Meta-Scale AI Deal: $14.3B Stake & Alexandr Wang's Move

What Scale AI sells

Scale AI's business is turning raw data into the labeled examples a model learns from. Its Data Engine annotates text, images, video and lidar sensor data, and for large language models its main product is reinforcement learning from human feedback: paid contractors review a model's candidate answers and rank, correct or rate them, and that graded data becomes the training signal used to steer the model toward the answers people preferred.Scale AI also sells red-teaming, where contractors try to provoke unsafe or wrong outputs, and model evaluation, where a model's answers are scored against a battery of prompts before release. [9]

Atlas interpretation: The reason a frontier lab pays for this rather than doing it in-house is scale of a different kind: the last stage of building a model like GPT-4 or Llama is not more scraped text, it is many thousands of hours of human judgment on the model's own outputs, delivered fast enough to keep a training cycle moving. Building and managing that labor force is a logistics business as much as an AI one, and Scale built it first and biggest. Both OpenAI and Google were paying Scale AI customers before this deal, which is what made a competitor's 49 percent stake in their vendor a problem once it happened. [9][1]

An investment, not an acquisition

Scale AI's own announcement describes this as an investment, not a sale, and is explicit that Scale remains "an independent leader in AI." Reporting on the deal terms puts the number at $14.3 billion for a 49 percent stake in Meta, valuing Scale AI at over $29 billion, with the stake carrying no voting rights. Alexandr Wang left the CEO role but stayed on Scale's board as a director, and Jason Droege, who had joined as chief strategy officer in September 2024, was named interim CEO. The proceeds were structured to give Scale shareholders and vested equity holders liquidity, which is a mechanism companies use when they want cash to reach employees and early investors without a full sale or an IPO. [1][4]

Atlas interpretation: The 49 percent, non-voting shape of the deal is not incidental. Antitrust attorney David Olson told Stanford's Cyber Policy Center the structure gives Meta "a lot of protection if someone comes after them," because a stake that stops short of majority control and carries no votes sits outside the transactions that automatically trigger a merger filing. Senator Elizabeth Warren pushed back publicly that the label does not settle the substance, calling for regulators to examine whether the arrangement squashes competition regardless of what it is called. Both things can be true: the structure was chosen with the review threshold in mind, and that does not by itself answer whether the deal concentrates power over AI training data in a way regulators should still contest. [5]

What Alexandr Wang actually became

Wang's title at Meta is Chief AI Officer. In that role he leads Meta Superintelligence Labs, the unit Meta stood up around his hire, and Meta's own executive bio describes him as shaping company AI strategy "from concept to execution." He was not the only senior outside hire that month: former GitHub chief executive Nat Friedman joined alongside him. Reporting from July 2025, a month after the deal, already had Wang's lab weighing whether to abandon Behemoth, Meta's largest open-weight model, in favor of a closed one, a philosophical reversal for a company that had built its AI identity on publishing model weights. The New York Times reported the discussions as preliminary, with teams pausing new tests on Behemoth while any decision waited on Mark Zuckerberg's sign-off. [8][6]

Scale AI's customers pull back

Within a week, both of Scale AI's largest rival-lab customers moved to cut the relationship. OpenAI confirmed on June 18, 2025 that it was phasing out its work with Scale AI; a company spokesperson attributed this to a pre-existing shift toward more specialized data providers rather than the Meta deal itself, a distinction reporters noted but could not independently verify. Google was reported days earlier, on June 14, to be backing away from roughly $200 million in planned 2025 spending with Scale AI and opening talks with competing data-labeling vendors. [4][3]

Atlas interpretation: Whatever the stated reasons, the mechanism is straightforward: Scale's RLHF workers see a customer's unreleased model outputs and, often, some of its prompts and evaluation criteria. Once Meta owned 49 percent of the company doing that work, OpenAI and Google were routing their most sensitive pre-release material through a vendor partly owned by a direct competitor. Pulling back is the obvious response regardless of whether Meta's stake carried votes. [4][3]

Reported turbulence inside the lab Wang built

By fall 2025, Meta had cut roughly 600 jobs across Meta Superintelligence Labs, its Fundamental AI Research (FAIR) group, and other AI product and infrastructure teams, while sparing the newly formed TBD Lab built around Wang's headline hires. On November 21, 2025, Yann LeCun, Meta's chief AI scientist for a decade and a co-founder of its original AI research division, announced he would leave at year's end to start an independent AI research company, ending a run that predated Wang's arrival by twelve years. [7]

Atlas interpretation: None of this proves the $14.3 billion bet failed. It does complicate the simplest version of the story Meta told in June 2025, that buying into Scale AI and installing its founder would buy Meta certainty on AI talent and strategy. A year that includes a live debate over reversing the company's open-weight strategy, hundreds of layoffs in the very unit built to fix Meta's AI standing, and the departure of its longest-serving AI scientist reads more like an organization still working out what it bought than one that solved its problem in a single deal. [6][7]