Comparing platforms by hourly rate alone is a mistake: the rate says nothing about how many hours the card is actually paid for. We went down this road twice: Clore in 2023, then Vast.ai from June 2024 — where the total has reached $23,638. Below are seven parameters that really affect a host's income, with numbers from our own fleet of six machines and ten cards.
A caveat about fairness. We saw Clore in 2023 and much has changed there since — so the Clore column is given as "how it was when we joined". Everything on Vast.ai is measured on our own fleet over 2024-2026; every machine's income is open on the "Fleet" page of our site, Vast Sale: our live fleet page.
The short answer
On most of the seven points Vast.ai gives more control: the host receives 100% of their rate, payouts are weekly and in money rather than a token, and there is enough demand for our fleet to hold 90% average utilization through August 2026. But "more predictable" does not mean "no trouble" — the two sections below cover what broke on our side and where we got it wrong.
Seven points to compare
- Demand and average utilization — how many of the month's 744 hours are paid for.
- Pricing — who sets the price, and whether the client sees a different number.
- Fees and payouts — in what, how often, above what threshold.
- Entry barrier — bandwidth, OS, verification and your own skills.
- Tooling and API — whether the fleet can be run by scripts.
- Support — how a disputed charge is investigated.
- Score — what drags it down, and what that costs in money.
| Parameter | Clore (as it was when we joined, 2023) | Vast.ai (our fleet, 2024-2026) |
|---|---|---|
| Utilization | never gathered enough data | 90% average in August 2026, six machines |
| Who sets the price | the host | the host; client sees +33% on top, not taken out of your rate |
| Payout currency | platform token or bitcoin | money |
| Payout frequency | never fully checked | weekly, $20 threshold |
| Bandwidth requirement | saw no explicit threshold | ~2.6 Mbit/s per 1 GiB VRAM, floor 100, cap 500 |
| Automation | not checked | host CLI and API |
| Cost of downtime | — | minutes offline cut the score; IP change during a rental ~ -0.10 |
1. Demand: counted in paid hours, not in enquiries
Through August 2026 our six machines with ten cards brought in $1,737 at 90% average utilization. The spread is almost fivefold: the machine with four RTX 4090s made $906 — 52% of the fleet's income from 40% of the cards — while a single RTX 5080 made $79. The row-by-row breakdown: our income breakdown card by card.
90% is a monthly average, not 90% every day, and it is made of demand, your price, your bandwidth and the machine's score. The platform is responsible for the first one only. We had a month when a card dropped by almost half, and the marketplace had nothing to do with it: we were holding the price high ourselves while preparing the machine for a rebuild.
2. Pricing: the host receives 100% of their rate
The main thing to understand before any arithmetic: the marketplace's markup is added on top for the client, not deducted from your money. Measured on our own machine: $0.21/hour from the host becomes $0.28/hour for the client — about +33%. You set the price yourself, spot is cheaper than on-demand, and both can run at once.
The median of the Vast.ai catalogue on 25 September 2026 for single cards available to rent, converted to the host's price:
| GPU | Client $/h (median) | Host $/h (median) | Host, range |
|---|---|---|---|
| RTX 5090 | 0.63 | 0.47 | 0.40 - 0.55 |
| RTX 4090 | 0.49 | 0.37 | 0.33 - 0.45 |
| RTX 5080 | 0.31 | 0.23 | 0.20 - 0.25 |
| RTX 3090 | 0.19 | 0.14 | 0.11 - 0.24 |
From there the arithmetic is simple: income = host rate × 730 × utilization. One RTX 4090 at the median $0.37 yields about $81 at 30% utilization, about $135 at 50% and about $189 at 70%. Plug your own numbers into the calculator at the income calculator; the cost side and the payback method are in our income breakdown card by card.
Beyond GPU hours a host sells storage and traffic. Storage is usually $0.10-0.20 per GB per month and accrues even while an instance is stopped — on a large disk that is comparable to the card's income. Traffic has a catch, and it is in point six.
3. Payouts: what exactly you are paid in
On Vast.ai payouts are weekly with a $20 threshold; until the threshold is met, earnings sit on your balance. That is boring — and that is the entire point.
On Clore in 2023 you could only take your earnings in the platform's coin or in bitcoin, and the fee was very high. We were escaping the volatility of mining and ran into it again through another door: the money arrived in a token whose price we do not influence, and getting out of it cost a noticeable share of the earnings. If the settlement currency jumps around, there is no stable income at any hourly rate.
4. Entry barrier: bandwidth, OS and verification
Technically almost any card can be rented out. The barrier is not the card, it is the environment around it.
- Bandwidth. The Vast.ai requirement scales with video memory: about 2.6 Mbit/s per 1 GiB of total VRAM, but no lower than 100 and no higher than 500 Mbit/s, up and down separately. One RTX 5090 (32 GB) hits the 100/100 floor; a server with four RTX 4090s (96 GB) already needs 250.
- Symmetry. Almost any plan clears the formal upload threshold, but in the catalogue's filters you lose to symmetric machines. In our measurements a server on Realtek 2.5GbE gives 706/625 Mbit while one on Intel I225-V gives 524/518: judge by the measurement. Test your own link: the network check.
- Blocking. In Russia you need a tunnel through your own VPS, and DPI is not the whole story: our download was throttled by the hosting provider's AS — OVH and M247 were choked, Private Layer was clean. Changing provider fixes it, tunnel settings do not. More: what network a host needs.
- OS. We install nothing but Ubuntu 22.04 Server, step by step: installing Ubuntu for Vast.ai. Newer releases on rental machines we consider raw.
- Verification. Until it passes, the machine is not visible to every client. What to check before moving is in the checklist: the pre-launch checklist.
5. Tooling: a fleet is run with scripts
For a single card this point does not matter; for six machines it decides everything. Vast.ai has a host CLI and API: show machines for the fleet's state, search offers with verified=any and rentable=any — the only way to see a machine's internal flags — and send_mach_info.py --speedtest to re-measure bandwidth.
A detail the documentation does not mention: the marketplace stores not your latest speed test but a moving average — roughly 0.9 of the old value plus 0.1 of the new one. One test run after a move changes almost nothing: it took us ten runs for the figure to climb from about 570 to 578/595 Mbit. The cheat sheet: our Linux command cheat sheet.
6. Support: tested by a disputed charge
Response time is a poor metric. The real test is what happens when the money does not add up. Our case: a machine priced at $3 per terabyte, a client in virtual-machine mode pushed 8.32 TB over 37 hours, and the report showed Upload $0.00 and Download $0.00. We verified the volume from three independent points, and it held up.
The explanation: traffic is billed only on docker rentals, and Vast does not meter traffic on virtual-machine rentals at all. The traffic price field can be filled in, but it will do nothing. Because of that we carried income that does not exist in our figures for months. If you plan on VM rentals, strike traffic out from the start.
7. The score: where downtime turns into money
The reliability score decides how often clients come to you, and it falls faster than it recovers. Changing the IP of a machine under rental costs about -0.10, while outside a rental it is not penalized at all. One episode where we noticed a broken link to the marketplace too late cost a drop from 0.99 to 0.79 on a machine with a client on it; on an empty one the same failure passed almost without loss.
Hence a rule we did not arrive at immediately: first check whether the machine has a client, and only then decide whether to fix it now or after the rental.
What actually breaks: our log
None of these failures has to do with the choice of platform — they happen on any of them. Yet it is these, not the difference in rate, that decide the final figure.
- The marketplace controller getting stuck. The agent memorizes a controller's address and clings to it even after it dies: our requests go out, no replies come back, and the machine is still listed and verified. On 25 August four machines went down like this; on 5 September, all six in one day. The fix is clearing the /var/lib/vastai_kaalia/controller_connection cache; the full diagnosis: the stuck controller write-up.
- The VM support flag is wiped by the platform itself: the hourly enable_vms.py job looks for a drm/renderD* node on the cards, and when every card is handed to vfio there is no such node. One of our machines stood without VM mode for ten days: from 19:22 on 11 September to 17:04 on 21 September.
- A dead memory module. memtest stayed silent; the machine fell over under load. We proved it by swapping modules: the failure disappeared once the suspect module was pushed out to swap.
- The agent hanging on a blocking HTTP request to the platform's web API: no timeout on the call, and the service packets stop going out. From the outside it looks like a bandwidth problem, while the link is fine.
- A tunnel loop before the DHCP address arrives: packets aimed at the VPS fell back into the tunnel, and within minutes 28,000 sockets piled up — the entire port range. The router's connection table overflowed, and on 20 September that triggered a false failover of four machines to the backup VPS.
Where we got it wrong
This is the section missing from retellings of other people's reviews. Four mistakes of ours; each cost money or time.
- We counted income from traffic. The $3 per terabyte went into our payback calculation, and on VM rentals it is zero — see point six; the calculation was inflated.
- We treated the wrong illness. When a card died after being passed through to a virtual machine, we blamed "the incurable RTX 4090 reset bug". The diagnosis is wrong: the reset bug belongs to Blackwell, whereas our card dropped into deep sleep when detached from the guest and never came back — which device rules fix. Because of that false diagnosis the four-4090 machine ran for a long time without VM mode.
- We bred our own workarounds instead of using the built-in ones. Our route-monitoring script pulled the marketplace's service traffic out of the tunnel and took five machines offline; we switched it off on 25 July. The rule now: the built-in tool first.
- We set thresholds far too sensitive. In the automatic failover to the backup VPS the threshold was three failed checks in a row. On 20 September interference on our home link produced exactly those three, and four machines moved to the backup address for no reason — and changing IP penalizes the score. The threshold is now nine.
How to compare for yourself, not through our article
If you are choosing between platforms right now, this is the order we would follow.
- Measure your link in both directions and specifically to the marketplace: the network check.
- Work out the income from the catalogue median for your card model, not from the best offer: the income calculator.
- Count costs after electricity. Our fixed costs are about 10,000 RUB a month for electricity and two VPS, roughly 6% of revenue. The method: our income breakdown card by card.
- Check the platform's rules for payout currency, threshold, frequency and withdrawal fee.
- Sanity-check the order of magnitude against open numbers: each machine's income is on our live fleet page.
- Start with one machine and watch it for a month: a single day shows nothing.
The bottom line: Vast.ai wins not by being "better", but by making it clear what your income is made of — the rate is yours, the markup is separate, payouts are in money, downtime and score are visible as numbers. The full breakdown of moving a fleet, with commands, installation and pricing, is in the paid guide: the full migration guide. If you want us to look at your hardware, bandwidth and electricity tariff, write to us: https://t.me/flux_vast.