Home/Pricing
Hosts set the marketplace price and competition pushes it down; we add a 15% commission on top and we do not hide it. Compute, storage and traffic are separate rows — on the invoice and on this page.
You pay by the second and stop whenever you want. No commitment.
1 · GPU
The rate is shown per hour, the metering is per second: a job that runs 9 hours 14 minutes is not rounded up to 10 hours. The figures below are example starting prices taken from today's pool and are illustrative; the live ones appear in the console.
| GPU | VRAM | Typical use | Starting price |
|---|---|---|---|
| RTX 3090 | 24 GB | Stable Diffusion, small LLMs | $0.21/hr |
| RTX 4090 | 24 GB | Fine-tuning, image generation | $0.42/hr |
| L40S | 48 GB | Inference, video | $0.79/hr |
| RTX A6000 | 48 GB | Long context, render | $0.88/hr |
| A100 PCIe | 80 GB | Training, 70B inference | $1.32/hr |
| H100 SXM | 80 GB | Serious training runs | $2.19/hr |
Example, illustrative prices; on the marketplace they move second by second with supply and demand. The price of the same card differs by region, by the host's scorecard score and by current utilization — in the console, every offer carries the reason its score is what it is.
2 · Storage and network
On some invoices storage and traffic are folded into compute; we keep them apart. A stopped pod's disk keeps accruing — we write that sentence inside the table, not in small print.
| Item | How it is measured | When it accrues | Example rate |
|---|---|---|---|
| Persistent volume (NVMe) | GB · month | For as long as the volume exists, including while the pod is stopped | $0.10/GB-month |
| Pod container disk | included | While the pod runs; the data goes when the pod is deleted | $0.00 |
| Network traffic — ingress | GB | While uploading data and images | $0.00/GB |
| Network traffic — egress | GB | While downloading output, while an endpoint responds | $0.01/GB |
The storage and traffic rates are illustrative and will be fixed in the console when the marketplace opens; we are not publishing a settled list today. What is settled is the shape: these two items are not folded into compute, they stand as separate rows on the invoice and they count against the budget cap.
Compute stops, storage continues. In the console that row is shown separately and counts against the cap; the question "why is money still going out" never comes up.
Spend updates every second; the burn rate ($/hr) and the end-of-month projection sit on the same card. An invoice surprise is not a category.
The cap counts storage and traffic too, not just compute. When it fills, pods are paused — not deleted.
3 · Commission
Instead of hiding that we are an intermediary, we chose to state our share: a 15% commission goes on top of the price the host set, and the remaining 85% goes to the host; host payouts will be weekly — the payout infrastructure comes online in Phase 1. The industry is known to sit in the 20–25% band; because our share is written on this page, it is a number you can argue with.
| Example: an RTX 4090 listed at $0.42/hr | Amount | To whom |
|---|---|---|
| What you pay | $0.42/hr | — |
| Kaldera commission (15%) | $0.063/hr | Kaldera |
| What the host receives (85%) | $0.357/hr | Host |
An example, illustrative breakdown. In the broker phase some cards come from integrated provider pools; there the cost structure is different, and the offer detail shows whose card it is. Whose GPU spins underneath is our job to handle, but it is not information kept from you.
4 · Simple cost math
The three calculations below are constructed and illustrative: they do not belong to a real customer, they are examples built by hand. The point is not a calculator but showing how the invoice is assembled.
A · An overnight fine-tune
| Item | Calculation | Example amount |
|---|---|---|
| Compute | 1 × RTX 4090 · 9 hr · $0.42/hr | $3.78 |
| Persistent volume | 100 GB · 1 day · $0.10/GB-month | $0.33 |
| Egress traffic | 5 GB · $0.01/GB | $0.05 |
| Total (example) | one night | $4.16 |
B · A week of training
| Item | Calculation | Example amount |
|---|---|---|
| Compute | 4 × A100 PCIe · 168 hr · $1.32/hr | $887.04 |
| Persistent volume | 500 GB · 7 days · $0.10/GB-month | $11.67 |
| Egress traffic | 50 GB · $0.01/GB | $0.50 |
| Total (example) | seven days without interruption | $899.21 |
C · An always-on inference endpoint
| Item | Calculation | Example amount |
|---|---|---|
| Compute | 1 × L40S · 730 hr (30 days) · $0.79/hr | $576.70 |
| Persistent volume | 200 GB · 1 month · $0.10/GB-month | $20.00 |
| Egress traffic | 300 GB · $0.01/GB | $3.00 |
| Total (example) | one month without interruption | $599.70 |
All three calculations are examples and illustrative; a real invoice is formed from the current price of the offer you picked, the number of seconds you actually ran and your actual traffic. If you want to compare the cost of an always-on endpoint against a model that scales to zero: Inference.
5 · The seconds we don't bill
This is not a gesture or a goodwill policy; it is a gate that sits inside the metering path. We write out how it works step by step so it can be argued with.
GPU visibility, driver response, container health and network reachability are checked. The result of every measurement is recorded.
Not as a discount row — they are absent as a line item: seconds marked unhealthy are not added to the compute total. Our billing invariant is this: an unhealthy second enters no line item at all.
If a node goes down, the job can be placed on another host; because the persistent volume stays the same, it resumes from the last checkpoint. The migration event leaves a record you can see — it is not done quietly.
If an unhealthy period is found after the invoice was issued, the amount is reflected as a refund item in the next period. In the console the "refunded last month" total is the sum of those rows, and clicking it opens which pod and which minutes.
The guarantee covers the health of the infrastructure. If your own code OOMs, your script errors out or your model produces a wrong result, the GPU was working and the charge applies. Hiding that distinction opens the guarantee to argument; writing it down closes it.
In the console, the intervals where the health check dropped are marked on the metrics chart: unbilled time is a visible band, not a footnote.
Where things actually stand
This site is a sample marketing page and the company is still being incorporated: every figure on the page is illustrative, not a settled rate card. We do not state customer counts, revenue or an uptime history — because there isn't one yet.
6 · Frequently Asked
Both. The rate is shown hourly so it stays readable; the metering is done per second. In the accounting layer the unit is micro-dollars ($0.42/hr works out to roughly 116.67 micro-dollars per second), because per-second billing does not fit into cents. Rounding happens only on the final invoice total.
No. No minimum rental period, no setup fee, no monthly commitment. Shut the pod down two minutes later and you pay for two minutes. There is also no committed-use discount and no reserved capacity rate — we did not plan them, and because they don't exist we don't list them.
It does not. The budget worker projects every 10 seconds; it sends a notice at 80% and 95%, and pauses pods at 100%. A paused pod is not deleted — its disk and persistent volume are kept. Raise the limit and continue from the second you left off. The cap is a brake, not a guillotine.
A persistent volume stays until you delete it, and the storage item keeps accruing. The free retention period for a paused pod's ephemeral disk is not settled yet — and we do not write an unsettled duration here. When it is settled we will announce it in the terms of service and in the changelog.
The period closes at the end of the month and a single invoice is produced, downloadable as a PDF. Line items are split into compute, storage and network traffic, and each row can be expanded per pod and exported as CSV. You can set up an automatic top-up when the balance drops below a threshold — it is not mandatory.
Because it is a marketplace: supply sets the price. Region, the host's scorecard score, network speed and current utilization all move it. In the console the default sort is not cheapest but best-scoring — price, reliability and network are weighed together, and each row shows why it sits where it does.
85% of the price you list stays with you and our commission is 15%; payouts will be weekly, with the payout infrastructure coming online in Phase 1. Your earnings depend on utilization and utilization depends on the scorecard score: a good host gets more work. Detail and an example earnings calculation: Become a Host.
Get started
You don't pay for seconds that don't work, a hard budget cap, per-second billing. Three promises, three mechanisms — all of them written on this page.
The console runs on a local development address: localhost:3000