News

Meta Invests in Scale AI While OpenAI and Google Step Back

• By John K. Waters
• 06/20/2025

In Silicon Valley's fast-shifting AI frontier, there are breakups that whisper, and others that shout across the landscape. OpenAI quietly began disentangling itself from Scale AI months ago, but Meta's $14.3 billion investment in the data-labeling firm turned the volume all the way up.

The partnership between OpenAI and Scale AI had never been front-page material. Scale, founded in 2016 by Alexandr Wang, built its business on wrangling armies of contractors to manually tag images, text, and video—grist for the AI model training mill. The firm counted OpenAI, Google, and Meta itself among its high-profile clients, supplying the kind of annotated data that early-generation AI models crave.

But those days are already past. Over the last year, OpenAI began searching for more "specialized" providers, pivoting away from Scale well before Meta made its move. An OpenAI spokesperson told Bloomberg that the ChatGPT maker had been moving on from Scale not out of competitive concern, but because the startup's services simply didn't match OpenAI's trajectory anymore: "more advanced AI models that can mimic human reasoning" and "agent-like models" needed a different class of support.

Then Meta entered the frame—not just with cash, but ambition.

Meta's Billion-Dollar Talent Grab
Meta's AI ambitions have shifted from steady iteration to outright acceleration. Sources told CNBC and Bloomberg that before the Scale deal, Meta made overtures to acquire Perplexity AI, a rising AI-powered search company that poses a potential threat to Google. That deal fizzled, but Meta didn't stop there. It pursued Safe Superintelligence (SSI), a boutique AI lab founded by ex-OpenAI scientist Ilya Sutskever and Daniel Gross. That deal also didn't land—but Meta did manage to recruit Gross, alongside former GitHub CEO Nat Friedman, into its ranks.

The $14.3 billion injection into Scale, which nets Meta a 49% stake but no voting power, comes with something more important: people. Scale's founder Wang left to lead Meta's new "superintelligence" unit, taking some of his former colleagues with him. That unit is tasked with what CEO Mark Zuckerberg has reportedly branded as Meta's moonshot: building next-gen AI powerful enough to compete with OpenAI's models.

"Meta offered OpenAI staff $100 million signing bonuses," OpenAI CEO Sam Altman said during a recent episode of the Uncapped podcast. "Their current AI efforts haven't worked as well as they hoped… but I respect being aggressive."

Meta's frustration is evident. While it has released solid research and competent models—like LLaMA 3—it hasn't made the same public splash as GPT-4 or Google's Gemini. In contrast to rivals investing in cloud compute infrastructure or foundation models, Meta is betting big on human capital and data annotation—especially the kind that Scale used to provide to OpenAI and Google.

Who Controls the Data Controls the Future
At the heart of the shake-up is a single, unglamorous truth: AI models are only as good as the data they learn from. And that data isn't just petabytes of text scraped from the web—it's also nuanced, hand-labeled, and often proprietary. As AI models evolve to become agents and reasoning engines, the demand shifts from volume to expertise.

Scale, once reliant on gig workers, had begun hiring PhDs and domain experts to label complex medical and scientific data. But even that shift wasn't enough for OpenAI. The company has since looked toward newer players like Mercor, a startup originally known for AI-assisted tech recruiting. Mercor now focuses on finding expert-level human intelligence for labeling data to train more contextually aware models.

Meanwhile, Meta's entanglement with Scale triggered concerns among other clients. Google, also a Scale customer, reportedly began cutting ties after the Meta deal, wary of giving a rival potential access—real or perceived—to its training pipelines. Scale, for its part, insists it's staying independent.

"Nothing has changed about our commitment to protecting customer data," said interim CEO Jason Droege.

That's cold comfort in an industry where talent and trade secrets are separated by only a few Slack exports.

Why This Matters
This isn't just about one vendor change. The Scale-Meta-OpenAI triangle reveals the increasingly zero-sum nature of AI's next phase. In the age of foundational models, controlling the supply chain of human-labeled data is strategic. As public data becomes increasingly commoditized—or litigated—companies are turning inward, investing in proprietary data creation and poaching high-value talent to manage it.

Meta is betting that data-labeled at Scale (and the people behind it) will give it a critical advantage. OpenAI believes it's moved past what Scale can offer. But as the lines between model developer, data pipeline owner, and cloud infrastructure provider blur, no one is staying in their lane.

And if the current AI wars are fought with models, the next ones may well be fought with the people and processes behind the labels.