Compute & Cost
Local Model
A model running on your own infrastructure: full control over your data, predictable costs, independence from a vendor's whims — in exchange for operating effort and, usually, somewhat lower peak performance. For law firms, clinics and public agencies it is often the only viable route. The quality gap to the big API providers is closing faster than most procurement policies age.
An example
A tax law firm runs a model on its own machine because client data may not leave the building. The calculation looks good at first: 12,000 euros for hardware instead of recurring fees. What the calculation left out was half a position for operations, updating with every new model generation, redundancy in case of failure, and the fact that the card sits idle at night and on weekends.
A closer look
A self-operated model on your own or rented hardware. It brings data sovereignty, fixed costs and independence from vendor changes — and demands operations: updates, monitoring, utilization, on-call readiness.
Why it matters
Self-hosting is the right choice when data protection or availability forces it — and rarely the cheaper one. An honest comparison includes hardware, power, operating effort, updates, resilience and utilization; it comes out in favor of self-hosting under steady, high load and almost never under fluctuating load. But what often decides the comparison is not in the calculation: with the weights in-house, no one can cancel your access or discontinue the model. That independence has value; it is just not a cost advantage.
Commonly misunderstood
The comparison is usually calculated too favorably. On top of hardware come staff, resilience and the distance to the best model currently available.