Renting out your GPU on vast.ai when it's not in use?¶
Machine translation of the French original.
Answer: No, not in our configuration. Analysis performed on 2026-10-22 for the RTX 4070 Ti 12 Go of vm-compute-01. The idea is appealing – the card is inactive for a good portion of the day, so why not use it for others during that time? The idea runs into three main issues, from most to least serious. The conclusion changes if any of them change: this is specified at the end.
1. The income barely covers the electricity¶
Prices obtained on 2026-10-22 using the public API of vast.ai (on-demand, 1 GPU):
| Card | Offers listed | Minimum price | Median price |
|---|---|---|---|
| RTX 4070 Ti | 21 | 0.107 $/h | 0.173 $/h |
| RTX 4070 | 20 | 0.080 $/h | 0.134 $/h |
| RTX 3080 | 55 | 0.067 $/h | 0.153 $/h |
These are requested prices from hosts, not actual prices received: an offer listed is, by definition, an offer that no one is renting at that time.
The calculation, with explicit assumptions:
- Consumption: A rented card runs at full load, consuming approximately 250-300 W more on the wall. At 0.30 €/kWh (assumption, residential Belgian rate), ≈ 0.08 €/h.
- Platform commission: Taken from what the host earns.
- Utilization rate: A machine with no history of reliability will be sorted by tenants based on this criterion. 20-40% is realistic initially, and the market for 12 Go consumer cards is saturated.
Rough estimate: 730 h × 30% utilization × (≈ 0.15 € − 0.08 € electricity − commission) ≈ 0 to 20 € per month, before tax – this is income that needs to be declared. It can be negative if utilization drops or if the equilibrium price falls.
2. Containers of unknowns on the main host¶
-
This is the key difference, regardless of the cost.
-
vm-compute-01 is not an isolated machine: it is the primary Docker host for the cluster (approximately 85 containers), with network mounts to storage and access to the local and private network. Hosting for vast.ai involves running the code for any tenant, with direct access to the GPU. A container escape or a GPU driver vulnerability then gives access to the entire cluster.
- vast.ai requires ports accessible from the internet to the host machine. Our principle is the opposite: a single entry point exposed, the front-end of the NAS (see reverse proxy chain). Opening ports to an internal machine breaks this principle.
- Doing this properly requires an isolated network (dedicated VLAN, firewall between this network and the rest). The current router does not allow this properly.
3. "When it's not in use" doesn't work as imagined¶
- During a rental, the card belongs to the tenant. A on-demand rental does not end before its expiry date: it is impossible to take it back because we need it now. Removing the rented machine reduces its reliability score, therefore reducing its future rentals.
- The hosting software reconfigures Docker and the NVIDIA runtime, and essentially assumes a dedicated machine. In our case, the same card is used for local inference, transcription, image recognition, and document processing, with a reservation mechanism that allocates video memory to each. The two systems would compete for the card.
What would change the answer¶
| If… | Then |
|---|---|
| the card is replaced with a larger one | selling it used will generate more profit, and immediately, than years of renting |
| you have a dedicated machine on an isolated network | the hosting becomes defensible; first measure the actual consumption with a meter |
| the median price received exceeds the electricity cost | recalculate section 1 using current prices |
How to recalculate the measurement¶
Prices can be read without counting, using the public API for searching for offers:
curl -s -G "https://console.vast.ai/api/v0/bundles/" \
--data-urlencode 'q={"gpu_name":{"eq":"RTX 4070 Ti"},"num_gpus":{"eq":1},"type":"on-demand","limit":500}' \
| python3 -c "import json,sys,statistics as s; p=sorted(o['dph_base'] for o in json.load(sys.stdin)['offers']); print(len(p), p[0], s.median(p))"
The actual idle time of the card can be read from its usage history (GPU occupancy rate over 30 days): this is the key factor, and not the impression that it "doesn't get used often," that limits the potential income.
Dernière vérification de cette page : 2026-10-22