Since June, Meta has embarked on one of the boldest and most expensive moves in its artificial intelligence strategy by committing a staggering $14.3 billion investment into Scale AI, a company best known for supplying meticulously annotated data to train advanced machine learning systems. As part of this monumental deal, Alexandr Wang—the young founder and chief executive of Scale AI—along with several key executives from his leadership team, were brought into Meta to take charge of Meta Superintelligence Labs (MSL), the ambitious internal unit envisioned to spearhead the company’s pursuit of artificial general intelligence. Despite the magnitude of this alliance and the promises it carried, the partnership between Meta and Scale AI has already begun to show visible fractures only a few short months later.
An early sign of these tensions emerged when Ruben Mayer, who previously held the role of Senior Vice President of Generative AI Product and Operations at Scale AI and was considered one of the critical recruits brought in by Wang to support MSL, abruptly departed Meta after barely two months at the company. According to individuals with direct knowledge of the matter, Mayer had been entrusted with overseeing the teams responsible for AI data operations within MSL, and he reported directly to Wang. However, he was notably absent from Meta’s most important and strategically central initiative, TBD Labs—a specialized group charged with constructing Meta’s superintelligence projects and attracting elite researchers, many of whom had defected from OpenAI. Mayer’s exclusion from this core initiative appeared to encapsulate a larger concern: that not all of Scale’s leadership imports were finding a place in Meta’s highest‑priority efforts. Despite requests for comment by TechCrunch, Mayer remained silent on the reasons behind his swift exit.
At the same time, Meta’s internal AI division has not confined itself to Scale AI’s services. Multiple sources revealed that TBD Labs, the company’s primary superintelligence research wing, has been collaborating actively with several outside vendors, including Mercor and Surge—two of Scale AI’s most prominent competitors in the data annotation marketplace. Although it is common practice for most industry players to maintain relationships with a variety of data suppliers to create redundancy and ensure flexibility, Meta’s situation stands out because the company has invested billions into one specific provider yet still relies heavily on its rivals. This contradiction is particularly telling: several researchers within TBD Labs have voiced dissatisfaction with the quality of Scale AI’s data, noting that they prefer datasets provided by Surge and Mercor, which they view as being more accurate and specialized.
Scale AI originally established itself with a crowdsourcing model. The company leveraged large pools of relatively inexpensive laborers performing straightforward annotation tasks—such as labeling images or tagging short text snippets. This approach initially served well for earlier AI systems, which required vast amounts of general data rather than highly nuanced input. However, as AI models have evolved into more complex systems demanding context‑specific expertise—ranging from legal reasoning to scientific analysis and medical evaluation—the limitations of Scale’s approach became more apparent. Success in today’s landscape requires data crafted by professionals with deep subject‑matter expertise. Competitors like Surge AI and Mercor designed their operations from the ground up to focus on highly paid, specialized contributors, an orientation that has allowed them to scale quickly in the current paradigm. In response, Scale AI launched the Outlier platform to attract experts in law, healthcare, and other intricate disciplines, but this pivot has required significant time and adaptation.
Adding further pressure to the situation, TechCrunch confirmed that shortly after Meta’s investment was announced, two of Scale AI’s other major clients—OpenAI and Google—abruptly ended their relationships with the data vendor. Within weeks of losing this key business, Scale AI conducted large workforce reductions, cutting approximately 200 positions dedicated to its core data labeling operation. Jason Droege, who recently replaced Wang as Scale’s chief executive, partially attributed this downsizing to changing market demands and shifting priorities. Nevertheless, the company attempted to offset these losses by expanding aggressively into government contracts, including winning a notable $99 million agreement with the United States Army.
While some industry observers suspected that Meta’s immense investment was less about acquiring access to Scale AI’s data and more about recruiting Alexandr Wang himself, uncertainties linger about the true value that Scale is bringing to its corporate benefactor. Wang, a founder steeped in AI infrastructure since 2016, was considered a rising star, and his arrival at Meta was seen as a magnet for elite researchers and engineers. Yet, according to voices within MSL, not all of the Scale executives hired alongside him play prominent roles in TBD Labs’ core missions, and many of Meta’s practical data annotation efforts continue to involve vendors outside of Scale AI.
Instead of stabilizing Meta’s AI program, the integration of new leadership and external recruits appears to have introduced organizational disruption. Reports from both former and current employees describe an increasingly chaotic environment. Researchers newly transplanted from OpenAI or Scale AI often express discontent at confronting the bureaucratic layers and slower tempo characteristic of a vast technology conglomerate like Meta. Meanwhile, members of Meta’s pre‑existing generative AI unit allege their scope has been diminished or redirected, leading to attrition and internal frustration.
This turbulence is unfolding against a backdrop of heightened expectations. After the underwhelming release of LLaMA 4 earlier in the year, a model criticized for failing to match the breakthroughs of competitors, Meta’s CEO Mark Zuckerberg reportedly grew dissatisfied with the trajectory of the company’s AI efforts. Determined not to fall further behind the likes of OpenAI and Google, Zuckerberg quickly pursued an aggressive course: he struck rapid partnerships, purchased smaller AI startups specializing in voice synthesis and generative imagery—such as Play AI, WaveForms AI, and Midjourney—and went on a high‑profile talent acquisition spree, managing to recruit prominent researchers from OpenAI, Google DeepMind, and Anthropic.
Underpinning this surge of activity is Meta’s considerable commitment to the infrastructure required to support world‑class AI models. The company has unveiled massive new data centers across the United States, the crown jewel of which is Hyperion, a $50 billion facility planned for Louisiana. Aptly named after a mythological titan associated with the sun, it symbolizes Meta’s aspiration to become a dominant force in the tech world’s next era.
Choosing Wang to lead such an audacious effort was nevertheless seen as unconventional. Unlike many of his peers in AI leadership positions, Wang is not recognized as a researcher or theorist of artificial intelligence but rather as an operator and founder who excels at building support systems and organizing data pipelines. Reports indicate Zuckerberg explored bringing in more traditional scientific leaders, including conversations with Mark Chen, OpenAI’s chief research officer, as well as attempts to acquire ventures led by AI luminaries Ilya Sutskever and Mira Murati. Each of those overtures, however, ultimately failed.
The instability has already had measurable consequences. Some researchers transplanted from OpenAI into Meta’s ranks have left within a short time, and multiple long‑standing members of Meta’s own generative AI team have resigned amid the restructuring. One of the latest to announce his departure is MSL researcher Rishabh Agarwal, who shared his farewell on the platform X. He acknowledged being inspired by the initial pitch from Zuckerberg and Wang about building transformative superintelligence, but ultimately decided to depart in search of his own risks and opportunities. He declined to elaborate further when asked for details. In recent weeks, other high‑ranking employees, including Chaya Nayak, director of product management for generative AI, and Rohan Varma, a research engineer, have also confirmed their exits.
Despite this enterprise‑wide volatility, Meta Superintelligence Labs is not putting its ambitions on hold. The group is already shaping the next generation of Meta’s AI model, and reports suggest that the company hopes to unveil it before the end of the current year. The central question looming over the initiative is whether Meta can stabilize its leadership structure, retain key talent, and meaningfully integrate external partnerships—particularly given the controversial dynamics surrounding Scale AI—before unveiling its latest advancement.
Sourse: https://techcrunch.com/2025/08/29/cracks-are-forming-in-metas-partnership-with-scale-ai/
