Scale AI, the data labelling and evaluation company, raised $1 billion at a $13.8 billion valuation, according to Reuters. Meta separately committed roughly $14 billion for a stake in the company, a structure that linked the social network's model roadmap to Scale's labelling capacity.
The financing confirms what model teams have said privately for a year: the binding constraint on frontier quality is increasingly the supply of high-quality human-labelled and human-evaluated data, not raw compute.
Why data is the new compute
Public internet text is largely exhausted as a training signal. The next gains come from expert-labelled reasoning data, reinforcement learning from human feedback, and evaluation pipelines that measure whether a model is actually getting better. Each is labour-intensive and hard to automate.
Scale's business is to industrialise that labour, with tens of thousands of contractors across multiple markets producing structured annotations.
The concentration risk
A small number of companies now sit between the largest model labs and the human data those labs depend on. That is a durable position, and a politically sensitive one, as regulators begin to ask who is doing the labelling and under what conditions.