Pricing
A rate card you can read in one screen.
What does Onrup cost?
Compute is metered per GPU-second above a one-minute floor, from $0.09 per GPU-hour. Training and serving are priced separately. Every job is costed and reserved against a spend limit you set before any compute is leased, so a run that would breach the ceiling never starts.
Compute
You choose a GPU class, not a machine. Capacity is sourced as serverless compute across multiple providers, which is what the rate reflects. Every figure below is the final amount you pay, inclusive.
| GPU class | VRAM | Training / hr | Serving / hr | Typical use |
|---|---|---|---|---|
| RTX 3080 | 12 GB | $0.09 | $0.11 | Fine-tuning models under 2B parameters |
| RTX 4000 Ada | 20 GB | $0.09 | $0.11 | Low-latency serving of small models |
| L4 | 24 GB | $0.17 | $0.20 | Power-efficient inference |
| RTX 3090 | 24 GB | $0.22 | $0.26 | QLoRA on 7-8B models |
| RTX 4090 | 24 GB | $0.38 | $0.43 | LoRA on 4-8B models |
| A40 | 48 GB | $0.42 | $0.48 | LoRA on 12-14B models |
| RTX 6000 Ada | 48 GB | $0.61 | $0.71 | Fast 14B LoRA |
| L40S | 48 GB | $0.78 | $0.90 | Multi-adapter serving on an 8B base |
| A6000 | 48 GB | $0.65 | $0.75 | 14B LoRA |
| A100 40 GB | 40 GB | $1.17 | $1.35 | 12-14B LoRA |
| A100 80 GB | 80 GB | $1.68 | $1.94 | QLoRA on 70B models |
| H100 80 GB | 80 GB | $2.59 | $2.99 | Fastest single-GPU runs |
| H200 | 141 GB | $4.55 | $5.25 | Mixture-of-experts training |
What that compares to
On-demand H100 80GB is the one figure enough vendors publish for a like-for-like comparison. Every number below comes from the vendor’s own pricing page on the date checked, and links to it.
| Provider | H100 80GB, per hour | Versus Onrup | Source |
|---|---|---|---|
| Onrup (serving) | $2.99 | — | This page |
| Modal | $3.95 | 1.3× the rate | 2026-08-06 |
| Together AI | $5.49 | 1.8× the rate | 2026-08-06 |
| Replicate | $5.49 | 1.8× the rate | 2026-08-06 |
| Baseten | $6.50 | 2.2× the rate | 2026-08-06 |
| Fireworks AI | $7 | 2.3× the rate | 2026-08-06 |
Vendors who do not publish a comparable per-GPU-hour rate are excluded rather than estimated. The full set, including why each one is priced the way it is, is on the landscape page.
Plans
Free
$0
no card required
Everything except leasing compute. Upload datasets, get a validation report, browse the catalogue and hold API keys. The point is to let you find out whether your data is usable before you pay anyone anything.
- API keys with per-scope permissions
- Dataset upload and validation reports
- Full model and template catalogue
- Cost forecasts before you commit
- Compute
- Not included
- Dataset storage
- 5 GB
- Concurrent runs
- 0
Pro
Most teamsCompute at cost card
plus a spend limit you set
The working plan. You set a spend limit; every run is authorised against it before compute is leased, so the limit is a gate rather than an alarm. Compute is billed per second above a one-minute floor, at the rates published on this page.
- Training runs on every GPU class
- Always-warm and scale-to-zero endpoints
- The evaluation gate before deploy or publish
- Hugging Face publishing with generated model cards
- Weight download for every finished run
- Per-run and per-endpoint cost meter
- Concurrent runs
- 3
- Concurrent endpoints
- 5
- Dataset storage
- 50 GB
- Spend limit
- You set it
Enterprise
Talk to us
invoiced, with terms
For teams that need higher concurrency, invoiced billing, or a contract before they can start. The product is identical; the paperwork and the ceilings are not.
- Everything in Pro
- Raised concurrency and storage ceilings
- Invoiced billing with payment terms
- Named support contact
- Security review and questionnaire support
- Concurrent runs
- Negotiated
- Concurrent endpoints
- Negotiated
- Dataset storage
- Negotiated
- Spend limit
- Negotiated
What is metered
Described the way you experience it rather than the way it is recorded.
Training compute
per GPU-second, one-minute minimum
Metered from the moment your job starts on the GPU to the moment it stops. Time spent waiting for capacity is not billed, and neither is teardown.
Serving compute
per GPU-second while a replica is live
An endpoint at zero replicas costs nothing. A warm one bills for the time it is warm, whether or not requests arrive — which is the honest way to describe reserved capacity.
Dataset storage
per GB-month above the plan allowance
Datasets, validation reports, checkpoints, weights and evaluation artefacts all count against the same allowance.
Evaluation and generation
per job, against a ceiling you set
Automated dataset generation and evaluation runs carry a hard cost ceiling declared before the job starts. The job stops at the ceiling rather than exceeding it.
The limit is a gate, not an alarm
Before any job leases compute, its maximum cost is estimated and reserved against your available headroom — which is your limit minus settled spend minus reservations already open. If the reservation cannot be made, the job does not start and the API returns a specific error saying so.
This is a different thing from a budget alert. An alert tells you about an overrun once the money is gone. A reservation prevents the overrun. It also means several jobs submitted at the same moment cannot each pass a check that they collectively breach.
Frequently asked questions
How is compute metered?
Per GPU-second, with a one-minute floor. Time spent waiting for capacity is not billed, and neither is teardown — metering starts when your job is on the GPU and stops when it comes off.
Why are training and serving priced differently?
They are different workloads. Training is bursty and bounded — a run starts, finishes and releases capacity. Serving is resident and continuous, holding capacity available whether or not requests arrive. Pricing them identically would mean overcharging for one of them.
What happens when I hit my spend limit?
New work is refused with a specific error code. Work already running continues to its natural end, because killing a run mid-flight would waste everything spent on it so far.
Does the plan change the compute rate?
No. The rate card is identical on every plan. A plan buys headroom and support, not a discount — a discount ladder just means the list price is wrong.
What does an idle endpoint cost?
Nothing, if it is set to scale to zero. An always-warm endpoint bills continuously for the hours it is resident, which is the honest description of reserved capacity.
Last verified 6 August 2026. Competitor rates are checked at source; each row links to the page it came from.
Start with the free tier
A magic link creates your account, your tenant and your first API key. No card until you ask for compute.