Training your own models: what it buys and what it costs

Calling somebody’s API is cheaper on day one, faster to ship, and better at almost everything. That comparison is easy to run and it is the one most teams run. The comparison worth having is the one at year three.

What renting commits you to

A price you do not set. Per-token pricing has moved in both directions and will move again, and a product whose margin depends on it inherits that volatility.

A model you do not keep. Versions are deprecated. Behaviour changes under a stable name. A prompt tuned over months against one version is not portable to the next, and the work of re-tuning is invisible until it is urgent.

A data path you have to explain. Every document sent for processing leaves the company. There are contractual answers to this, and they are answers, but they are somebody else’s promise rather than your architecture.

Renting can still be the right call. This is the list of what you trade away, and it is worth having written down before the decision rather than after.

The three conditions

Training your own is worth it when all three hold, and it is a bad idea when any one fails.

The task is narrow enough to be learned. Five defined operations, not general assistance. See small models on closed tasks.

You have data nobody else has. If your training set is public data, you are reproducing at great cost what someone else already did better. The case rests on proprietary material.

The work outlives the hype cycle. Training is a capability you build over years. If the requirement might disappear in eighteen months, rent.

What it costs

Hardware you own and have to keep busy, which is a different discipline from paying per call. Engineers who understand training rather than prompting. And an evaluation practice, because a model you trained yourself comes with no external benchmark and no vendor claiming it works. You have to know that it does, which means baselines and a held-out set of real documents.

Where we stand

We train our own models in France, on our own machines. Our language models and our world model are work in progress, and nothing has been published so far. That status is on the research page and it is deliberately unflattering, because the alternative is a claim we would have to walk back.

A country that controls neither its software nor its artificial intelligence ends up depending on those who do. So we build them ourselves. That is the reason underneath the arithmetic, and it does not replace the arithmetic.