YesPress · San Francisco · Founded 2016
The Company That Feeds Every AI Model Worth Knowing
$29 billion. 240,000 workers. One fridge camera that never caught the thief.
YesPress · San Francisco · Founded 2016
$29 billion. 240,000 workers. One fridge camera that never caught the thief.
The Origin Story
In 2016, a 19-year-old MIT dropout named Alexandr Wang tried to build an AI-powered camera to catch a food-stealing roommate in the act. The roommate escaped justice. But the experience of drowning in unlabeled video frames planted the seed for what would become Scale AI - the company that quietly became indispensable to every major AI lab on the planet.
Wang and co-founder Lucy Guo started Scale AI in Y Combinator's Winter 2016 batch - the same year AlphaGo defeated world Go champion Lee Sedol and the modern deep learning era began in earnest. Their insight was simple and correct: building great AI requires enormous quantities of clean, labeled data. And nobody wanted to do the labeling.
"The fridge camera never solved the crime. But it built a company worth $29 billion."
- The origin of Scale AI's founding thesisWang comes from a family of physicists - both his parents worked at Los Alamos National Laboratory on nuclear weapons research. That background, with its emphasis on rigor, measurement, and getting things precisely right, shows up in Scale's obsession with data quality and evaluation. When you are training models for the US military or for OpenAI's latest flagship, imprecision has consequences.
Lucy Guo, a designer and former Quora colleague of Wang's, departed the company in its early years. Wang ran it solo, scaling from a handful of contractors to a network of 240,000+ gig workers and a team of roughly 1,200 full-time staff. By 2021, at age 24, he had become the world's youngest self-made billionaire.
FAST FACTS
◆ Wang's parents: nuclear physicists, Los Alamos National Lab
◆ Dropped out of MIT at 19
◆ Y Combinator Winter 2016
◆ Same year as AlphaGo's famous win
◆ 19-year-old with a stolen sandwich and a vision
◆ Wang net worth ~$3.6B (Forbes, April 2025)
"The picks-and-shovels play of the AI gold rush."
Every miner needs a pickaxe. Scale sells the pickaxes.
What They Build
Scale AI is not one product. It is a stack of tools and platforms built around a core belief - that AI is only as good as the data it learns from, and the evaluations used to measure it. Whether you are a startup training a language model or the US Department of Defense deploying AI for military logistics, Scale has something for you.
The backbone. Dataset curation, labeling, and synthesis powering supervised fine-tuning and RLHF. This is what most of the world's major AI models were trained on.
Full-stack generative AI tooling for enterprise: data prep, prompt orchestration, evaluation, and safety checks. The on-ramp for companies not yet ready to build from scratch.
Agentic AI for mission-critical workflows. Built specifically for government and defense, Donovan helps operators make high-stakes decisions faster - and with better information.
Safety, Evaluations and Alignment Lab. Expert-driven leaderboards and benchmarks for frontier AI models. If you want to know which AI is actually smart, you check SEAL.
The human engine. 240,000+ gig workers globally handling computer vision, autonomous vehicle data labeling, and more. Founded 2017. The invisible workforce behind visible AI.
Fine-tuning, but for LLMs. Outlier recruits experts - not just crowd workers - for high-quality annotation of language model training data. The difference shows up in model quality.
Red teaming and adversarial evaluation. Scale attacks models to find weaknesses before bad actors do. Part research, part security, part stress test for the AI you are about to deploy.
Launched March 2026. Expanded research division covering model capabilities, post-training evaluation, enterprise deployment, and risk oversight. Built on what SEAL started.
Advertisement
The Money
Scale AI's funding history reads like a compressed history of AI hype cycles - except that each round was justified by real revenue growth. The Series G was the story of 2025: Meta wrote a $14.3 billion check for a 49% non-voting stake, implying a $29B enterprise value and cementing Scale as one of the most valuable private AI companies anywhere.
The catch: the moment Meta became a major shareholder, OpenAI, Google, and xAI quietly walked away as customers. When your main investor is also a competitor of your biggest clients, partnerships get complicated. Fast.
"Meta bought 49% of Scale AI. OpenAI, Google, and xAI promptly left. Silicon Valley is nothing if not consistent."
The Roster
Most data companies serve startups. Scale AI serves governments, militaries, and the companies building the models that will define the next decade. That breadth is unusual. The Pentagon trusts Scale to evaluate AI for battlefield use. OpenAI trusted Scale to fine-tune GPT-3.5. Qatar's government signed a five-year deal. General Motors used it for autonomous vehicle data. The customer list is a who's who of anyone serious about AI.
KEY PARTNERS & CLIENTS
The US DoD relationship is significant. February 2024 saw Scale selected to test and evaluate large language models for military use. Then came Thunderforge (March 2025) - a multimillion-dollar deal building AI agents to plan and help execute movements of ships, planes, and military assets. This is not typical enterprise SaaS.
THUNDERFORGE PROJECT
Announced March 2025. Scale AI builds AI agents to plan and coordinate movements of military ships, planes, and assets. The US Department of Defense does not hand out contracts like this to companies it does not trust.
Data labeling to defense contractor. The arc of Scale AI is something.
The Record
Scale AI moves fast and accumulates firsts. Here is what the scoreboard looks like.
The Benchmark
In January 2025, Scale AI's SEAL lab and the Center for AI Safety released "Humanity's Last Exam" - described as the hardest publicly available benchmark for AI models. It contains roughly 2,500 expert-level questions contributed by over 1,000 researchers from more than 500 institutions across 50 countries. The questions span mathematics, medicine, law, linguistics, biology, physics, and more.
The result: the world's best frontier AI models score below 10%. This is not a marketing exercise. It is a genuine attempt to find the ceiling of what current AI can do - and to measure how close we actually are to it. The answer, apparently, is: not as close as the press releases suggest.
"2,500 questions written by the world's experts. The world's best AI models score under 10%. Consider that."
2,500+
Expert-level questions
500+
Contributing institutions
<10%
Top model accuracy rate
Latest Updates
MARCH 2026
Scale Labs launches - an expanded research division built on SEAL, covering model capabilities, post-training evaluation, enterprise deployment, and risk oversight.
JUNE 2025
The Meta deal closes. $14.3B for 49%. Alexandr Wang departs as CEO to lead Meta Superintelligence Labs. Jason Droege - the person who built Uber Eats to $20B revenue - becomes CEO.
JUNE 2025 · FALLOUT
OpenAI, Google, and xAI cut ties. Data-sharing concerns with a Meta-controlled vendor drove the departures. Three of Scale's biggest customers left within weeks of the deal.
MARCH 2025
Thunderforge announced. A multimillion-dollar DoD contract to develop AI agents for planning and executing movements of military ships, planes, and assets.
FEBRUARY 2025
Qatar partnership signed. A five-year deal to improve Qatar's government services using AI-based tools and training programs.
JANUARY 2025
"Humanity's Last Exam" released with Center for AI Safety. 2,500 expert-level questions from 50 countries. Current frontier models score below 10%.
The Scrapbook
The fridge camera. Wang tried to build an AI system to catch a food-stealing roommate. He never caught them. But the experience of handling unlabeled video data became the founding thesis for Scale AI. The roommate remains at large.
Nuclear heritage. Wang's parents are both physicists who worked at Los Alamos National Laboratory on nuclear weapons programs. The founder of the world's largest AI data company comes from a family that built the bomb.
Same year as AlphaGo. Scale AI was founded in 2016 - the exact year AlphaGo beat Lee Sedol and announced that the modern AI era had arrived. Wang timed his entry well, or was prescient, or both.
The world's largest invisible workforce. Most people have never heard of Remotasks or Outlier. But 240,000+ workers on those platforms have contributed to training the AI models that millions of people use every day.
Youngest self-made billionaire. At 24, in 2021, Wang held that title. His estimated net worth was ~$3.6 billion as of April 2025, per Forbes. He started the journey by trying to catch a thieving roommate.
The $250M federal check. In January 2022, Scale secured one of the largest US federal AI contracts issued at the time. The government was paying attention before most of the industry was.
How To Work With Scale AI
If you are building an AI model and need high-quality labeled training data, Scale Data Engine is where you start. If you are a large enterprise looking to fine-tune a model for your specific use case, the GenAI Platform is built for that. If you are a government agency deploying AI for critical decisions, Scale Donovan is the product designed specifically for operational workflows where mistakes are expensive.
For AI researchers and teams evaluating model quality, SEAL's leaderboards offer expert-driven benchmarks that cut through marketing claims. "Humanity's Last Exam" is publicly accessible and remains the hardest standardized test for frontier models.
"If you are building serious AI, at some point you will likely need what Scale sells. The question is only when."
For developers and data teams, Outlier recruits domain experts for annotation work - a step above general crowd labeling. If you have budget and need precision, this is the route. If you are looking to contribute as a worker, Remotasks and Outlier both offer gig opportunities in AI data work.
Scale AI is enterprise-first and not designed for individual developers without institutional backing. Pricing is not publicly listed - you contact the sales team. Given the client list (OpenAI, DoD, Meta, Google), the minimum engagement threshold is likely significant. This is infrastructure for organizations building AI at scale - the name is not an accident.
Links & Sources