The question nobody asks before renting a GPU
Most guides open with a table of graphics cards. Wrong starting point. The first question is how many hours a month your workload actually runs, because that single number decides whether you belong on an hourly marketplace, a managed GPU cloud, or a fixed-monthly dedicated server. Get it wrong and you pay double. Get it very wrong and you pay five times what the person next to you pays for the same silicon.
Every price in this guide comes from provider pricing pages checked in October and November 2026. Marketplace listings on Vast.ai move daily, sometimes hourly, so treat each number as a snapshot rather than a contract.
What actually needs a GPU
Nginx, a WordPress site, a small API, even a Minecraft server: none of this touches a GPU, and a $5 KVM VPS handles it fine. GPU compute earns its price on a short list of jobs:
- LLM inference and fine-tuning. A 7B to 13B model fits on a 24 GB card, 30B class wants 48 GB, and 70B wants 80 GB or aggressive quantization.
- Image and video generation. Stable Diffusion and its relatives run on 24 GB and run faster on 48.
- Training models from scratch. This is H100, H200 and B200 territory, usually eight cards at a time.
- Rendering and simulation. Blender scenes, molecular dynamics, CFD.
Everything else is cheaper on CPU. That sounds obvious. The number of GPUs billed 24/7 to run cron jobs that would fit on a $2 VPS says otherwise.
The 2026 price landscape, card by card
Entry level: RTX 3090, 4090, 5090
The 24 GB RTX 4090 is still the default hobbyist card in late 2026. Vast.ai listings opened around $0.12 per GPU-hour in October, with a median closer to $0.36. RunPod's Community tier sat at $0.34 and its Secure tier at $0.89. Hostinger, a name regular VPS buyers already know, charges $0.38. Theta EdgeCloud undercut the field at $0.26. The newer 32 GB RTX 5090 started near $0.29 on Vast.ai with a $0.44 median. An RTX 3090, still serviceable for 24 GB inference work, traded as low as $0.07 on marketplace listings.
Rule of thumb: if your fine-tune fits in 24 GB, you should rarely pay more than $0.40 per GPU-hour for it in 2026.
Mid range: L40S, RTX 6000 Ada, A100
Once you need 48 GB, prices roughly double. RunPod Community rents an L40S at $0.79 per hour, the Secure tier at $1.09. CoreWeave lists L40S at $2.25 per GPU but only on eight-GPU nodes. The A100 80GB spans a remarkable range: about $0.40 on the cheapest Vast.ai offers (median $0.71), $1.39 on RunPod Community, $1.79 Secure, $2.79 at Lambda, then $3.67 to $4.10 at the hyperscalers.
Training class: H100, H200, B200
The H100 80GB is the card the whole market prices against. Vast.ai from $1.33 with a median near $2.07. RunPod Community at $1.99 PCIe or $2.69 SXM; Secure at $2.89 PCIe and $3.99 SXM. Lambda sells a single H100 PCIe at $3.29. DigitalOcean GPU Droplets land at $4.41. CoreWeave asks $6.16 per GPU on an eight-GPU node. GCP lists on-demand rates up to $11.06 and Azure up to $12.29. Identical silicon, a nine-fold spread between the cheapest and the priciest listing.
The B200 with 180 to 192 GB of memory starts around $4.38 to $5.31 on Vast.ai, runs $6.69 at Lambda, $6.79 to $7.99 at RunPod, and peaks at $16.11 on GCP.
One trend worth flagging: a Chinese brokerage research note counted H100 one-year rental rates up nearly 40% since October 2025, with domestic cloud vendors raising AI compute prices anywhere from 5% to 400%. Cheap H100 hours are getting harder to book, not easier.
Marketplace, managed cloud, or dedicated: pick by utilization
Three tiers, three risk profiles.
- Marketplace (Vast.ai, RunPod Community, Lium). The cheapest hours anywhere. The hardware belongs to third-party hosts, listings disappear without notice, and interruptible offers can be reclaimed mid-job. Good for checkpointed training and batch rendering, bad for anything that must stay up.
- Managed GPU cloud (RunPod Secure, Lambda, TensorDock, Hyperstack). Typically 30 to 50 percent more per hour. In exchange you get owned datacenter capacity, and in Hyperstack's case zero egress fees.
- Fixed monthly (Hetzner, reserved capacity elsewhere). Hetzner rents an RTX 4000 Ada server at 184 euros a month, an L40S at 349, an RTX 6000 Ada at 439. No hourly flexibility, but the math at high utilization is brutal in your favor.
Run the numbers on an always-on L40S. Hetzner fixed monthly: 349 euros, roughly $380. RunPod Secure at $1.20 an hour comes to $876 a month at full 24/7. At 30% utilization, meaning eight hours a day, RunPod drops to about $288 while Hetzner still costs $380. The crossover sits near 35% utilization. Below it, hourly billing wins. Above it, fixed monthly wins by a factor of two to four. No tier wins everywhere.
The billing traps that decide your real bill
Per second, per minute, or per hour
RunPod and Vast.ai bill per second. Lambda bills per minute. A 16-minute experiment costs about 27% of the hourly rate on RunPod and the full hour on a per-hour platform. If your workload is short bursts, billing granularity is worth more than a 10% headline discount.
Idle billing
Most platforms keep charging as long as the instance exists, GPU busy or not. Hostinger is candid about this: destroying the instance is the only way to stop compute charges. The stopped-but-not-deleted pod is the classic end-of-month surprise.
Egress and storage
RunPod advertises no egress fees, Hyperstack lists free egress, and Lambda charges nothing for ingress or egress on its filesystems. The hyperscalers charge, and moving a 300 GB checkpoint out of the wrong cloud can cost more than the compute that produced it.
Prices that move without much warning
Lambda posted new rates effective November 16, 2026, lifting a single H100 SXM to $4.89 an hour. Fly.io shut down its GPU Machines product entirely on August 1, 2026. Locking your architecture to one vendor's spot availability is a plan with an expiry date attached.
A short decision path
- Estimate your monthly GPU hours honestly. Not the optimistic number.
- Under about 35% utilization: rent hourly. Marketplace if the job survives interruption, managed cloud if it does not.
- Over 35%: price a fixed-monthly dedicated server before renting hourly again.
- Match VRAM to model size before anything else. A cheaper card that cannot hold the model is not cheaper.
- Confirm billing granularity, idle policy and egress fees before your first launch, not after.
The gap between a $0.12 marketplace 4090 hour and a $12.29 Azure H100 hour is not quality. It is utilization fit, risk and contract terms. Work out your hours first and the rest of the decision mostly makes itself.