Stop comparing GPUs by the hour.
The $/GPU-hour spreadsheet compares one attribute of a product that differs in ten: whose silicon it is, where it sits, who else is on it, whether it vanishes mid-job, and who answers when it breaks. For a steady workload, compare months against months, with the hidden lines surfaced.
In this article 3 sections
- The GPU comparison everyone runs is a column of hourly prices. The column compares silicon-time and hides everything that decides whether your workload survives: tenancy, jurisdiction, hardware provenance, preemption, and who answers at 3 a.m.
- The hourly column does not say whether you are renting a consumer card, a shared machine, or dedicated datacenter silicon, or in which jurisdiction it sits.
- For a steady production workload the honest comparison is total monthly against total monthly: the hourly rate times 720, plus the engineering time the raw box leaves on your team, against one flat managed number.
- And if your workload genuinely bursts for an afternoon, hourly is the right product, from an hourly provider. We sell the other thing, deliberately.
The artifact is always the same. A spreadsheet, six providers down the side, one column that matters: dollars per GPU-hour, sorted ascending. The meeting gravitates to the top row, because that is what sorted columns are for. Nobody in the room asks the only question the column cannot answer: what is this number a price of?
Because a GPU-hour is not a unit of the same good. The price compares one attribute of a product that differs in ten, and every attribute it hides is one your production workload will meet personally.
What the hourly number doesn't encode
Whose silicon is it: a datacenter card, or a consumer card? What tenancy: yours alone, or shared with strangers whose workloads you cannot see? Which jurisdiction: an EU rack under EU law, or somewhere the listing does not state? What continuity: reserved, or interruptible the moment someone outbids you mid-job? And when it dies at 2 a.m., what happens: a warranty chain and an engineer, or a refund of the hour? Every one of those answers is priced in. That is why the cheap number is cheap, and the market is being honest about it, in the fine print.
Silicon
Tenancy
Jurisdiction
Continuity
Support
The price
The comparison that actually decides it
For a steady workload, the only kind this product is for, convert everything to the same unit: a month. An hourly rate running around the clock is the rate times roughly 720 before any hidden line item. Then add the ones the raw box leaves on your team: the driver fights, the CUDA version pinning, the serving-stack tuning, the on-call. Our managed GPU scope explains which operational work sits with us; the GPU sizing brief helps define a configuration to compare against your current spend. Against the summed column stands one flat managed quote with the agreed operations inside it. We are not going to fill in the left column for you, because it is your spreadsheet and the rates move monthly. Bring your workload to the GPU Servers page for a scoped quote.
| THE HOURLY SPREADSHEET | THE MONTHLY ONE |
|---|---|
| Rate times the hours you guess you'll use | One flat number, known in advance |
| THE LINE THAT DECIDES REAL BUDGETSYour engineers' stack-babysitting time: uncounted | The managed layer: included, and it's our on-call, not yours |
| Preemption and re-queue time: uncounted | Reserved: the machine is yours around the clock |
| Provenance and jurisdiction: unpriced risk | Owned datacenter silicon, EU rack, named engineers |
When hourly is the right answer
If your workload is a research burst, an afternoon of fine-tuning experiments, a batch you can checkpoint and resume when the spot instance dies, then hourly elasticity is the right product and the hourly clouds do it well. Buy hours, from someone built to sell hours. This post is for the other workload: the one that runs every day, carries data with a jurisdiction, and needs dedicated datacenter silicon in an EU rack.
“A consumer card, a shared machine, or dedicated datacenter silicon in an EU rack: the hourly price will not tell you which one you bought.”
Straight answers
Why are marketplace GPU prices so much cheaper?+
Because the price reflects the product: it may be consumer cards, shared or interruptible capacity, an unstated jurisdiction, and no support chain.
How do I compare hourly GPU rental with a dedicated monthly server?+
Convert to the same unit: the rate times roughly 720 for around-the-clock use, plus the engineering time a raw box leaves on your team, plus re-run time for preemptions, against one flat managed number. Steady workloads usually stop being an hourly product at that point. Bursty ones don't, and should stay hourly.
Is spot or interruptible GPU capacity usable for production?+
For checkpointable batch work, sometimes, and that is a legitimate engineering decision. For a serving endpoint your product calls, an instance that vanishes by design is not a reliability tier. It is a different product.
Are you cheaper than the hourly clouds?+
For an afternoon, no. For a steady 24/7 workload with the operations included, run the monthly comparison above. The format is the answer.
Running it every day? Compare months, not hours.
Bring us the model and the workload. You'll get one flat number with the engineering inside it, on owned EU metal with your name on it.
See GPU ServersYou’ve reached the end.
Put the reading to work
What are
you running?
Tell us about the workload, traffic and recovery requirements. Start with the server builder, or talk through the details with an engineer.
Written by the engineers who build and run the infrastructure it describes.Talk to an engineer