You set the price and Clore.ai pays you for every minute the machine is rented, in BTC, CLORE, USDT or USDC. Listed on the per-minute marketplace, a single A100 40GB can earn up to ~$920/month per card. The board's structural advantage is MIG: it is one of the few GPUs NVIDIA lets you cut into seven isolated instances, which puts it in front of multi-tenant renters that very few cards can serve.
The board brings MIG, HBM2e and NVLink to a marketplace where renters filter on exactly those, and it pays by the minute in the currency you choose.
NVIDIA supports Multi-Instance GPU on the A100, A30, H100, H200, B200 and GB200, and on the RTX PRO Blackwell parts, and on nothing else. A renter who needs seven hardware-isolated instances on one board has a very short list of cards to choose from, and the A100 40GB is the most accessible entry on it. That is demand driven by a capability rather than by the hourly rate.
Earnings accrue for each rented minute and settle to your balance. Balances can be withdrawn in BTC, CLORE, USDT or USDC. Network rewards, where you qualify for them, are paid in CLORE.
An eight-board HGX baseboard is a single server with eight GPUs, not eight listings, and renters filter on GPU count when they want a coherent node. Price the machine, not the card, and let the marketplace divide it.
Every server carries an internal quality score, its Maximum Fair Price. Locking CLORE against that score does two things: it qualifies the machine for a share of block-reward emissions, and it lets you keep more of every order paid for in BTC, USDT or USDC. Skip it and you are still paid for every rented minute.
Shares the emission allocation with Tier 2 and starts raising what you keep on rentals paid in anything but CLORE.
The other half of the emission allocation, and a further step up in what you keep on BTC, USDT and USDC orders.
Renters pick between the order types you enable. Both bill per minute against prices you set, and both pay into the same balance.
Publish a spot price below your on-demand one and the board picks up checkpointed fine-tunes and batch jobs whenever it is free.
A reserved order stays with the renter until they stop it, which is what long fine-tuning runs and MIG-partitioned serving boxes want.
Read the host docs on getting paid. Payouts and fees →
~$920 is what one A100 40GB brings in when it is rented nonstop through the month at the high end of what renters currently pay for this GPU on Clore.ai. Your daily USD price is yours to set, so where your listing sits in that range is your call.
The figure for a whole A100 40GB on the marketplace, billed by the minute and paid in the currency you pick.
An HGX baseboard lists as one server, and every A100 on it earns toward the node. Partial GPU rental also lets an on-demand renter book a subset of the eight.
Platforms that need isolated instances have a short list of cards to choose from. Mention MIG in your listing and the A100 40GB is on it.
Ready to list? The install takes one generated command. Host onboarding guide →
The install is one command that the dashboard generates for you. Everything before it is making sure the box is the kind of box the agent expects.
MIG is configured on a machine that is already running, so the base image comes first. Ubuntu 22.04 LTS on x86_64 is the recommendation and 24.04 is supported; 18.04 and 20.04 are not. A working NVIDIA driver goes on before the agent. Budget 8 GB of RAM (16 recommended) and 32 GB of disk. HiveOS rigs are supported through their own flow.
Register, choose Add Server, and the dashboard builds a one-line installer carrying your init token. Run it as root. Docker is installed as part of setup.
After a reboot the machine reports in and appears online under My Servers. If you installed without a token, pair it afterwards with /opt/clore-hosting/clore.sh --init-token.
Publish a spot price, an on-demand price, or both, and adjust them whenever you like. Locking CLORE against the server's MFP score is optional and can wait.
Clore.ai lists and bills a server by its GPUs, so an A100 40GB goes to a renter as a whole card, and there are no per-slice listings. MIG is what the renter does with it: inside their own rental they can split the board into as many as seven isolated instances, which is exactly why multi-tenant platforms look for this card. On a server with several A100s, on-demand renters can also take some of the cards rather than all of them through partial GPU rental, always in whole GPUs.
As one server with eight GPUs, which is what a renter filtering for a coherent NVLinked node is looking for. The node is billed by those GPUs, so price it as eight cards. A renter who wants the whole fabric books all eight, an on-demand renter can take a subset of the boards through partial GPU rental, and whoever rents a board can partition it with MIG inside their own rental.
They filter on it because it is a hard boundary rather than a preference. A 70B model quantised to INT8 is about 70 GB and simply will not load on your board, and no price makes that work. What you are selling instead is everything that fits inside 40 GB with headroom, which is most fine-tuning and most serving of models up to around 30B quantised. The 80 GB part attracts different work at a different price; yours is not a cheaper version of it.
Every minute under an order is credited to your Clore.ai balance, and withdrawals go out as BTC, CLORE, USDT or USDC. That holds whether a renter reserves a whole node or, through partial GPU rental on on-demand orders, only some of its boards. Locking CLORE adds MFP rewards, paid in CLORE.
Per day, in USD, with separate spot and on-demand prices, and you can revise both at any time. The figure on this page, up to ~$920/month per card, is the upper end to aim for with a reliable, clearly described listing. Decide first whether you are pricing single boards or a full NVLinked node, because each draws a different renter.
Three kinds of work. Multi-tenant platforms partition the board into as many as seven isolated MIG instances. Fine-tuning runs use BF16 and TF32 at 312 dense TFLOPS and prefer on-demand orders so they are not interrupted. And renters who want a coherent NVLinked node book whole HGX baseboards, since nobody can assemble one from separate listings.
Useful to know when you write a listing description. These are the three shapes of demand this board attracts, and each one is about capability.
They need hardware isolation, not a scheduler. Say MIG in your listing and you are visible to a search almost nothing else answers.
Host docs →Jobs that run for days and care more about not being interrupted than about the hourly rate. On-demand listings capture this work; spot listings lose it.
Payouts and fees →Nobody can assemble an NVLinked node out of separate listings, so if you have one, that is the product. List it whole.
MFP mechanics →Monthly earning figures per card, beside the partitioning and memory specs that multi-tenant and training renters check. Route shows where each card's top figure comes from.
Start with the install guide, which takes an A100 from a working driver to a listed machine; after that come the MFP lock, how payouts work and the docs index.
The rent page carries the same 21 Sep 2026 snapshot from the other direction: what was listed, what it cost, and what a MIG slice is actually good for.
Install the agent on an Ubuntu 22.04 or HiveOS machine with a working driver, set your own price, and get paid for every rented minute in BTC, CLORE, USDT or USDC.