Every NVIDIA L40S can earn up to ~$930/month per card on Clore.ai. That top figure is the bare-metal contract route, open to operators supplying 8 or more cards on 30-day terms. One server can also join the per-minute marketplace, where you set the price and each rented minute is paid in BTC, CLORE, USDT or USDC. FP8 tensor cores and ray-tracing hardware let the same card serve models on one order and render scenes on the next.
Eight-bit arithmetic, ray-tracing hardware and 48 GB of ECC memory. Each draws a different kind of renter, and every one of them pays by the minute.
Fourth-generation Ada tensor cores run eight-bit floating point natively, something no Ampere card can do. A renter serving an FP8 model inside a 48 GB budget has a short list of cards to choose from, and the L40S is on it.
The L40S keeps the ray-tracing hardware that the H100 dropped, so a Blender queue and an inference endpoint can share the same machine on different orders. Two kinds of renter can book the same card.
Each rental takes all 48 GB and every tensor core, one renter at a time. Multi-GPU servers talk over PCIe, so list the GPU count accurately and the orders that fit will find you.
Both are optional and both come on top of what the card earns per minute. CLORE locked against your L40S server collects MFP rewards while it is rented on the marketplace, and it raises what you keep on marketplace orders paid in BTC, USDT or USDC. How much you may lock follows your server's MFP score.
Those two tiers share a slice of network emissions, and each is sized against your server's daily rental price. The reward therefore grows with what you charge rather than with how much you lock.
Tier 1 fills first, then Tier 2, then Tier 3. The more of that capacity you fill, the more of each BTC, USDT and USDC payment stays with you.
Operators with eight or more cards can supply bare-metal contracts for the top figure. Anyone with a single server can list it per minute and name the price.
It is what the highest L40S bare-metal contract rate comes to over a full month around the clock, in blocks of 8 GPUs or more on 30-day terms. Nothing to switch on in the console: contracts are agreed with Clore, so talk to us about supplying bare-metal capacity.
List one machine and it can be booked at any hour, at the daily price you set, with every minute of an order paid.
Marketplace orders pay in Bitcoin, CLORE, USDT or USDC, whichever you enable, so income lands in the coins you want to hold.
The host docs explain when and how each currency pays out. Payouts and fees →
The top figure assumes a card that is working around the clock. Three inputs decide how close your own listing gets to it.
Set in USD per day in the dashboard and converted to Bitcoin and CLORE automatically. Spot is entered separately and lower, for renters who accept interruptible orders.
An L40S that stays online can sell any hour of the month. Uptime, a clear description and a price near the market are what fill those hours.
Lock CLORE against the server for MFP rewards on rented minutes, plus more of every per-minute BTC, USDT and USDC payment kept.
The pricing guide walks through daily prices, spot floors and USD tracking. How to price a server →
The agent goes onto a machine that already runs a supported OS, and a token from your dashboard registers it as it installs.
Ubuntu 22.04 or 24.04 LTS on x86_64. The documented minimums are 8 GB of RAM, 16 recommended, and 32 GB of disk with 30 GB free. The NVIDIA driver must be installed and working before the agent goes on.
My Servers, Add Server, then either the quick one-step registration or the manual route. Either produces an install command carrying a token unique to that machine.
Choose on-demand and spot prices per currency. A traffic-light badge on each server compares your price with what renters pay for the same GPU, and you can adjust it at any time.
MFP locking is optional and can be added whenever you like. Capacity follows your server's score: MFP × 2,000 for Tier 1, × 4,000 for Tier 2, × 3,000 for Tier 3.
Yes. Partial GPU rental covers on-demand orders on rigs whose GPUs are all one model, which lets a four-card L40S server carry more than one order at once. You can opt a server out if you prefer to rent it whole.
It brings renters whose serving stack has an FP8 path, which the Ampere A6000 cannot offer at all. Both cards hold 48 GB with ECC and both land a 70B model at INT4 with cache to spare, so on capacity they are interchangeable. The L40S adds fourth-generation Ada tensor cores with FP8 and 864 GB/s against the A6000's 768. Renters with eight-bit models search for FP8-capable cards by name, and at 48 GB the L40S is one of two such cards on Clore.ai.
Two kinds of work, often on the same card. FP8 inference is the first: a 32B model in eight-bit float takes about 32 GB and leaves around 16 GB for its cache. Rendering and video generation are the second, using the RT cores the H100 left out. A 70B model also fits at INT4, at roughly 40 GB of weights.
As whole cards. Without MIG each L40S is rented complete, one renter at a time, and without NVLink multi-GPU renters on your server work over PCIe. Neither stops anything from running, and single-card jobs, which is most of what a 48 GB board attracts, never touch either feature. State both in the description and renters book with the right expectations.
Bare-metal contracts start at 8 GPUs on a 30-day term and are agreed with Clore directly rather than switched on in the console. Operators running eight or more L40S cards reach out to support to put that capacity forward, while a single L40S in a single server can go on the per-minute marketplace today.
In the currency the renter used, credited as the order runs and billed per minute against the daily price you set. Bitcoin, CLORE, USDT and USDC are all accepted for marketplace orders. Locking all three MFP tiers raises what you keep on the Bitcoin and stablecoin ones.
The two cards an L40S owner most often weighs it against, and the features that send renters to the Ada card.
Eight-bit tensor cores for serving and ray-tracing hardware for rendering, fed by 864 GB/s of memory bandwidth.
Payouts and fees →Same capacity, but the Ampere card has no FP8 units, so renters with eight-bit models pick the Ada card.
A6000 hosting →The H100 brings HBM3 and MIG. The L40S keeps the graphics pipeline Hopper left out, so render work stays on its side.
H100 hosting →Datasheet specs for the cards an L40S operator compares it with. Follow a row for that card's host page.
H100 figures are the SXM5 part
Payouts, lock capacities and install steps on this page all trace to the four documents below. Where this page and those disagree, follow them.
The renter page covers what fits in 48 GB, FP8 serving, and how the card compares with the HBM parts.
One install command and a daily price put a single L40S on the per-minute marketplace. Operators with eight cards or more can also supply contract capacity at up to ~$930 a month per card.