Almost every piece of GPU pricing advice ever written, including most of ours, assumes a renter buys graphics cards by the hour. OctaSpace does not work that way. A renter takes a whole node, the price is set for the whole node, and a hard limit applies to that whole-node figure no matter how many GPUs are inside it. For a multi-GPU host that last detail is not a footnote. It decides whether the machine belongs on this marketplace at all.
Check these details against the source
The unit of sale is the node, not the GPU
On Vast.ai you list a machine and renters can take a slice of it: one card out of four, with the price quoted per GPU per hour. Minimum GPU grouping exists precisely because slicing is the default. On OctaSpace there is no slicing. A renter who wants your four-GPU node takes all four, for the duration, at one price.
Everything follows from that. You have one price to set rather than a per-card rate. Your competition is other whole nodes with the same card in the same quantity, not the general population of that GPU model. And your machine either earns everything or nothing at any given moment, with no partial occupancy in between.
Three prices, not one
An OctaSpace node carries three separate figures, and hosts arriving from a per-GPU marketplace tend to set the first and ignore the rest:
| Field | What it charges for |
|---|---|
| Base | The node itself, per hour. This is the headline rate renters compare. |
| Storage | Disk the renter uses |
| Traffic | Data moving in and out |
The same logic that applies to storage and bandwidth on any marketplace applies here: renters compare the total they will pay, not the headline, and an outlier on a secondary field quietly loses you rentals without ever producing an error message. The full argument is in storage and bandwidth pricing.
The limit that surprises multi-GPU hosts
OctaSpace caps a node's base price at $1.00 per hour. Not per GPU. Per node. The platform stores prices as integers of one ten-thousandth of a dollar, and the maximum accepted base value is 10,000 of those units, which is exactly one dollar.
Divide that by your GPU count and the consequence appears:
| GPUs in the node | Highest possible base | Which is, per GPU hour |
|---|---|---|
| 1 | $1.00 | $1.0000 |
| 2 | $1.00 | $0.5000 |
| 4 | $1.00 | $0.2500 |
| 6 | $1.00 | $0.1667 |
| 8 | $1.00 | $0.1250 |
Where the division does not land on a whole number of ten-thousandths, the usable maximum is the value just below it: our agent writes $0.1666 per GPU on a six-card node rather than rounding up past the limit.
Read the eight-card row against the rest of the market. An eight-card RTX 4090 node cannot be listed above roughly twelve and a half cents per GPU hour on OctaSpace, however strong demand is, because the node price cannot exceed a dollar. That is a fraction of what a single 4090 commonly clears for elsewhere. The limit is not a problem for a one-card or two-card node, where a dollar an hour is generous. It is the single most important fact about the platform for anyone considering putting a dense server on it.
This is a base-price limit, not a revenue limit
How rental status is actually detected
One more mechanical detail, because it explains behaviour you might otherwise find strange. OctaSpace's marketplace feed does not carry a plain rented-or-not flag on a node. What it carries is the GPU performance state: P0, P1, or P2 for a card doing work, P8 for one sitting idle.
So "is this competitor rented" is inferred from whether their cards are busy. In practice that is a good proxy and it is the best signal the platform offers. It is worth knowing about for one reason: a node running idle jobs, described in how to become an OctaSpace host, is working, and working looks like rented from the outside. A market that appears busier than it is will read as stronger demand than it is.
What this means for how you price
- Compare against your own configuration. A four-card node competes with four-card nodes of the same GPU model. A median taken across every node carrying that card mixes single cards with dense servers and tells you nothing useful, for the same reason explained in the market clearing price.
- Convert to per-GPU to think, convert back to write. Your costs are per card, your electricity is per card, and your break-even from the price floor method comes out per card. The listing is per node. Multiply your per-GPU floor by your GPU count before you type anything.
- Occupancy is all-or-nothing. There is no partial rental to soften a quiet week, so the case for pricing to stay rented rather than to look impressive is stronger here than on a marketplace where cards sell individually.
- Check the limit before you buy dense hardware. If the plan is an eight-GPU server, price it against the table above before it arrives, not after.
What GPU Autopilot does with all this
Our agent handles the conversion so you do not have to. It reads the OctaSpace market, filters to nodes matching your GPU model and your GPU count, converts every whole-node price to a per-GPU basis so the comparison is like for like, works out a target on that basis, then multiplies back up and writes a whole-node base price. Your floor is enforced on the result, and so is the platform's dollar limit: a target that would exceed it is brought back down rather than refused, so the node keeps pricing.
Two Vast.ai features do not exist here, and it is better to say so than to let you go looking. OctaSpace machines use Market Match; the other strategy modes are Vast.ai only. And there is no interruptible or spot price on OctaSpace, so the two-tier approach described in the interruptible pricing guide has no equivalent. Your floor, your Max Price, and the region filter work here the same way they do on Vast.ai.
Where to go next
If your node is not up yet, how to become an OctaSpace host covers the install, the unusual filesystem requirement, and what verification actually checks. If you are deciding between marketplaces for a particular machine, compare real demand rather than headline rates: the GPU IQ guide covers reading utilization across platforms, and how QuickPod pricing works describes a per-GPU market for contrast, which is a useful thing to hold next to this one.