Each NVIDIA A40 on Clore.ai can earn up to ~$500/month per card. That top figure is the bare-metal contract route, for operators who supply 8 or more A40s on 30-day terms. A single server can also go on the per-minute marketplace, where you set the price and every rented minute is paid in BTC, CLORE, USDT or USDC. Forty-eight gigabytes of ECC memory and NVLink pairs are what long training and serving jobs go looking for.
A datacenter build, a fast card-to-card link and 48 GB of error-corrected memory. Each one brings in renters who book for hours at a time.
The A40 is a datacenter part, designed for server chassis and a duty cycle that never pauses. That is the profile a month-long contract or a multi-day marketplace order needs, and it suits the block contracts behind the top figure.
Two bridged A40s exchange data at about 112.5 GB/s, faster than PCIe, which is what tensor and pipeline parallel work leans on. List the pair as two 48 GB cards joined by a fast link and the renters splitting models across GPUs will find it.
Each A40 rental takes the full card, so every order puts all 48 GB and 10,752 CUDA cores to work for one renter. INT8 and INT4 serving, 70B-class models and large-scene rendering are the natural fits.
MFP Lock is optional. CLORE locked against your A40 server raises what you keep on marketplace rentals paid in BTC, USDT and USDC, and it earns MFP rewards in CLORE while the server is rented per minute. Both sit on top of what the per-minute route already brings in.
Ten percent of network emissions goes to each of those two tiers, and locked servers share it for the minutes they carry an order. The more your listing earns per day, the more room each tier has to pay out.
Tier 1 takes MFP × 2,000, Tier 2 MFP × 4,000 and Tier 3 MFP × 3,000. The fuller the lock, the more of every BTC, USDT and USDC payment you keep.
Supply a bare-metal contract for the top figure, or list a server per minute at a price you choose. Per-minute income arrives in the currencies you accept; contract income is paid in USD stablecoin or BTC.
That is an A40 earning around the clock at the top bare-metal contract rate for this card, in blocks from 8 GPUs on 30-day terms. Clore sets these contracts up directly instead of through the console, so talk to us about supplying bare-metal capacity.
Register a single machine, publish a daily price and go live. Renters book by the minute and every rented minute is paid.
Bitcoin, CLORE, USDT and USDC. Enable the ones you want and renters pay in whichever of them they hold.
Payout timing and currencies are set out in the host docs. Payouts and fees →
A single A40 server can start earning on the marketplace today, and a fleet of eight or more can move up to a contract. Each step is a setting or a conversation, with no new software to learn.
Install the agent, set a daily price and the A40 is in front of renters. The market-position badge in My Servers shows how your price sits against rented servers carrying the same GPU.
Mass onboarding through cloud-init, SSH, Ansible or HiveOS registers a whole row of servers at once, and pricing by GPU model updates every A40 you own in one step.
With eight or more A40s you can take the contract route, which pays up to ~$500 a month per card for the length of the term.
Fixed crypto prices, USD tracking and spot floors are all covered in the pricing guide. Pricing guide →
The agent installs on an existing Ubuntu system, or a HiveOS rig attaches through a flight sheet. Either way your token registers the machine while it installs.
The installer expects a working NVIDIA driver. From My Servers, Add Server produces a one-line command carrying that machine's token, run as root.
If you have a bridged pair, list two 48 GB cards with a 112.5 GB/s link between them. Precise listings bring renters back for the next job.
On-demand first, spot lower if you want interruptible orders as well. USD tracking keeps the crypto amount in step with the exchange rate.
Once eight or more A40s are online, the bare-metal route opens up alongside your marketplace listing.
No. The per-minute marketplace takes a single server carrying a single A40, and you set its price. Eight cards is where the bare-metal contract route begins, the route behind the ~$500 figure, so a fleet can start small and grow into it.
Ubuntu 22.04 or 24.04 LTS on x86_64, or a HiveOS rig, with a working NVIDIA driver already in place. CLORE's documented minimums are 8 GB of RAM with 16 recommended, and 32 GB of disk with 30 GB free. A40 cards are designed for server chassis, which is where most of them already run.
A bridged pair gets roughly 112.5 GB/s between the two cards, which is faster than routing the same traffic over PCIe and genuinely useful for tensor or pipeline parallelism. Your server is registered through the dashboard as one machine with the GPUs it has, so there is no separate NVLink listing type to choose. What matters is what you write in the listing, and what you must not write is that a bridged pair offers 96 GB. It does not. The two cards keep 48 GB each and a model larger than that still has to be split by the renter's framework.
It does not, and the badge is precisely why people assume otherwise. Inside the Ampere generation the partitioning feature went to two products only, the A100 and the A30. The A40 is neither. It is GA102 silicon, the same die as the RTX A6000 sitting beside it in CLORE's pro tier, and that die never received the feature. Verified against NVIDIA's supported-GPU list on 21 September 2026. In listing terms it means one order per card, and every renter gets all 48 GB to themselves.
Contracts start at 8 A40 cards on a 30-day term. They are put together with Clore directly, so an operator with eight or more cards contacts support to offer the capacity. Smaller fleets list per minute today and can move to a contract as they grow.
A number in US dollars per day, entered in the dashboard, which the platform converts to Bitcoin and CLORE at the prevailing rate. If you want interruptible orders as well, you set a second, lower figure for spot. Orders then bill per minute against whichever applies, and you can change either figure at any time from My Servers.
What renters bring to a 48 GB Ampere card with ECC and NVLink. Capacities come from parameter counts rather than benchmarks, because speed belongs to the renter's own stack.
A 70B model quantized to INT4 fits on a single A40 with room left for its cache, and serving endpoints tend to run for long stretches.
Host setup →Renters splitting a model across GPUs want a link faster than PCIe, which makes a 2-GPU A40 server the listing they filter for.
Pricing guide →Scenes that need every gigabyte resident keep a card busy for hours, and each of those hours is billed by the minute.
Payouts and fees →Datasheet specs for the pro-tier cards an A40 operator is most likely to run alongside it, plus which of them can go on a bare-metal contract. Follow a row for that card's host page.
NVLink on the three Ampere cards joins them in pairs, with each card keeping its own 48 or 24 GB
Install requirements, payouts and MFP capacities were taken from these four pages. Where a number here differs from one there, theirs is correct.
The renter page covers the contract terms, the 48 GB field it competes in, and the NVLink misreading in more depth.
Install the agent, set a daily price and go live on the marketplace, or bring a fleet of eight or more to a bare-metal contract paying up to ~$500 a month per card.