What is scale AI?
Scale AI manufactures training data for AI models — labeling images, writing expert examples, and rating model outputs via a large global contractor workforce. Founded in 2016 by Alexandr Wang, it supplied OpenAI and Google. In June 2025 Meta paid $14.3 billion for a 49% non-voting stake, valuing Scale at $29 billion; Wang left to run Meta's superintelligence lab.
Why — the first-principles explanation
Every AI model needs three inputs: compute, algorithms, and data. Compute you can buy with money — Nvidia will sell to anyone. Algorithms leak fast; research papers and departing employees spread them within months. Data is the one input capital alone cannot conjure, because the useful kind isn't lying around on the internet. It has to be made, by people, one item at a time. Scale's business is being the factory that makes it.
The work sounds unglamorous and is strategically enormous. A self-driving car model needs millions of street photos where someone has drawn a box around every pedestrian and labeled it "pedestrian." A language model needs humans ranking which of two responses is better, over and over — that's the raw material for RLHF, the process that turns a raw text predictor into something that follows instructions. As models got smarter, the labelers had to get smarter too: Scale shifted from anyone-with-a-mouse toward paying PhDs and lawyers to write expert-level questions and answers, because you can't teach a model graduate physics with crowdworkers who don't know physics.
Scale's structural position was unusual and turned out to be its vulnerability. It sat in the middle, supplying data to nearly every major lab — OpenAI, Google, and others — which meant it saw the shape of everyone's work. That's a great business right up until a customer decides a neutral supplier holding that much insight is intolerable. When Meta bought in, Google and other customers reportedly moved to pull work away, precisely because the neutral middleman was no longer neutral.
The deal structure is the tell. Meta paid $14.3 billion for 49% and no voting power. Buying 51% would have triggered a lengthy antitrust review; 49% non-voting is a way to get the founder, the talent, and the data pipeline without formally acquiring the company. Alexandr Wang — who dropped out of MIT and founded Scale at 19 — left to lead Meta's superintelligence effort, with his personal stake reportedly worth around $5 billion. Scale stayed nominally independent under new CEO Jason Droege. Whether this is an acquisition or an investment is, in practice, a question about paperwork.
An example that makes it click
Imagine a school where the students are astonishingly fast readers but have never met a human being. They can absorb every book ever written in a weekend. But nobody has told them which answers are actually good — what's polite, what's true, what a real doctor would say.
So the school hires graders. Thousands of them, all over the world. The graders don't teach; they mark. Two answers, side by side: which is better? Circle the pedestrian in this photo. Write out how a radiologist would explain this scan. That pile of markings is what turns a fast reader into something useful, and someone has to physically produce it. Scale AI is the agency that staffs the graders. It supplied graders to nearly every school in town — which was a wonderful business, until the biggest school bought half the agency, and the other schools started wondering who was reading their marks.
Key facts
- Meta's investment closed on June 13, 2025: $14.3 billion for a 49% stake with no voting power, valuing Scale AI at approximately $29 billion.
- Scale AI was founded in 2016 by Alexandr Wang, an MIT dropout, and co-founder Lucy Guo; Wang was 19 at founding.
- Wang stepped down as CEO to join Meta's superintelligence team as part of the deal; his personal stake was reported at roughly $5 billion. Chief strategy officer Jason Droege was promoted to CEO.
- Scale's core business is producing training data — image labeling, expert-written examples, and human preference ratings used in RLHF — via a large global contractor workforce.
- Before the Meta deal, Scale ran data labeling work for customers including OpenAI and Google; several reportedly moved to reduce reliance on Scale afterward because Meta's stake compromised its neutrality.
- The 49% non-voting structure is widely read as a way to secure talent and data access without triggering the antitrust review a majority acquisition would invite.
▶ The 60-second explainer (script)
Scale AI is the company that manufactures training data for artificial intelligence. Here's why that matters more than it sounds. Every AI model needs three things: compute, algorithms, and data. Compute you can buy — Nvidia sells to anyone. Algorithms leak; papers publish and researchers move. Data is the one input money alone can't create, because the useful kind isn't sitting on the internet. It has to be made by people, one piece at a time. That's Scale. Humans drawing boxes around pedestrians so a self-driving car learns what a pedestrian is. Humans ranking which of two chatbot answers is better — that's the raw material for the process that turns a text predictor into an assistant that follows instructions. As models got smarter, Scale started hiring PhDs and lawyers, because you can't teach a model graduate physics using labelers who don't know physics. Founded in 2016 by Alexandr Wang, an MIT dropout, it supplied nearly every major lab — OpenAI, Google, the Pentagon. Then June 2025: Meta paid fourteen point three billion dollars for forty-nine percent, with no voting power, valuing Scale at twenty-nine billion. Look closely at that structure. Forty-nine and non-voting sidesteps the antitrust review a majority buy would trigger. Wang left to run Meta's superintelligence lab. And Scale's great strength — being the neutral supplier to everyone — evaporated the moment one customer owned half of it.
What authoritative sources say
People also ask
Did Meta buy Scale AI?
Not formally. Meta paid $14.3 billion for 49% with no voting power, so Scale remains nominally independent with its own CEO. Functionally Meta got the founder, key talent, and deep data access — which is why many describe it as an acquisition in everything but paperwork.
Why 49% and not a full acquisition?
A majority stake would likely trigger a lengthy antitrust review. A 49% non-voting stake avoids that threshold while still delivering the strategic goods: the founder, the team, and the data pipeline.
What does Scale AI actually do day to day?
It coordinates a large global contractor workforce that labels data — drawing boxes around objects in images, transcribing and annotating text, writing expert question-answer pairs, and ranking model outputs to produce human preference data for RLHF.
Is Scale AI still working with OpenAI and Google?
Its position weakened substantially after the Meta deal. Multiple customers reportedly moved to reduce their reliance on Scale, since a supplier half-owned by a direct competitor is a hard thing to keep sending your training data to.
Who owns Scale AI now?
Meta holds 49% without voting rights, acquired June 2025. The remainder stays with prior investors, employees, and founders including Alexandr Wang, who retains his stake while working at Meta.
The same question, asked other ways
- What does scale AI do?1,900/mo