GPU Cloud News still treats “$X/hr” as one product. It is not. On-demand is a published hour you can start. Spot is a bid against leftover capacity. Reserved is a contract. A quote is a meeting. If a GPU cloud mixes those in one table, the cheapest cell is usually the one you cannot run.
I built GPUBeacon so the catalog can keep the flags honest. Filter on-demand GPUs first. Spot stays visible. It does not get to impersonate a price. If you are about to rent an H100 off a roundup that does not say interruptible, stop.
On-demand is the product
On-demand means you pay the listed hourly rate and the machine is yours until you stop it, within the provider’s actual policy. No bid. No “maybe later.” No 12-month commit hiding in a footnote. That is the hour I compare across Vast.ai, RunPod, Lambda, CoreWeave, and the hyperscalers.
It is also the hour that survives a checkpoint failure, an eval job, a debug loop, and a serving process that has users. If the workload cannot die, on-demand is not a luxury. It is cheaper than restarting.
Spot is a bid
Spot, interruptible, preemptible, community leftover: different labels, same object. The provider can take the GPU back. Your step is gone. If you checkpoint cleanly and you can wait, the bid can be rational. If you cannot, the discount is a fiction. You did not save 40%. You bought a lottery ticket.
- Training with frequent checkpoints and a queue you can tolerate: spot can win.
- Hyperparameter search with many short runs: spot can win.
- Serving, eval, debugging, un-checkpointable fine-tunes: pay on-demand.
- A “cheap H100” that is spot next to an on-demand H200: not a comparison.
Marketplace hosts add a second failure mode. The bid is not only the cloud reclaiming capacity. The host can vanish. I use Vast.ai when the job can die. I do not use it when the job is the product. RunPod community vs secure is the same fork with a nicer UI.
Reserved and quotes are not on-demand
A reserved cluster can beat on-demand if you will fill it. That is a capacity plan, not a rental. A hyperscaler quote can beat a specialist cloud if you already live in that account. That is procurement, not GPU Cloud News. I still want those numbers labeled. I refuse to rank a 1-year commit against a 2-hour on-demand row.
Lambda’s reserved clusters and CoreWeave’s contracted capacity are real products. They are not the on-demand index. If a blog post leads with a reserved dollar and a screenshot of an H100, it skipped the only honest sentence: you cannot click that today without a contract.
How vendors hide the flag
- One “from” price, three SKUs behind it.
- Spot default, on-demand in a dropdown.
- Waitlist with a green availability badge.
- Per-GPU price on an 8× node next to a 1× PCIe.
- Regional bait: the cheap hour is in a region your data cannot enter.
I read pricing pages in a fixed order: model and VRAM, GPU count and interconnect, region, stock, then billing flag, then the dollar. If the flag is missing, I assume the worst. How we compare hours is the same order the catalog uses.
A simple rule for the bill
Take the on-demand hour times the expected runtime. That is the ceiling you can defend. Spot only wins if (bid × runtime × probability you finish) plus the cost of restarts is lower. People skip the probability. That is how a “cheap” weekend becomes three weekends and a ruined eval.
Serving is even simpler. If a preemption takes you off the internet, you did not save money. You bought an incident.
Where this shows up on real cards
H100 is the card with the most mixed tables. Marketplace on-demand, marketplace spot, managed on-demand, hyperscaler on-demand, reserved. Same badge, five products. Start with how to rent an H100, then look at H100 vs H200 vs B200 only after the billing flag is honest.
Consumer cards make this worse. A spot 4090 looks like a steal next to an on-demand L40S. One is a gaming GPU on a host. The other is a datacenter inference card. Different ECC, different duty cycle, different failure mode. GPU Cloud News that ranks them on TFLOPS is wasting your time.
What GPUBeacon does with this
We index published on-demand GPU rentals across specialist clouds and hyperscalers. Spot is a flag. Waitlists are a flag. Affiliate links are marked. Commissions never change the order. If you want the market without the fog, stay on the catalog and the provider list.
I take a side in the copy because a comparison without a verdict is a brochure. Pay on-demand when the job cannot die. Take spot when you can checkpoint and you can wait. Never let a vendor mix them so the spreadsheet lies to your manager.
FAQ: on-demand vs spot in the GPU cloud
Is spot always cheaper?
The bid is lower. The finished job often is not. Restarts, queue time, and a dead serving process eat the delta. Compare expected complete cost, not the cell in the table.
Is community cloud the same as spot?
Sometimes. Community often means someone else’s hardware with a weaker isolation story. It can still be on-demand billing. Read the policy. Do not assume “community” means interruptible, and do not assume “secure” means in stock.
Should I use AWS Spot for H100?
Only if your scheduler already speaks Spot and you can reclaim. For a one-off fine-tune, specialist on-demand is less theatre. Hyperscaler Spot is a platform feature, not a GPU Cloud News bargain bin.
How do I see on-demand only on GPUBeacon?
Use the on-demand toggle, then open a model page. The from-price on GPU cards is an on-demand sample. Spot rows stay labeled. If a provider hides the flag, we still will not pretend it is the same hour.
The cheaper hour is the one that finishes. Filter on-demand, match the SKU, confirm stock, then rent. That is GPU Cloud News you can actually use.
Some outbound links in these articles may be affiliate links. Rankings on GPUBeacon stay independent.
