estimate · free · no signup
Three questions. A real plan, not a quote.
The model we would deploy, the exact GPU instance, what your own cloud will bill you for it, and the arithmetic against your current seat bill. It moves as you move the slider.
Nothing to submit. Cloud prices are published on-demand list rates as of 2026-08-22; you pay them to your own provider, never to us.
- model
- Qwen3-30B-A3B
- quantisation
- AWQ 4-bit, 32k context
- Host
- 1x L4 24GB · g6.xlarge
- region
- eu-central-1
- expected throughput
- 265 tok/s at 4 concurrent
- your cloud bills you
- $676/month
- setup, one-time
- $3,900
- your bill today
- $1,000/month
Qwen3-30B-A3B Only 3.3B parameters are active per token, so it answers fast on a small GPU and serves a whole company from one machine.
At 50 seats a Deployment costs you $12 more in year 1 than seats do.
year 1, a Deployment $12,012 = $3,900 setup + $676/month of Host × 12
year 1, seats $12,000 = 50 × $20/seat/month × 12
It flips at 51 seats on this cloud and stays flipped to 200.
We would rather say it here than in the invoice: at this size you are buying control, not savings. Your data stops leaving your account, and everyone gets a login instead of the few people on the company plan. That is worth money to some companies and not to others.
If confidentiality matters more to you than cost, reply and tell us what you are bound by.
Throughput is expected, from the model and card specs — not our own benchmark. We replace it with a measured figure from your Host in the Run Log. We size the Host to hold 20.5GB and stay above 22 tok/s per person using it at once.
the crossover
Where a Deployment starts costing less than seats.
One GPU running around the clock costs the same whether 20 people or 200 use it. Seats do not. That is the whole argument, and it has a crossing point we publish rather than hide.
| seats | seat bill | us | diff |
|---|---|---|---|
| 25 | $6,000 | $10,012 | +$4,012 |
| 50 | $12,000 | $12,012 | +$12 |
| 80 | $19,200 | $12,012 | −$7,188 |
| 120 | $28,800 | $15,012 | −$13,788 |
| 200 | $48,000 | $24,264 | −$23,736 |
Year 1, AWS eu-central-1. Seat bill at $20/seat/month.
Year 1 on AWS eu-central-1: setup plus 12 months of the Host you pay your own cloud for, against 12 months of seats at $20/seat/month. The steps are the setup bands at 41 and 101 seats, and the Host moving up a card.
What we decided for you, and why we did not ask.
the model
Picked from the job, not the headcount. More people need more throughput, which is a GPU decision. Swapping in a bigger model would put an 80GB card on your bill for no gain your staff would notice.
the quantisation
4-bit weights with an FP8 KV cache. It is the setting that fits a real context window on a card you can afford, and we will not ship a quantisation we would not use ourselves.
the GPU class
The cheapest card in your region that holds the model and stays above 22 tok/s per person using it at once. Never the biggest card that fits.
the serving stack
vLLM, or SGLang where it measurably wins on your model. You do not choose this, and you do not have to learn it.