An H100 on Clore.ai can earn up to ~$1,750 a month per card, and that ceiling belongs to the contract route: a single card at the best H100 rate in Clore's bare-metal catalogue, counted for every hour of the month, supplied as nodes of 8 GPUs or more on terms from 30 days and agreed directly with the Clore team. The second route is yours to run: list servers on the per-minute marketplace, set spot and on-demand prices, and earn for each minute a renter holds the machine, paid out in Bitcoin, USDT, USDC or CLORE.
Serving and training teams both want this card, the contract route is sized for full nodes, and on the marketplace every price is yours.
Hopper brought FP8 and the Transformer Engine, and serving teams pick this card for them: a 70B model quantised to FP8 needs about 70 GB for its weights, so one 80 GB H100 holds it and still has room for cache. Training teams add a second reason, the NVLink 4 domain that ties eight SXM5 modules together. Both are shopping for the node you already run.
On the Clore.ai marketplace you publish a spot price, an on-demand price or both, change them whenever you like, and are paid for the minutes your server is rented. Earnings credit to your balance and withdraw in Bitcoin, USDT, USDC or CLORE.
Per-minute listings suit capacity you want to fill on your own terms. The bare-metal contract product is the other route, and it is where H100 demand concentrates: whole nodes of 8 GPUs on terms from 30 days, sold in five countries, and the source of the up to ~$1,750 a month per card, paid in USD stablecoin or BTC.
MFP Lock sits alongside your per-minute rental income. You lock CLORE against a server's Maximum Fair Price, and once a day has passed the server starts drawing MFP rewards, paid in CLORE for each minute a marketplace order runs on it. The same lock raises what you keep on marketplace rentals paid in BTC, USDT or USDC. Whether to lock, and how much, is entirely your choice.
Tier 1 holds 2,000 CLORE per point, Tier 2 holds 4,000 and Tier 3 holds 3,000. Across all three tiers the maximum total lock is the server's MFP multiplied by 9,000 CLORE.
Both tiers earn for every minute the server is rented on the marketplace, starting 24 hours after the lock. Tier 3 adds to what you keep on BTC, USDT and USDC rentals.
Spot and on-demand renters both pay by the minute. You choose which of the two your server offers and set the price for each.
Spot renters run at the spot price you publish and give way when an on-demand order arrives, so a single H100 server can serve both kinds of tenant.
On-demand renters hold the server for as long as they need it at your on-demand price, which suits serving teams and long training runs.
Both order types credit your balance minute by minute. The host docs cover payouts and fees →
The contract route sets the ceiling at up to ~$1,750 a month per card. On the marketplace, these three are yours to move.
Spot, on-demand or both, changed whenever you want. The rate on your listing is the rate you chose, and you can revise it the moment demand moves.
Billing runs per minute of an active rental, and partial rental lets on-demand renters take some of the cards in a multi-GPU server while the rest stay available.
The marketplace lets renters sort and filter by reliability, and its training and inference presets set a minimum. A node that stays online climbs those lists.
Running full 8-GPU nodes? Talk to us about supplying bare-metal capacity →
There is no Clore operating system, no bootable image and no PXE step. The hosting agent installs onto an existing Ubuntu server or HiveOS rig, and the driver has to be working before you start.
An x86_64 machine on Ubuntu 22.04 or newer, with 22.04 LTS recommended for driver compatibility. The documentation is explicit that nvidia-smi has to return before you install. Minimum 8 GB of RAM and 32 GB of disk; NVIDIA GPUs only.
One command, pulled from Clore's public installer repository. Appending your init token pairs the machine to your account, and the server shows as online after a reboot.
The installer detects HiveOS on its own, so the command does not change. Only the OS preparation differs, and the host docs cover the driver and disk steps specific to HiveOS.
Configure the server, publish spot or on-demand prices, and adjust them as demand moves. Earnings accrue per rented minute in BTC, USDT, USDC or CLORE.
Three kinds of job, mostly. Serving teams load 70B-class models at FP8, which fit on one 80 GB card with room for cache. Training teams run FSDP or DeepSpeed across all eight cards and lean on NVLink 4 at 900 GB/s inside the node. Smaller teams split a card with MIG inside their own rental. Each of them reads the listing for the same details, so name the module, the memory and the fabric precisely and the right tenant finds you.
Both work. The marketplace lists servers, and a server with one H100 in it is a server you can price and rent out per minute. Renters come to this card for the 80 GB memory budget and, for training, for the NVLink domain that only a full node provides. That is why the contract route treats the node as the unit, 8 GPUs on terms from 30 days, and why a multi-GPU server listed per minute also accepts partial rental: on-demand renters can take some of its cards at a proportional price.
Renters running distributed training care a great deal, and renters running single-card inference do not care at all. A whole-server rental hands over every card in the chassis, so the fabric the renter gets is whatever the box physically has; there is nothing to switch on. What you can do is be exact in the listing about whether these are SXM5 modules on an NVSwitch baseboard or PCIe cards with optional bridges, because those are different products at different prices, and a tenant who finds out afterwards is a tenant who does not come back.
They suit different buyers. SXM5 gives a renter 16,896 CUDA cores, 3,350 GB/s of HBM3 and a 900 GB/s NVLink domain in an eight-GPU HGX chassis, which is the node shape the contract route is built around and where the up to ~$1,750 a month per card applies. PCIe gives 14,592 cores and roughly 2,000 GB/s in a standard dual-socket server, and it goes on the per-minute marketplace at whatever price you set.
MIG is genuinely available on this part. NVIDIA's supported-GPUs table lists H100-SXM5 and H100-PCIE with up to seven instances at both 80 GB and 94 GB. Whether you can sell those slices as separate listings is a different question, and Clore's host documentation describes listing servers rather than MIG instances. Treat partitioning as something a renter does inside a rental they already hold, not as a way to multiply your listings, and check the host docs before you plan revenue around it.
One command, run on a machine that already has a working NVIDIA driver, which the documentation is explicit about: nvidia-smi has to return before you start. The installer comes from Clore's public installer repository, pulls Docker in as part of the install, and installs the hosting agent; appending the init token flag with your token pairs the machine to your account, and the server shows as online after a reboot. The same command detects HiveOS, so a rig and an Ubuntu server take the same path. Ubuntu 22.04 LTS is the recommended base, 24.04 is supported, and the documented floor is 8 GB of RAM and 32 GB of disk.
The three physical properties of an H100 node that decide whether a renter picks yours.
Enough to hold a 70B model at FP8 on one card with cache to spare, the most common reason a renter shortlists an H100.
Host onboarding →A whole-server rental hands over the interconnect as well as the silicon. For a distributed-training tenant that is most of what they are paying for.
Multi-GPU guide →Supported by the silicon and usable by a tenant inside their own rental, which widens the range of jobs a single card can take.
Training guide →Figures from NVIDIA's Hopper resources. Name the exact module in your listing, because these are the rows renters check first.
A listing that names its module precisely is easier to choose. Renters get the full FP8 argument for this card on the H100 rental page.
Installation, payouts and MFP Lock, documented by Clore and kept current there.
The renter page covers the FP8 argument, the SXM5 and PCIe split, and the region-by-region contract rates.
Install the agent on an Ubuntu server or a HiveOS rig, pair it with your init token, publish your own spot and on-demand prices, and get paid for the minutes the machine is rented.