The B200 carries the highest per-card ceiling on Clore.ai: up to ~$3,700 a month per card. That figure is the contract route, one card at the top of Clore's B200 bare-metal rate for every hour of a month, supplied as whole 8-GPU pods on terms from 30 days in the USA, Japan and Slovenia and agreed with the Clore team. Renters pay for Blackwell's FP4 path, 180 GB of HBM3e per card and NVLink 5 across all eight. The per-minute marketplace takes the same machine at prices you set.
Renters contract this card for three properties that only a whole pod delivers, and Clore sells it exactly that way.
Blackwell is the first NVIDIA generation with a four-bit tensor path. DeepSeek-V3's 671B experts shrink to about 336 GB at FP4, two B200 cards where FP8 needs four, and inference teams contract whole pods to serve models of that class. That capability is the core of what your B200 pod sells.
The bare-metal route contracts B200 in whole eight-GPU units, from a single pod up to 1,000 GPUs, and it is the route behind the up to ~$3,700 a month per card. Each contract is set up with the Clore team directly rather than through a console toggle.
180 GB of HBM3e a card at 8,000 GB/s, and an NVLink 5 domain moving 1.8 TB/s between GPUs. A whole-server rental hands over the interconnect with the silicon. NVIDIA's MIG table also lists the B200 with up to seven instances, which a tenant can use inside a rental they already hold.
MFP Lock connects a listed server to the network's CLORE emission. Put CLORE against the server's Maximum Fair Price and from the next day it collects MFP rewards on every minute it is rented through the marketplace. Filling the tiers also means you keep more of each marketplace rental paid in BTC, USDT or USDC. It is optional and works alongside whatever the server earns from per-minute renters.
As the MFP Lock documentation sets them out. The lock itself is sized from the server's MFP, and the tiers fill in order.
Before that the server earns rental income only. The ceiling across all three tiers combined is the server's MFP multiplied by 9,000 CLORE.
Contract income is paid in USD stablecoin or BTC, your choice; marketplace listings credit your Clore balance per minute and withdraw in BTC, USDT, USDC or CLORE.
Supplied to Clore's bare-metal product in units of eight cards, on terms that run from a month to 1,800 days and are agreed with the team before the pod goes live. The catalogue describes each pod with 1.5 TB of RAM, eight 3.84 TB NVMe drives and 3.2 Tbps Ethernet or RoCE, the shape buyers expect.
List any B200 server with spot and on-demand prices of your choosing. Each rented minute is credited to your balance while the order is live.
How balances withdraw, coin by coin: payouts and fees →
Contracts set the top of the range, and the marketplace gives the same hardware a second way to earn.
The top figure for one B200 in a contracted pod, and the highest per-card ceiling of any GPU on Clore.ai.
List a pod or any B200 server with your own spot and on-demand rates, and revise them as often as you like.
Commit a pod for a single month or for nearly five years, with the terms in between offered in all three regions.
B200 pods ready now? Talk to us about supplying bare-metal capacity →
Joining Clore.ai takes an installer and a token on any Ubuntu or HiveOS machine you already run.
Create an account and open Add Server in the dashboard. It writes the installer line for you, with your init token already in it.
The install documentation treats a working NVIDIA driver as a precondition rather than a step. x86_64, Ubuntu 22.04 or newer, 8 GB of RAM and 32 GB of disk as the documented floor.
Pulled from Clore's public installer repository, with your init token appended to pair the machine. Docker is pulled in as part of the install, and the same command works on a HiveOS rig.
Spot, on-demand or both, priced by you. Earnings credit per rented minute and withdraw in BTC, USDT, USDC or CLORE.
Mostly the largest models in circulation. DeepSeek-V3's 671B mixture of experts is about 336 GB at FP4 and fits on two cards; Llama 3.1 405B at FP8 is about 405 GB and fits on three; sharded training runs use all eight inside one NVLink 5 domain. Blackwell's four-bit path and 8,000 GB/s per card are the reasons these tenants choose a B200 over a Hopper node.
Through the support page rather than a console setting. Contract supply is arranged with the Clore team: share your pod count, region and the terms you can offer. B200 contracts run in whole 8-GPU pods on terms from 30 days to 1,800, in the USA, Japan and Slovenia, and that route is where the up to ~$3,700 a month per card comes from.
The fabric is not something you add to the node; it is the node. B200 pods are built around NVLink 5 at 1.8 TB/s between GPUs, with NVIDIA quoting 14.4 TB/s of aggregate NVLink bandwidth for an eight-GPU HGX B200 system, and a whole-server rental hands that interconnect to the tenant along with the cards. What you can get wrong is describing it. A renter budgeting for a full NVLink domain who receives something else has bought the wrong machine, so be exact about the chassis in the listing.
On the contract side the unit is the pod: Clore's configurator starts B200 at 8 GPUs. The per-minute marketplace lists servers of any size, so a smaller B200 machine can go there at your own price, and on a multi-GPU server on-demand renters can take a subset of the cards, priced in proportion to how many they use. Most of what a B200 tenant pays for is the NVLink 5 domain, which is why full pods lead.
On Clore.ai a B200 earns up to ~$3,700 a month per card against up to ~$2,750 for an H200, both on the contract route. Renters pay the difference for FP4, for 8,000 GB/s of HBM3e against 4,800, for 180 GB against 141, and for NVLink 5 at twice the speed of the Hopper link.
Marketplace rentals are billed per minute and credited to your Clore balance while the order runs, and the balance withdraws in Bitcoin, USDT, USDC or CLORE. MFP Lock is an optional extra on listed servers, paying MFP rewards in CLORE and leaving more of each BTC, USDT and USDC rental with you. Contract income is paid in USD stablecoin or BTC, your choice, on the terms agreed when the pod is placed.
What tenants load onto a pod, sized in gigabytes of weights before any cache.
Blackwell's four-bit path halves the FP8 footprint, so one pod can hold several copies of a model this size.
Host onboarding →Three of the eight cards hold the weights; the other five stay free for KV cache or a second service.
Bare metal →Sharded training keeps all eight cards in constant exchange, and inside one pod that exchange runs over NVLink 5.
All documentation →Top monthly figure per card on Clore.ai, with the contract shape behind it. Specifications from NVIDIA.
Each figure is the top of that card's range on Clore.ai, per card per month. Every other row links to that card's own host page.
The installer, MFP Lock and payouts, as Clore documents them.
The rental page covers FP4, the 180 GB question, the term ladder and how an eight-card contract is configured.
Register as a host, install the agent on a machine whose driver already works, pair it with your init token and publish your prices, or bring whole pods to the contract route.