Economic scoping and sizing
The pilot cost three thousand euros. Scaling it costs two hundred thousand a year, and nobody did the maths.
The problem
The pilot cost three thousand euros. Scaling it costs two hundred thousand a year and nobody did the maths. Generic model leaderboards tell you nothing about your own bill.
How we proceed
- 01
Measuring cost per useful token on your real use cases, not on generic leaderboards. A model that is excellent on average can be mediocre on your data.
- 02
Comparing several model mixes at equivalent performance, separating what genuinely requires a large model from what never did.
- 03
Hardware sizing: memory, bandwidth and storage, which are the real bottleneck far more often than raw compute.
- 04
Buy, self-host or hand over as a managed service, each scenario with its total cost of ownership over three years.
Deliverables
- Costed economic model
- Recommended model mix per use
- Infrastructure sizing
- Three-year comparison of the scenarios
Indicative duration
Two to three weeks.
Who this is for
Finance departments, CIOs, procurement departments.