketplace, and it raises what you keep on per-minute rentals paid in BTC, USDT or USDC. Renters pay the price you set either way.
CLORE locked in the first two tiers shares in the CLORE emission for every minute the server is rented through the marketplace. Rewards are paid in CLORE.
Every tier you fill raises what you keep on rentals paid in those three currencies, and a full lock across all three gives the most. Tier 3 works on this side alone.
Marketplace earnings build up while each order runs, and they leave your balance in the coin you pick.
Renters are billed minute by minute on spot and on-demand orders alike, and your earnings land in your Clore balance while the order is still running.
Withdraw in whichever of the four suits your treasury, straight from the balance your rentals fill.
Contract income is paid separately, in USD stablecoin or BTC, your choice. Currency by currency, the host docs cover it under payouts and fees →
The contract route sets the top of the range. The per-minute route keeps every setting in your hands, and one machine can serve either.
Offer spot, on-demand or both, and move either rate whenever you like. Lead the listing with the 80 GB memory line, because that is what renters filter on.
The top of the range. Clore's A100 bare-metal catalogue is built on 80 GB SXM nodes sold in blocks from 8 GPUs, so a populated eight-board node already has the shape a contract takes.
The hosting agent goes on with a single installer line. From there you can list per minute straight away, while contract supply is agreed with the Clore team.
Eight or more boards in one node? Talk to us about supplying bare-metal capacity →
There is no queue to join and no approval to wait for. The agent installs with one command that the dashboard writes for you, on a machine you have already prepared.
Ubuntu 22.04 LTS on x86_64 is recommended and 24.04 is supported; 18.04 and 20.04 are not. The NVIDIA driver goes on first. Minimum 8 GB of RAM, 16 recommended, and 32 GB of disk. HiveOS rigs have their own path.
Sign in, choose Add Server, and copy the installer line it produces. It carries your init token. Run it as root; Docker is installed as part of setup.
The machine reports in after a reboot and shows as online. A machine installed without a token can be paired afterwards with /opt/clore-hosting/clore.sh --init-token.
Set spot, on-demand or both, then tune them as demand moves. Owners of eight or more boards can open the contract conversation at the same time.
Yes, on the per-minute marketplace. When every card in a server is the same model, an on-demand renter can take a subset of them, two boards of eight for example, at a price that scales with the number of cards, and several renters can share one machine at once. Spot orders always take the whole server, and a host can switch partial rental off for any individual rig.
It puts you in front of anyone whose model will not fit in 40 GB, which at current model sizes is a real and specific group. Seventy billion parameters at 8 bits is roughly 70 GB of weights: it loads on your board and does not load on the smaller one, at any price. That is a capability filter rather than a performance preference, which is the strongest kind of differentiation a listing can have.
The informed ones do, because decode throughput tracks memory bandwidth almost linearly and a 25 percent gap is visible in tokens per second on the same model. In practice, though, capacity is the reason people pick this board and bandwidth is the bonus. Lead your listing with the memory and mention the bandwidth second.
You can list a machine with two boards and 160 GB of total memory, and you should describe it exactly that way rather than as a 160 GB GPU, because NVLink is an interconnect and not a memory merger. A renter still has to shard the model across two devices. Saying two boards, 160 GB total, 600 GB/s peer bandwidth is both accurate and more attractive to the people who know what to do with it than a number that implies something the hardware does not do.
For most eight-board owners the contract route is the one to explore first. It is where the up to ~$1,500 a month per card comes from, and a populated node is exactly the block size it starts at, on terms from 30 days. Contracts are arranged with Clore rather than switched on in the console, so the first step is a conversation through the support page. A per-minute listing keeps you free to set and change your own prices, and the same agent install serves both.
Marketplace rentals are billed per minute and credited to your Clore balance as they run, and the balance withdraws in Bitcoin, CLORE, USDT or USDC. Locking CLORE in MFP is an optional extra on that per-minute route: it earns MFP rewards on rented minutes and means you keep more of every rental paid in BTC, USDT or USDC. Contract income is paid separately, in USD stablecoin or BTC, your choice.
What one board can earn, what renters load onto it, and the shape contract buyers ask for.
The most a single A100 80GB earns on Clore.ai, reached by supplying it to the bare-metal product inside an 8-GPU block.
Host docs →It loads on an 80 GB board and on no 40 GB one, which is why renters with a 70B to serve filter on memory before anything else.
The 40 GB page →Blocks start at eight boards in the USA, Japan and Slovenia, and the catalogue's A100 nodes are 80 GB SXM machines.
Bare metal →The top monthly figure per card for each GPU on Clore.ai, with the route that produces it. Specifications from NVIDIA.
Installation, MFP rewards and payouts, straight from Clore's documentation.
The rent page walks through what 80 GB actually buys a renter, why a 70B at INT8 is the dividing line, and how an A100 bare-metal block is configured.
Put the agent on any Ubuntu 22.04 or HiveOS machine whose driver already works and list it at your own price, or bring a node of eight boards or more to the contract route.