Cloud vs. on-premise: the model calculation
Cloud AI means renting, on-premise means owning. Move the sliders to see from when your own hardware pays off — and how large the saving is over three years.
= 3,000,000,000 tokens per month
After about 22 months on-premise is cheaper.
Savings over 36 months: €20,400
Illustrative model calculation — not a price statement. All values are freely adjustable; real costs depend on hardware, utilization and rates. On-premise cost is modelled as hardware (one-off) plus running operation, cloud purely usage-based.
›Basis of the default values
The starting values are not Quinta prices but publicly available market anchors — deliberately conservative and overridable with the sliders at any time:
- •Cloud rate €0.60 per 1M tokens — based on the list prices of low-cost API models. OpenAI lists GPT-4o mini at $0.15 (input) and $0.60 (output) per 1M tokens; more capable models cost several times more. openai.com
- •Since 1 June 2026 GitHub Copilot also bills by token: input, output and cached tokens are converted to AI credits at each model's API rate (1 credit = $0.01), base rate $0.80 per 1M input and $3.20 per 1M output tokens — on top of a fixed monthly credit allowance per seat (Business from $19, Enterprise from $39). That puts it close to the usage model this calculator reflects. docs.github.com
- •Hardware (€30,000) and operation (€400 / month) are illustrative assumptions for a small DGX-class inference server; real figures depend on configuration, electricity price and utilization.
Reference prices as of July 2026, stated in US dollars. Provider prices can change at any time.
More on the topic in the article “On-premise vs. cloud AI: the honest cost calculation”.
Let’s work through your real case together.
In the demo we look at your volume, your hardware and your use cases — with solid numbers instead of estimates.