The Secret $29B Company Powering ChatGPT and the US Military
Most people can name five AI chatbots. Almost nobody can name the company that made those chatbots possible. Scale AI doesn't make AI — it makes AI trainable. It's the infrastructure layer that OpenAI, Google, Meta, Apple, and the US Pentagon all pay for, and until a year ago, it served all of them simultaneously without conflict. Then Meta invested $14.3 billion for 49% of the company — and everything got complicated. Here's the full story of what Scale AI is, how it works, and why the most important AI infrastructure company you've never heard of is now at the center of a competitive conflict that's reshaping the whole industry.
Scale AI processes raw data into high-quality labeled training datasets using a hybrid of AI automation and 240,000+ human contractors — making it the essential infrastructure layer for every major AI model.
The phrase "data is the new oil" gets used constantly in AI coverage. Scale AI is the refinery.
Raw data — unstructured images, video, text, 3D sensor feeds — is nearly useless for training AI models. Scale AI takes that raw data, processes it with a combination of automated ML and human annotation, and returns clean, labeled, structured datasets that AI models can actually learn from.
What Scale AI Actually Does
Scale AI's core business is taking data that machines can't learn from and turning it into data that they can. It sounds simple. At the scale required by frontier AI labs, it's an extraordinarily complex logistics and quality control problem.
The Four Core Scale AI Services
- Data Labeling & Annotation: Images, video, text, audio, 3D sensor data (LiDAR for autonomous vehicles). Human contractors annotate raw data — drawing bounding boxes, classifying objects, transcribing speech, identifying sentiment. Scale uses automated pre-labeling ML to reduce human workload, then routes difficult cases to specialists.
- RLHF (Reinforcement Learning from Human Feedback): Human raters compare pairs of AI responses and indicate which is better. This feedback trains the reward model that makes AI assistants like ChatGPT and Gemini helpful, harmless, and honest rather than erratic. Every major RLHF-trained model has used services like Scale's.
- Model Evaluation & Testing: Scale Validate — systematic testing of AI models for accuracy, safety, bias, and reliability before deployment. Increasingly important for enterprise and government clients who need certified AI performance, not just benchmark scores.
- Agentic AI Environments: Scale RL Environments (launched February 2026) — training and evaluating AI agents in simulated tool-use and computer-use workflows. Nearly half of all new data-training projects now involve agentic components.
Alexandr Wang — The Founder Who Became the Most Valuable Hire in AI History
You can't understand Scale AI without understanding Alexandr Wang. Born in Los Alamos, New Mexico to Chinese immigrant physicists, he dropped out of MIT after his freshman year at age 19 to build Scale AI with Carnegie Mellon dropout Lucy Guo.
By age 24, Wang was the world's youngest self-made billionaire. By 28, he was the most aggressively recruited person in Silicon Valley — and Zuckerberg won.
The $14.3 Billion Hire
In June 2025, Meta invested $14.3 billion for a 49% stake in Scale AI. The investment values the company at $29 billion — more than double Scale AI's prior valuation of $14 billion. Investors including Accel received $2.5 billion in payouts. Nvidia, Amazon, and others with earlier stakes also received liquidity.
Wang stepped down as CEO of Scale AI and joined Meta as Chief AI Officer, leading Meta Superintelligence Labs (MSL). He remains a director on Scale AI's board — a detail that created immediate controversy (more on that below).
Jason Droege — former founder of Uber Eats and Scale AI's Chief Strategy Officer — became Scale AI's Interim CEO. Wang's estimated net worth: $3.2–$3.6 billion.
The Conflict Nobody Is Explaining Clearly
Here's what most articles about the Meta deal miss: Scale AI's entire business model depends on being trusted by competing AI labs simultaneously. When you pay Scale AI for data labeling and RLHF services, you're sharing your AI model's training priorities, your data characteristics, and your development direction with Scale's infrastructure.
🚨 The Client Exodus After the Meta Deal
After Meta took a 49% stake and Wang joined Meta's board with ongoing Scale board participation, several major Scale AI clients reduced or paused their engagements:
- Google: Had planned to spend $200 million on Scale AI services. Notably decreased engagement after the Meta deal — the most significant client reduction reported.
- OpenAI: Reduced engagements due to data confidentiality concerns. OpenAI was one of Scale AI's most prominent early clients.
- xAI: Also reduced involvement over competitive conflict concerns.
The fear: with Wang on both Scale AI's board and inside Meta's AI lab, competitors worry that their AI development priorities — embedded in the data they send Scale to label — could indirectly inform Meta's strategic decisions. Scale AI disputes this, citing independent governance. But the perception damage was significant enough to drive structural client diversification away from Scale.
Thunderforge — The Scale AI Story Nobody Is Covering
While the Meta deal dominated headlines, the most strategically significant Scale AI development of 2025 may be in a completely different direction: the US Department of Defense.
💡 Scale AI Won the Pentagon's Most Important AI Contract
In March 2025, Scale AI won the prime contract for Thunderforge — the DoD's flagship AI agent program for military planning and operations. Partners: Anduril Industries and Microsoft. This is not a data labeling contract. Thunderforge is the US military's primary AI agent infrastructure for battlefield decision-making and planning automation.
Scale AI has secured over $300 million in Department of Defense contracts in total, including a separate $500 million Pentagon contract for processing military data and aiding decision-making. The government platform is Scale Donovan. Wang personally testified before US House Armed Services committees and met with heads of state to position Scale as essential national AI infrastructure before his departure to Meta.
The strategic implication: Scale AI is now simultaneously the training data provider for commercial AI labs AND the prime contractor for US military AI agents. These two business lines create different kinds of risk and create an unusual position where a $29B AI infrastructure company is embedded in national security at the same time it's 49% owned by a social media platform.
How Scale AI's Human Infrastructure Actually Works
The most important thing to understand about Scale AI's business is the invisible workforce making it run: 240,000+ contractors globally, primarily in Kenya, the Philippines, and Venezuela, managed through Scale AI's Remotasks and Outlier platforms.
These contractors label images, transcribe audio, rate AI responses, test model outputs, and annotate 3D sensor data. Their work is what makes every RLHF-trained AI model possible — including ChatGPT, Claude, Gemini, and Llama.
💡 The Physical AI Push Nobody Is Watching
While everyone focuses on language model data, Scale AI quietly launched a major physical AI initiative in September 2025: a robotics data program hiring contractors globally to record point-of-view demonstrations for startups training AI-powered robots. The Physical AI Data Engine is integrated into Universal Robots' UR AI Trainer, targeting UR's 100,000+ industrial deployments. It's already logged 100,000+ production hours. A large-scale industrial robotics dataset is planned for late 2026. Scale Labs — Scale AI's research arm (launched March 2026) — is specifically studying advanced AI systems in real-world environments.
Scale AI's New Products in 2026
| Product | What It Does | Launched |
|---|---|---|
| Scale Launch | AI data management platform for enterprise ML pipelines | 2026 |
| Scale Validate | Model testing and evaluation — accuracy, safety, bias | 2026 |
| Scale Rapid | Automated data annotation — faster, lower-cost pipeline | 2026 |
| Scale RL Environments NEW | Train and evaluate AI agents in simulated tool-use and computer-use workflows | Feb 2026 |
| Scale Labs NEW | Research hub: advanced AI systems, agentic evaluation, AI safety oversight | Mar 2026 |
| Scale Donovan | Government/defense AI platform — federal agencies and DoD | Ongoing |
| Physical AI Data Engine | Robotics training data — UR AI Trainer integration, 100K+ production hours | Sep 2025 |
| Outlier NEW | Expert contractor platform — domain specialists for complex annotation | Active |
Scale AI's Journey — From $14B to $29B in One Deal
Frequently Asked Questions
What is Scale AI?
Scale AI is a San Francisco AI data infrastructure company (founded 2016, Alexandr Wang and Lucy Guo) that provides high-quality training data, model evaluation, and RLHF services to the world's major AI labs and US government. Clients include OpenAI, Meta, Microsoft, Nvidia, Toyota, and the US DoD. Valued at $29 billion after Meta's $14.3 billion investment for a 49% stake in June 2025. Revenue: $870M in 2024, projected $2B in 2026. Often called the "picks and shovels" of the AI gold rush — the infrastructure every major AI model depends on.
How does Scale AI work?
Hybrid approach: automated ML pre-labeling combined with 240,000+ human contractors (Kenya, Philippines, Venezuela) on Remotasks and Outlier platforms. Raw data (images, video, text, audio, 3D sensor) is submitted by clients. Scale automates what it can, routes complex cases to human annotators, and returns clean labeled datasets. Also provides RLHF (human feedback that trains AI to be helpful), model evaluation testing, and agentic AI training environments. Nearly half of all new data-training projects now involve agentic components.
What is the Scale AI and Meta deal?
In June 2025, Meta invested $14.3 billion for a 49% stake, valuing Scale AI at $29 billion. Founder Alexandr Wang departed to join Meta as Chief AI Officer, leading Meta Superintelligence Labs (which produced Muse Spark in April 2026). Jason Droege became Interim CEO. The deal created controversy: Google (planned $200M spend), OpenAI, and xAI reduced engagements over data confidentiality concerns. Scale AI remains independent with Meta as minority shareholder.
What is Scale AI's government business?
Scale AI has secured $300M+ in DoD contracts. In March 2025, Scale won the prime contract for Thunderforge — the DoD's flagship AI agent program for military planning and operations (with Anduril and Microsoft). Earlier: $500M Pentagon data processing contract, $250M federal agencies access contract (2022). Government platform: Scale Donovan. Wang personally testified before House Armed Services committees and lobbied internationally for AI investment before his Meta departure.
Who founded Scale AI and what is Alexandr Wang doing now?
Alexandr Wang (MIT dropout at 19, born Los Alamos, NM) and Lucy Guo (CMU dropout) founded Scale AI in 2016. Wang became world's youngest self-made billionaire at 24. In June 2025 he left Scale AI to join Meta as Chief AI Officer at 28, leading Meta Superintelligence Labs. Net worth: $3.2–$3.6 billion. He remains a Scale AI board director. Current Scale AI Interim CEO: Jason Droege (former Uber Eats founder and Scale's former Chief Strategy Officer).
Scale AI is the infrastructure story the AI industry talks about least and depends on most. The Meta deal changed its competitive position significantly — some clients left, but the government business accelerated, the Physical AI platform is growing, and the projected $2B in 2026 revenue suggests the model is still intact despite the conflicts.
The most honest way to think about Scale AI: it's the unsexy layer that makes the sexy AI products possible. Wherever you see a headline about a new frontier model, somewhere in that model's training history, human contractors on Remotasks or a Scale AI pipeline labeled the data that made it work.
