How we calculate

The four steps

• We read your cloud bill. Seat plans become seats times price; usage plans become tokens per month. Everything is net of VAT, because invoices from US providers to EU companies are usually reverse charge.
• We size hardware for your peak load, not your average. Peak concurrent sessions and the model class set how much memory and throughput you need, and we pick the smallest catalogue entry that carries it.
• We cost that hardware over the period: purchase and setup once, electricity and maintenance every month.
• We compare the two curves month by month, and report the month the local curve crosses below the cloud curve. If it never crosses inside 120 months, we say so.

The formulas

• Memory needed = model weights + KV cache per session times peak concurrent sessions. Weights depend on the model class you need, not on your team size.
• Throughput needed = peak concurrent sessions times tokens per second per stream. Chat and coding assistants have different demands, so we size them separately.
• Monthly electricity = rated power times power factor times hours per day times days per month times price per kWh. Running around the clock costs roughly three and a half times what business hours cost.
• Monthly maintenance = a yearly share of the hardware price, divided by twelve. The share rises with setup size, because more machines mean more to keep running.
• Break-even = the first month where cumulative local spend falls below cumulative cloud spend. Cumulative, not monthly: the hardware is paid on day one and has to be earned back.
• Total cost of ownership = purchase + setup + all monthly costs over the period. We show the same period for both sides so the comparison is fair.

Assumptions you can change

These four are on a slider in every report. Move them and the report recalculates in front of you.

• Electricity price: 0.25 €/kWh net by default, adjustable from 0.20 to 0.32. German business tariffs sit in that band.
• Power factor: 70% of rated power by default. Hardware rarely draws its full rating under inference load.
• Maintenance: from 8 to 15 percent of hardware price per year, depending on setup size.
• Cloud price change: 0% per year by default, meaning we assume today's prices hold. You can test a fall or a rise.

Assumptions you cannot change

These are fixed because changing them would make the comparison meaningless rather than more accurate.

• VAT is 19 percent and is excluded from every calculation. You can switch the report to gross for display, which does not change the underlying numbers.
• The default period is 36 months, with 60 months available. Beyond that, hardware replacement starts to dominate and a three year comparison stops being honest.
• Hybrid setups assume 20 percent of your load stays in the cloud, for the cases local hardware cannot serve.

Where the numbers come from

• Hardware prices are list prices from German and EU vendors, including the EU premium over US list. Each catalogue entry carries its own source and date.
• Cloud plan prices come from provider pricing pages. Usage based plans are converted with a blended price per million tokens, derived from published input and output rates.
• Power draw figures are manufacturer ratings, scaled by the power factor above.
• Prices as of 17 Sep 2026. We re-check them when providers publish changes.

What this does not tell you

An honest cost model has edges. Here are ours.

• It says nothing about output quality. An open model on your own hardware is not the same model you use today, and for some work the difference matters more than the cost.
• Throughput figures are derived from published measurements, not from a benchmark we ran on your workload. Treat them as estimates with a margin.
• It does not price your time. Someone has to own the machine, and that cost lands on a person rather than an invoice.
• It ignores resale value and hardware refresh. Both would improve the local case, so the numbers you see are the conservative ones.