ML Becoming an API Product

This points to ML becoming an API level product, not a specialist function inside every software team. As pre trained models cut the data needed for a useful classifier from thousands of labels to roughly 100, the hard part stops being model science and becomes packaging, evaluation, and deployment into a simple workflow a product manager or developer can run in an afternoon. That is the opening for Nyckel, and for Scale at the high end, to become the Twilio layer for narrow ML tasks.

• The key economic shift is lower setup cost. Nyckel describes customers uploading their own examples, labeling about 100 samples in roughly 20 minutes, then training and deploying in seconds. That replaces the old workflow of hiring labelers, writing instructions, stitching together MLOps tools, and maintaining inference infrastructure.
  1 sacra 2 sacra
• This does not mean all ML collapses into a simple API. Scale grew by serving data hungry, high risk workloads like autonomous driving and defense, where huge visual datasets, audit trails, and edge cases still justify a full stack and human labeling operation. The self serve API model fits narrower text and image tasks first.
  2 sacra 4 sacra
• The market is splitting in two. Nyckel is built for startup CTOs and product managers who want custom classification without becoming an ML team. DataRobot and similar platforms sell broader automation, governance, and deployment to enterprises. Scale sits between those worlds, moving from labeling into an end to end stack as revenue scaled from $215M ARR in 2022 to $760M in 2023 and $1.5B in 2024.
  1 sacra 3 sacra 4 sacra

Over the next few years, the winners will be the companies that hide ML behind the cleanest data in, prediction out workflow and then widen from classification into adjacent functions like search, detection, and extraction. As that happens, more product teams will buy ML the way they buy payments or messaging today, as a usage based API instead of a headcount heavy internal capability.