The AI startup’s biggest clients, which include OpenAI and Google, won’t likely want to trust their data to a company almost half owned by Mark Zuckerberg’s tech empire.
Scale AI’s shock $15 billion deal to sell a 49% stake to Meta, with its CEO Alexandr Wang leaving the startup he founded to lead a new AI lab at the tech giant, has raised the specter that the startup’s $14 billion-valued data labeling business could implode.
At play is the concern that Scale, which has been the dominant player in labeling data to help major tech companies and AI startups train their models, could share details about the types of data that leading AI players have used to build their most cutting edge tech with Meta.
As one former Scale employee told Forbes: “They all want to cut Scale off now. Scale as a business, once it becomes part of Meta, entirely collapses.”
OpenAI, one of its most prominent customers, has already been winding down its work with Scale, according to four sources with knowledge of its business; two sources said this has been happening for months and OpenAI has been vetting potential new partners. OpenAI declined to comment.
The surprise move, which would value the company at $28 billion, left people inside Scale scrambling and confused, the former employee said. Some staff are concerned over what visibility Meta might have into past projects, even though most contracts with the company stipulate data be deleted after projects are completed, the person added.
Scale initially declined to comment. After publication, Scale AI spokesperson Joe Osborne said, “This reporting seems to be fueled by smaller competitors and couldn’t be more off the mark.” He denied that OpenAI has pulled back its spending with Scale.
Scale’s smaller rivals are already jostling for position, and are welcoming any customers worried about conflicts of interest when it comes to data and privacy. Invisible Technologies cofounder Francis Pedraza told Forbes his company is committed to staying independent. Startup Turing, which already supplies data for model training to OpenAI, Anthropic and Google, sees the deal as an opportunity to be “Switzerland” and become an impartial distributor of data to the frontier AI labs. CEO Jonathan Sidharth told Forbes that “clients want to work with someone neutral who could support all the labs equally.” One investor who has backed a Scale competitor said the deal would create new avenues for other firms to “capture the open space left by Scale AI.”
Access to high quality human data is crucial for training powerful AI models, so much so that data has been considered a moat for AI juggernauts like OpenAI and Anthropic in the AI race. Scale AI, which booked $870 million in revenue in 2024, captured the data labelling market early by supplying troves of human-labelled data to companies like Cohere, OpenAI and Microsoft. But Meta owning almost half of Scale would change that equation.
Scale built its business on the back of a massive clickworker army, largely of workers based overseas who would teach AI models by providing context for masses of data. But this kind of work has become a commodity. “Anyone who puts up a team can compete with you, and it comes down to price really quickly,” said Kevin Guo, cofounder of a startup that got out of the data labeling business, in a 2023 interview about Scale.
One senior AI company executive described Scale as “the bulk food section of the AI training market.” Multiple sources familiar with Scale’s business told Forbes the startup has suffered from quality problems. “They overpromise and they oversell, and then they underdeliver very often,” said the former employee.
Wang, who is one of the world’s youngest self-made billionaires at age 28 with an estimated net worth of $3.6 billion, will head up a new lab at Meta focused on building so-called superintelligence — an AI system that outperforms human capabilities, The Information reported. To build out the team around him, Zuckerberg is apparently offering researchers from Scale AI as well as OpenAI, Anthropic and Google Deepmind north of $10 million a year to join the new lab. Zuckerberg is closely involved in assembling the team and has gone to great lengths like setting up a WhatsApp group called “Recruiting Party,” personally reaching out to potential recruits and rearranging desks for researchers to sit near him, Bloomberg reported.
The deal is yet to close, and could be blocked by regulators. If it does, it will be a windfall for Wang and Scale’s early investors, including Accel and Index Ventures. But it’s not clear what will happen to Scale at that point. “This was great for Alex and early investors, terrible for everyone else including employees and former employees,” said a former senior Scale AI employee. “Unclear how the deal helps Scale.”
The most powerful tech giants in the world have spent the last several years trying to out-jockey each other for AI supremacy. Meta, which first launched its AI lab in 2013, has struggled to keep up with Google, OpenAI and Anthropic with its open source Llama models. The company’s AI reputation took a hit in April when it was accused of artificially booting the benchmark scores of its Llama 4 models (Meta denied the claim). A glossy new hire like Wang could be a shot in the arm for the company’s AI efforts.
Zuckerberg has also started pushing Meta toward defense contracts, and Scale’s government arm could dovetail with those efforts (though as Forbes previously reported, that part of its business has previously struggled to gain traction).
Meta is the latest company to hire away the CEO of a splashy AI startup. Last year, Microsoft poached Mustafa Suleyman, a founder of the lab DeepMind, and his senior team from Inflection, the AI lab Suleyman started in 2022, leaving a shell of a company behind. Months later, Amazon inked a deal with Adept to license its tech and hire CEO David Luan and his founding team. Google made a similar gambit to re-hire researcher Noam Shazeer — who co-invented the transformer architecture that underpins generative AI during his tenure at Google — from his startup Character.AI.
CEO of Invisible Matt Fitzpatrick noted that as models get more complex, they need higher levels of expertise. Scale, Turing and Invisible have all moved away from low-paying clickwork to emphasize more sophisticated tasks performed by PhDs and highly educated professionals. Meta’s deal reinforces the importance of human workers when it comes to training AI, he said. “This is a 10-year bet that this is going to be really important for a long time,” Fitzgerald said.
John Paczkowski, David Jeans and Iain Martin contributed reporting.
This story has been updated with additional comment from Scale.
Read the full article here
