The cheapest cloud GPU listings start at $0.11 an hour — a T4 on Google Cloud spot, in gputracker's April 2026 budget tier across 54 providers. The five platforms with the lowest catalog averages span $0.34–$1.12/hour, and the whole low end of that list is marketplace or spot inventory: third-party hardware, no SLA, reclaimable mid-job.
01. How cheap does the bottom tier go?
The budget tier runs from $0.11 to $0.46 an hour.
gputracker's April 2026 sweep of 54 providers puts four listings in its budget tier, and not one of them is an on-demand datacenter contract:
| Card | Where it is listed | Hourly price | Listing type |
|---|---|---|---|
| T4 | Google Cloud spot | $0.11 | spot |
| RTX 4000 Ada | Vast.ai | $0.17 | marketplace |
| RTX 4090 | Vast.ai spot | $0.38 | marketplace spot |
| L4 | Google Cloud spot | $0.46 | spot |
The T4 and the L4 are older datacenter inference cards, the 4000 Ada is a workstation card, and the 4090 is the consumer flagship with 24GB of VRAM. What the four listings share is not the silicon but the pricing mechanism: three of the four are spot listings, and spot listings are cheap because the price moves when the host wants the hardware back. That is the whole trade compressed into one table row.
The 4090 lands at $0.38 on spot — inside its on-demand range, near the bottom.
gputracker's budget list carries the 4090 at $0.38/hour on Vast.ai spot. Coloprice's October 2026 sweep of on-demand 4090 listings runs $0.13–$0.74 across platforms, so a spot row near the bottom of that range is not a typo; it is the auction doing exactly what auctions do. The same mechanism that pulls the number down can pull the machine out from under the job, and that part of the trade is section 03's subject.
02. Which platforms average the lowest prices?
Vast.ai tops gputracker's cheapest-platforms list at a $0.34 catalog average.
Averaged across every listing in its catalog, gputracker's April 2026 ranking of the five cheapest platforms reads:
| Rank | Platform | Catalog-average hourly price |
|---|---|---|
| 1 | Vast.ai | $0.34 |
| 2 | RunPod Community | $0.44 |
| 3 | Shadeform | $0.61 |
| 4 | Lambda | $0.89 |
| 5 | RunPod Secure | $1.12 |
Read the order as an inventory ranking, not a quality ranking. The platforms in the top three rows list machines owned by individuals and small operators; the two at the bottom run inventory in facilities the operator controls. The catalog average is largely a census of whose hardware is for rent.
One card shows the tier gap: $0.34 versus $0.74 on the same platform.
A single SKU makes the split visible in one snapshot. Coloprice's October 2026 prices list the 4090 at $0.34 on RunPod Community and $0.74 on RunPod Secure — same card, same operator, same week. The difference is who owns the rack it sits in and what happens on the other end when the machine goes away. That difference follows the same line the catalog averages draw, one card at a time.
A catalog average is not your price.
An average blends every card, every listing type, and every host on the platform into one number. It works as a pointer and fails as a quote: the listing you actually book has its own price, its own billing increment, and its own rules about who can take the machine back and when. Treat the ranking as a map of where the cheap inventory lives, then read the specific listing before the meter starts.
03. What does the low end cost you?
Everything at the cheap end of the list is marketplace or spot inventory.
gputracker's write-up of its own ranking reads as a caution as much as a deal sheet: the low rows are marketplace and spot listings, which means hardware the platform does not own, no service-level agreement standing behind the price, and a reclaim right that can end a job mid-run. That is the mechanism the discount is made of. None of it is hidden — and none of it appears in the hourly number either.
No SLA means nobody owes you a rerun.
On a marketplace listing, the contract is the listing page: a price, a card, a host reputation score. If the machine dies an hour into a job, the compensation is that the billing stops. A managed datacenter listing costs more precisely because there is an operator on the other side of it — the $1.12 catalog average at the bottom of gputracker's list is, in large part, that difference being priced in.
Spot means the host can take the card back.
Spot and bid-priced listings are cheaper because they are interruptible by design: when demand rises, the host reclaims the card and the process on it stops. For a queue of rerunnable renders, an interruption costs minutes. For a job whose state lives only in that machine's memory, it costs the entire attempt. The discount is payment for carrying that risk yourself, and the honest way to shop the bottom of the list is to price the retry, not just the hour.
04. Floor or median: which number belongs in the budget?
The 4090 floor is $0.14; the median is $0.49 — 3.5× apart.
On Vast.ai, the RTX 4090 floor sits at $0.14/hour and the median at $0.49/hour. Coloprice's tracker puts the same market floor at roughly $0.13–$0.14, so the two sources agree on where the bottom is. The floor is one listing — one host, one machine, one price that usually disappears the moment somebody books it. The median is where an hour settles when you need a machine that actually exists.
Budget-tier prices are floor-type numbers.
The $0.11 T4 and the $0.38 4090 in section 01 are single rows in an auction, not steady quotes. The rule the 4090 market teaches applies to the whole budget tier: plan on the settle price, read the listed bottom as luck. That does not make the budget tier useless — it makes it a place to watch and pounce from, rather than a number to build a monthly plan on.
The full floor-and-median board lives in one place.
We keep the card-by-card floor and median table on our GPU rental price comparison page, and the working method — the median of the smallest card that finishes the job, with the meter off when you stop — in the cheap GPU rental guide. Both quote the same numbers every article on this site uses, so a median here and a median there never disagree.
05. Which tier should the workload pick?
Pick by the cost of one interruption, not by the bottom of the price list.
The decision variable is what a reclaim costs you, and it sorts cleanly:
- Rerunnable queues and overnight batches: the bottom tier. An interruption costs a rerun, and the spot discount has already paid for it.
- Deadline work with a client waiting: the managed tier. Lambda's $0.89 and RunPod Secure's $1.12 catalog averages buy operator-controlled inventory and a restart you do not have to sit through.
- Multi-card enterprise jobs: a different market entirely — the same gputracker snapshot prices the enterprise tier at $5.25/hour for an H200 on CoreWeave.
If the deliverable is files rather than a machine, the tier question shrinks.
Most platforms in the tables above rent a server you administer: an OS, a disk, a meter that runs while you set up. When the deliverable is finished images or clips, administration time is overhead the hourly rate never shows. Choosing a tier well reduces that overhead; choosing a task-shaped product removes it — which is a different row than any of the five above.
Where SpareHive fits in this table.
SpareHive is not in the live price table. The planning band sits roughly 50–60% under the $0.49 RTX 4090 median, list price unset, rentals not open. Join the waitlist at sparehive.com.
FAQ
What is the cheapest cloud GPU per hour?
Listings start at $0.11/hr — a T4 on Google Cloud spot in gputracker's April 2026 budget tier. Everything else near that number on the same list is spot or marketplace inventory too.
Is a $0.14 RTX 4090 a price I can plan on?
No. $0.14/hr is the floor and $0.49/hr is the Vast.ai median, 3.5× apart. Plan on the median; treat the floor as a listing that is usually gone by the time you click it.
Why do marketplace platforms average less than managed ones?
Because the inventory is not the same thing. Marketplace rows are third-party machines with no SLA behind them and a reclaim right; managed rows are racks the operator controls. The $0.34 and $1.12 ends of gputracker's average list price that difference in.
Does cheap mean the job will fail?
Cheap means the risk moved to you: third-party hardware, no SLA, interruptible by design. Whether that trade works depends on what a restart costs your job — a rerunnable queue does fine at the bottom of the list, and a one-shot deadline job does not.