GPUs & Pricing
Every GPU we offer — what it is, what it's for, and what it costs per second.
You pay per second of GPU time, only while your work runs. The pricing page always shows this table live, with a cost simulator.
Rates at a glance
| GPU | VRAM | $ / second | ≈ $ / hour |
|---|---|---|---|
| CPU | — | $0.000042 | $0.15 |
| T4 | 16 GB | $0.000180 | $0.65 |
| L4 | 24 GB | $0.000244 | $0.88 |
| A10G | 24 GB | $0.000337 | $1.21 |
| L40S | 48 GB | $0.000596 | $2.15 |
| A100 | 40 GB | $0.000641 | $2.31 |
| A100-80GB | 80 GB | $0.000763 | $2.75 |
| H100 | 80 GB | $0.001207 | $4.35 |
| RTX PRO 6000 | 96 GB | $0.001263 | $4.55 |
| H200 | 141 GB | $0.001892 | $6.81 |
| B200 | 192 GB | $0.002604 | $9.37 |
| B300 | 288 GB | $0.002958 | $10.65 |
Faster is often cheaper per job: a card with 2× the hourly rate that finishes 3× sooner costs less per generation. Judge by cost-per-result, not cost-per-hour.
The lineup, card by card
Everyday tier — images and editing
CPU — no GPU at all. For opening a workspace to edit workflows, testing node wiring, and running CPU-only nodes. At $0.15/hr it's the cheapest way to build before you spend GPU money.
T4 (16 GB) — the budget classic. Comfortable with SD 1.5 and small models; the right pick for high-volume batch jobs where per-image cost matters more than speed.
L4 (24 GB) — a newer, more efficient small card. Runs SDXL acceptably; good middle ground when T4 is too small and A10G feels like more than you need.
A10G (24 GB) — the image workhorse. Fast enough that SDXL images land in seconds, cheap enough to iterate freely. If you're generating images and not sure what to pick, start here.
Mid tier — big image models, first video steps
L40S (48 GB) — the value sweet spot of the mid tier. Twice the VRAM of A10G for less than 2× the price: Flux-class image models, upscaling passes, ControlNet-heavy pipelines, short/light video work.
A100 (40 GB) — a datacenter classic. Similar capacity territory to L40S; pick whichever is cheaper for your workload (usually L40S).
A100-80GB (80 GB) — doubles the memory for pipelines that stack several models at once (base + refiner + ControlNets + LoRAs) or larger video models on a budget. The most affordable way into the 80 GB class.
High-performance tier — heavy pipelines, quantized models
H100 (80 GB) — much faster than the A100-80GB at the same memory size. When your 80 GB workload is compute-bound and iteration speed matters, the upgrade pays for itself in wall-clock.
RTX PRO 6000 (96 GB) — a Blackwell-generation workstation card. Its superpower is native support for modern quantized formats (FP4/INT8): models shipped as quantized weights run their fast path here rather than emulating it. Excellent price/performance for quantized image and video models that fit in 96 GB.
Video flagship tier — where modern video models live
Modern video generation (multi-second clips, native audio, high resolution) is the most demanding workload on the platform: the model weights are large AND the video itself occupies enormous working memory during sampling. This tier exists for it.
H200 (141 GB) — the video default. Huge memory plus very high bandwidth means current video models fit comfortably and sample fast. If a workflow says "video" and you don't have a reason to pick something else, pick H200.
B200 (192 GB) — flagship Blackwell. More bandwidth than H200 and native quantized-format support — the fastest option we offer for long clips and high resolutions, and often the cheapest per video on big jobs despite the higher rate, because it finishes so much sooner.
B300 (288 GB) — the ceiling. Blackwell Ultra with 288 GB for the workloads nothing else holds: maximum duration, maximum resolution, full-precision weights of the largest video models. If a job out-grows the B200, this is the answer; otherwise it's more card than you need.
Choosing in 3 questions
- Does the model fit? Weights + working memory must fit in VRAM. Out-of-memory error = move up a tier.
- Image or video? Images: A10G → L40S → H100 as models grow. Video: H200 by default; B200/B300 for long or high-res jobs.
- Quantized weights? (filenames with
fp4,int8,nvfp4,awq…) — Blackwell cards (RTX PRO 6000, B200, B300) run these natively and are usually the best price/performance for them.
What jobs actually cost
| Job | Sensible GPU | Typical cost |
|---|---|---|
| One SDXL image (~30 s) | A10G | |
| Flux image (~20 s) | L40S | |
| Video upscale pass (~3 min) | L40S | |
| 5 s video clip (~5 min) | H200 | |
| 15 s video @ 1K (~25 min) | B200 |
Your exact numbers depend on workflow, steps, and resolution. The ⏳ In progress meter on the Usage page shows real cost accruing live, and after one run you'll know that workflow's price almost exactly.
Good to know
- One run = one GPU. Raising Max Parallel GPU runs more jobs simultaneously (each on its own card) — it never speeds up or enlarges a single job.
- Long video jobs: set the machine's Workflow timeout high enough for the whole job before running (options up to 2 hours) — see the Billing FAQ.
- Changing GPU: Machine → Settings → GPU → save (auto-rebuilds). Sessions pick a GPU at session start.