Inference

Splits one AI job across whichever models and chips run each part cheapest.

The routing layer across heterogeneous silicon — the software that decides which model runs on which chip, sitting between the chip makers and the teams buying inference.

Hosted open-model inference at realtime, async and overnight-batch prices.

Open-weight inference sold by latency tier, plus an open-source gateway teams self-host — the layer where a buy-versus-rent decision turns into a per-token bill.

One API key for open models that grows from per-token to reserved hardware.

The clearest published attempt to make inference pricing checkable rather than trusted — they write down the rule that sets each rate and ship calculators that show where per-token billing stops being the cheap option.

Compresses models so the same output costs less GPU time.

Compression sold as a product rather than a feature — the clearest live test of whether "we make your model cheaper to run" is a business you can charge for, or something the serving layer eventually swallows.

Deploys and scales models as APIs on the customer’s own servers or cloud.

UbiOps runs and scales AI models on infrastructure the customer already owns — the orchestration layer a European organisation reaches once it has decided its models are not leaving its own hardware.

One compiled inference server that runs the same model on any vendor’s chip.

The clearest live test of whether inference is really portable — one server, one model, five accelerators, and published numbers for each, which is the only way to price a task instead of a token.