Skip to main content
← All Posts

GPU cloud pricing is hard to compare on purpose

GPU cloud pricing is hard to compare. These six normalisations turn any offer into one comparable number: €/GPU-hour.

By Leafcloud
Published on

GPU cloud pricing is hard to compare on purpose

We've noticed GPU cloud pricing is hard to compare. Credits, compute units and subscription tiers all resist landing in the same spreadsheet.

Six normalisations turn any offer into one comparable number: €/GPU-hour.

Every provider quotes in its own currency of convenience. Monthly subscriptions denominated in "units". Prepaid credits with an expiry date. Hourly rates with a minimum commitment. "From" prices attached to hardware you will never actually be allocated. None of these are directly comparable, which suits everyone quoting them. So before you compare anything, normalise.

GPU Cost Comparison

40h
#1Leafcloud
0.00
/month
Free tier — €0 until you exceed 50 hours
#2Google Colab Pro
11.99
/month
1 subscription needed
#3RunPod (on-demand)
18.00
/month
Per-hour pricing, no free tier
#4Lambda Labs (spot)
22.00
/month
Spot pricing, may be preempted

1. Name the actual GPU

Class, generation and VRAM, and whether the allocation is dedicated or shared and preemptible. A slice with guaranteed VRAM and an "up to" fraction of a card are different products.

2. Compute the effective €/GPU-hour

Divide what you actually pay per month by the hours you actually compute. Subscription tiers look cheap until the divisor is realistic.

3. Price the idle time

Per-second billing with a monthly minimum is a subscription wearing an hourly rate. Work out what an idle week costs you. If the answer is not zero, the hourly rate is not the price.

4. Count the peripherals

Storage, egress and static IPs. Egress in particular, because that is the line that turns a cheap training run into an expensive model download.

5. Check expiry and lock-in

Do the credits expire? Does the rate require a term commitment? An expiring credit is a discount only if you use it.

6. Price the interruption risk

Spot and preemptible prices are real savings for restartable jobs, and a hidden tax on everything else. Price your own checkpointing discipline realistically before you count that discount.

Worth Considering

If your workload is large-batch, interruptible training, spot capacity on bigger cards elsewhere can win on raw price per FLOP. If your usage is a couple of hours a month, a free notebook tier is the right answer. And check the currency. Most GPU pricing is quoted in dollars, so make sure you're comparing apples to apples, including any conversion fees.

join the waitlist for the free tier and test it on your own workload. If you would rather talk the numbers through first, book a call with our team.

Related Articles