Each RTX 6000 Ada can earn up to ~$1,100/month per card on Clore.ai, with the top figure coming from the bare-metal contract route for operators with 8 or more cards on 30-day terms. A single card can start on the per-minute marketplace today, where you set the price and are paid for every rented minute in BTC, CLORE, USDT or USDC. With 48 GB of ECC memory, FP8 tensor cores and 18,176 CUDA cores, it is the card renters with eight-bit models look for.
It is the fastest 48 GB board in the pro line, it speaks FP8, and it can earn on two routes. Three things worth knowing before you list it.
Serving stacks built on eight-bit floating point need Ada-generation tensor cores or newer. A renter with an FP8 model and a 48 GB budget searches for this card by name, and your listing is what they find.
Enable BTC, CLORE, USDT or USDC and renters pay in whichever they hold. Orders bill per minute against your daily price, so every rented hour of a card this capable counts toward the month.
Most of what a 48 GB board attracts runs on one card: FP8 serving, adapter fine-tuning, diffusion and rendering. List the GPU count accurately and those orders come straight to it, with PCIe carrying anything that spans two cards.
Once the RTX 6000 Ada is earning, you can lock CLORE behind the server. That raises what you keep on every marketplace order paid in BTC, USDT or USDC and pays MFP rewards in CLORE for the minutes it is rented on the marketplace. It works on a machine that is already running, and you decide if and when.
Each tier you fill raises what stays with you on Bitcoin, USDT and USDC orders, and partial fills count in proportion. A full lock gives the largest effect.
Tiers 1 and 2 share in the network emission while your server carries an order, each capped at your own daily rental price, so a higher price leaves more room.
Bare-metal contracts carry the top figure. The per-minute marketplace lets a single card start earning today, at a price you set.
Earned at the top RTX 6000 Ada contract rate around the clock, for blocks from 8 GPUs on terms of 30 days. Contracts are made with Clore directly, outside the dashboard, so talk to us about supplying bare-metal capacity.
Register the server, set a daily price for on-demand and a spot floor beneath it, and get paid in BTC, CLORE, USDT or USDC for each minute an order runs.
How and when each currency pays out is covered in the host docs. Payouts and fees →
The ~$1,100 figure describes a card working around the clock. These are the controls in your dashboard that keep an RTX 6000 Ada close to it.
Track USD on any currency and the crypto price follows the exchange rate for you. A benchmark from rented servers of the same model sits right beside your price in My Servers.
On a multi-card server of identical GPUs, on-demand renters can take some of the cards while the rest stay open for the next order. Partial rental is on by default.
Give the server a default background container and it runs whenever no rental is active, stepping aside the moment a renter arrives.
Install, pricing and payouts are covered in one place. Full host guide →
The agent goes onto the Ubuntu or HiveOS system the card already runs under, and your dashboard token registers the machine as it installs.
Ubuntu 22.04 or 24.04 on x86_64, or HiveOS with the same line. NVIDIA GPUs only. Ubuntu 18.04 and 20.04 are past support and will not work.
Your dashboard generates it when you start the add-server flow and fills it into the command. Installed without one already? Run /opt/clore-hosting/clore.sh --init-token <token> afterwards.
Set on-demand and spot prices per currency, and adjust them from My Servers whenever the market moves.
MFP Lock works on a running server. Set the machine up first and add the lock when there is income to build on.
For anything that has to load in one allocation, yes, and it is not close: 48 GB on one board holds a model two 24 GB boards cannot hold at all without the framework splitting it. For everything else the pair wins on total throughput and on resilience, because one failure does not end the order. The listing question underneath the hardware one is simpler: capacity filters are how renters narrow the marketplace, and 48 GB clears a filter that 24 GB does not.
It removes one filter, and a narrow one. Renters doing gradient-heavy multi-card training look for a peer interconnect, your board has no connector for it, and those orders go elsewhere. Single-card work, which is most of what a 48 GB board attracts, never touches NVLink at all. The practical advice for your listing: do not advertise multi-card scaling you cannot deliver, and do not expect the missing bridge to hold back the demand you actually get.
Mostly after. Marketplace filters are built around the things that gate a job outright: GPU model, card count, VRAM, location. FP8 is a property of the model you get when you filter for RTX 6000 Ada, not a separate checkbox. A renter who cares about FP8 already knows which boards have it and searches for the board. That matters to you because it means your listing is found by name, and the name is doing all the work.
Mostly single-card work that needs a lot of memory and fast arithmetic: a 32B model served in FP8 with room left for its cache, adapter fine-tuning, Flux and SDXL image generation, and Blender renders that lean on all 18,176 CUDA cores. Jobs like these tend to book the card for hours at a time.
They start at 8 cards on a 30-day term and are agreed with Clore rather than created in the dashboard. With a full block of cards, contact support to offer the capacity; with one or two, list per minute and set your own price today.
Every rented minute is billed against your daily price and credited as the order runs. You choose which of Bitcoin, CLORE, USDT and USDC to accept, and MFP Lock, if you use it, raises what you keep on the non-CLORE ones and adds rewards in CLORE.
Operational patterns rather than benchmarks, because throughput belongs to the renter's stack and not to your machine. Each of these asks something different of the box the board is sitting in.
A 32B model in FP8 sits resident for the life of the order. Steady power, steady memory, few restarts. From your side this is the best kind of tenant: predictable and long.
Read the guide →Fine-tuning that stays inside one board never needs NVLink, which is why a single-card listing competes well for this work, and long jobs mean long bookings.
Read the guide →A 70B model in eight-bit float is about 70 GB and needs more than one board, while four-bit fits here with room for its cache. Say so in the listing and every renter arrives with a model that fits.
Read the guide →Specs from NVIDIA datasheets for the pro and 24 GB cards a host is most likely to own. Follow a row for that card's host page.
Lock tiers, payouts, the install command and supported operating systems all trace back to these four. Where this page and the docs disagree, believe the docs.
The renter page works through FP8 memory budgets and what 960 GB/s does for generation speed.
One install command on the machine the board already lives in, then a price you choose. Each card can earn up to ~$1,100 a month on contract, and a single card can start on the per-minute marketplace today.