Scale AI

What Scale AI does

Scale AI is a company that produces the data used to train machine-learning systems. Its contribution is less visible than the models themselves: gathering, labeling and evaluating examples so that models have something to learn from, and organizing the human feedback that shapes how a trained model behaves. Scale supplies data to organizations including OpenAI, Meta and Microsoft, and its public-sector work includes programs for the US Department of Defense^[2][1].

Origins

The company was founded in 2016 by Alexandr Wang and Lucy Guo, who had worked together at Quora, and it went through the Y Combinator startup program^[1]. Guo was fired from Scale in 2018^[1]. Wang stayed on as chief executive until 2025^[1].

How the labeling work is organized

Much of the actual labeling is outsourced. Scale does this through subsidiaries: Remotasks, which focuses on computer vision and autonomous vehicles, and Outlier, which focuses on annotating data for large language models (LLMs)^[1]. The two areas call for different kinds of labor. Computer-vision systems for vehicles learn from images in which objects have been marked out, so the job is largely one of drawing and checking boxes and outlines across very large numbers of frames. LLM work is more often textual: writing example answers, comparing two candidate responses, or checking a model's output against a specification. Both are labor-intensive, and both grow with the number of examples a model is trained on.

Financing history

Scale's valuation rose quickly. In August 2019, after Peter Thiel's Founders Fund made a $100 million investment, the company's valuation passed $1 billion^[1]. By July 2021 it had reached $7 billion, after a financing led by Greenoaks, Dragoneer Investment Group and Tiger Global^[1]. In May 2024 Scale closed a $1 billion financing at a $13.8 billion valuation, a mix of primary and secondary money, led by existing investor Accel^[2]. New investors in that round included Cisco Investments, DFJ Growth, Intel Capital, ServiceNow Ventures, AMD Ventures, WCM, Amazon, Elad Gil, Meta and Qualcomm Ventures^[2]. Several names on that list are also buyers of computing or of AI systems; Meta, for one, appears both as a customer of Scale's data and as an investor in the company^[2].

Work on human feedback

Scale says it partnered with OpenAI on the first experiments in reinforcement learning from human feedback (RLHF) on GPT-2, and then scaled those techniques to InstructGPT and beyond^[2]. RLHF is a way of tuning a model with human judgements rather than only with text to imitate: the model produces several candidate answers, people rank them, and those rankings are used to train a reward signal that the model is then adjusted to score well on. The method became a standard step in turning a raw language model into an assistant that follows instructions, and it depends on large amounts of consistent human comparison work — which is what a data company supplies.

The 2025 Meta deal

In June 2025 Meta spent more than $14 billion for a 49% non-voting stake in Scale AI^[1]. Wang left to join Meta, and Jason Droege replaced him as chief executive^[1]. The structure is unusual: a large stake without voting control, alongside the departure of the founder to the investor. It tied the data supplier closely to one of the large model developers it serves, while leaving it formally independent under a new chief executive.