How Does Scale AI Make Money: Data Labeling, RLHF, and the Meta $14.3B Stake

Scale AI made $870 million in 2024 revenue and won over $300 million in Department of Defense contracts before Meta paid $14.3 billion for a 49% stake in June 2025. That's the short answer. The longer answer is more interesting.

Scale AI built its business selling the unglamorous input every AI model needs: labeled, ranked, and human-verified data. Founded in 2016 by Alexandr Wang, the company grew from processing images for self-driving car startups into the default vendor for reinforcement learning from human feedback (RLHF) at OpenAI, Google, and Meta, plus a growing book of Pentagon contracts. Then, in one of the largest strategic investments in AI history, Meta bought 49% of the company for $14.3 billion — and in doing so, arguably broke the business model that made Scale AI worth buying into in the first place.

How Does Scale AI Make Money?

Scale AI makes money by charging AI labs and government agencies for turning raw data into the labeled, ranked, and human-verified training material that large language models are built on. Commercial clients like OpenAI, Google, and Microsoft pay per-project or contract fees for data annotation and RLHF work — human evaluators scoring model outputs, writing reasoning traces, and building the reward signals that make post-training work — processed through a hybrid of automated tooling and a global contractor workforce of over 100,000 people.

Layered on top of that core annotation business is a smaller applications and SaaS segment, generating an estimated $200-300 million of Scale's roughly $2 billion in 2025 revenue, plus a Department of Defense book worth more than $300 million across multiple contracts. Roughly 70% of revenue comes from U.S. customers. If you're tracking how the broader private AI market is being priced, see our AI valuations dashboard.

Pricing runs on two structures depending on the client. Smaller or one-off engagements pay per-annotation rates that scale with task complexity — a simple bounding-box image label costs far less than a multi-step chain-of-thought reasoning trace written by a licensed professional. Larger frontier-lab accounts instead sign annual or multi-year project-based contracts with committed volume, which is what lets Scale plan capacity against a workforce of over 100,000 global contractors rather than staffing project-by-project.

Scale AI's Revenue Mix: Data Labeling, Government, and Applications

The core of Scale AI's business is still data labeling and RLHF, an estimated 60% of revenue, where clients pay per-annotation or project-based fees for everything from image and video tagging to chain-of-thought supervision for reasoning models. Government and defense work makes up roughly another 20%, anchored by the Thunderforge program — the Pentagon's flagship AI agent initiative for military planning, which Scale AI won as prime contractor in March 2025 alongside Anduril and Microsoft.

The remaining share comes from Scale's Generative AI Platform, a developer-facing SaaS and API layer that lets enterprises build and evaluate their own AI applications without hiring a labeling team directly. CEO Jason Droege has described the underlying data business as "very, very large" relative to the applications segment, which suggests Scale is still, at its core, a data-labor company wrapped in enterprise software branding.

Scale AI's $300M+ Pentagon Contract Book

Government work has become one of Scale AI's most durable revenue streams precisely because it isn't exposed to the same competitive dynamics as its frontier-lab business. The relationship dates to January 2022, when Scale won a $250 million contract giving federal agencies access to its annotation and evaluation tools. Since then, the company has added a $99 million Army R&D contract, a five-year, $100 million ceiling agreement with the DoD's Chief Digital and Artificial Intelligence Office (CDAO), and a $500 million Pentagon contract to process data supporting military decision-making.

The most significant recent win is Thunderforge, the DoD's flagship AI agent program for military planning and operations, which Scale AI captured as prime contractor in March 2025 in partnership with Anduril and Microsoft. Unlike the commercial frontier-lab relationships that unraveled after Meta's investment, government contracts run on multi-year terms and procurement cycles that are far slower to unwind — a structural hedge Scale is now leaning on more heavily than it may have expected a year ago.

Why Meta Paid $14.3 Billion for Scale AI

On June 13, 2025, Meta closed a $14.3 billion investment for a 49% non-voting stake in Scale AI, implying a total company valuation of roughly $29 billion — one of the largest strategic AI transactions ever completed. The deal brought Scale founder Alexandr Wang into Meta to lead a new superintelligence research lab, while Jason Droege stepped in as CEO of an operationally independent Scale AI that retained its board seat structure and existing customer contracts, at least on paper.

Meta's rationale was straightforward: access to the specialized, expert-vetted datasets required to train competitive large language models, at a moment when pretraining on scraped web text had hit diminishing returns and RLHF-driven post-training became the new frontier for model improvement. But the deal solved Meta's data problem by creating a much bigger one for Scale — turning its most valuable asset, neutrality among competing AI labs, into a liability almost overnight.

Structurally, the deal was built to look like a minority investment rather than an acquisition — Meta holds no board votes and Scale continues operating with its own leadership, a design choice that likely reflected antitrust sensitivities around a straight buyout of the market's dominant data vendor. In practice, the non-voting structure didn't matter much to Scale's frontier-lab customers, who judged the arrangement on the economics of ownership, not on governance paperwork.

Scale AI vs Mercor vs Surge AI: The Post-Meta Market Split

Within weeks of Meta's investment closing, Google, OpenAI, and xAI all cut or significantly reduced their contracts with Scale AI, unwilling to route their proprietary training data and evaluation pipelines through a vendor now half-owned by a direct competitor. That business didn't disappear — it moved to rivals Mercor and Surge AI, both of which have since posted faster revenue growth than Scale, despite Scale still carrying the higher headline valuation from the Meta transaction.

Figures blended from TechCrunch, Forbes, CNBC, and Sacra's competitive analysis of the expert-data labeling market, 2025-2026. Private-company revenue and headcount figures are estimates and may not be independently audited.

The Bull and Bear Case for Scale AI's Business Model

The bull case is that Scale AI is still the most diversified player in the expert-data market, with a $300 million-plus government book that's structurally insulated from the commercial-lab churn that hit its 2026 numbers, plus deep customer relationships built over nine years that Mercor and Surge AI, both founded within the last six years, haven't had time to replicate. Scale also retains its brand as the category-defining player, which still matters for winning new enterprise and government business even after the Meta deal's fallout.

The bear case is that Meta's $14.3 billion check bought Scale a permanent conflict-of-interest discount: OpenAI, Google, and xAI represented a meaningful share of the frontier-lab revenue that pushed Scale to an $870 million 2024 run rate, and that business is largely gone for good. A $29 billion valuation set at the moment of maximum optimism, against 2026 revenue guidance of roughly $1 billion, implies a price-to-sales multiple north of 25x on a business now growing slower than two of its closest competitors. Track how the broader AI infrastructure spending picture is evolving on our Big Tech Earnings Tracker.