Meta’s $14.3B stake triggers Scale AI customer exodus, could be a windfall for rivals like Mercor

Scale AI has been attempting to quell fears about its company sovereignty and data security after its ‘acqui-hiring’ by Meta, but customers appear to be defecting anyhow, and competitors are being rewarded with a slice of the limelight.

Meta is investing $14.3 billion in the data labeling and model evaluation startup, which gives the social media giant a 49% stake in the company, and is bringing Scale’s founder and former CEO Alexandr Wang onboard to work on AI “superintelligence.”

Within days of the news of the deal, OpenAI said it would be phasing out its work with Scale, although not explicitly because of the Meta deal. For the last several months, the AI leader has been backing away from the relationship and opting for competitors like Mercor, reportedly because Scale doesn’t have the expertise it needs for its increasingly advanced models.

Others are also purportedly hitting the brakes on their relations with Scale, including xAI and Google, the latter over concerns that Meta could access information about its AI developments.

Scale’s interim CEO Jason Droege has pushed back, emphasizing in a blog post that the company will remain “unequivocally an independent company” and will not provide Meta with access to its internal systems.

Despite this assurance, an analyst understands industry concerns.

“Meta’s move signals a trend toward vertical integration and supplier lock: Owning the data annotation pipeline to secure control over the quality, provenance, and scalability of training data,” said Thomas Randall, AI lead at Info-Tech Research Group. “Moreover, OpenAI’s pullback shows how quickly partnerships in this space can shift based on alignment, data strategy, or concerns about competition.”

Rivals in the data labeling game

Data labeling is a critical step in AI development, as it involves tagging raw data to provide context for models so they can continue to learn and iterate.

The Meta-Scale deal underscores the importance of the capability, and, perhaps counterintuitively, has drawn much more attention to rival, potentially superior data labeling companies. This includes five-year-old startup Surge, which reportedly had more than $1 billion in revenues last year. Others in the growing space include Turing, Snorkel, Invisible, Toloka, CloudFactory, and Label Your Data.

However, Droege asserted that Scale is “one of the only providers capable of serving customers at volume” with the “largest network of experts training AI.” Going forward, the company will focus on building out its applications business units and will continue to be model-agnostic and human-driven, he said.

“The spike in competition from players like Surge, Turing, and Invisible gives enterprises more leverage, but also more responsibility,” said Info-Tech’s Randall. These vendors differ significantly when it comes to workforce models, automation levels, and quality controls, he noted. Enterprise leaders should evaluate providers not just on price or throughput, he advised, but on whether they offer robust annotation auditability, support for domain-specific edge cases, and alignment with ethical AI practices.

“The quality of labeled data is a leading indicator of model performance and a lagging indicator of strategic oversight,” said Randall. “The enterprises that succeed in AI won’t just be the ones with the best models, but the ones with the most intentional, resilient data ecosystems.”

Not just about selecting a labeling company

But the ultimate conversation around data labeling is a little more nuanced and complex, analysts note.

Hyoun Park, CEO and chief analyst with Amalgam Insights, pointed out that Scale has built its reputation on text and image labeling, and its ability to identify global talent. This is a “powerful fit” for Meta, as Facebook, Instagram, and its other applications and services have massive amounts of data that can be further tagged and indexed to support large language models (LLMs) and AI, based on Meta’s ownership of accounts and digital assets.

“Scale works well with social networks and other media-based websites with self-refreshing and original media creation that can be labeled and used to train models on an ongoing basis,” he noted.

For OpenAI, Google, Anthropic, and other LLM providers selling directly to businesses and large organizations, however, the competitive landscape is quickly shifting. It is no longer enough to simply take in and process general data; providers must be able to automate code and conduct higher-level tasks, said Park. When digging deeper into programming, healthcare, legal services, and other specialized fields, they need subject-matter expert data.

Enterprises must be able to contextualize their own internal data and jargon, and have the ability to trust their AI enough to allow it to take action, he said. This means that the AI needs to be trained well enough to understand the common sense ramifications of the requests it receives, and the data that it accesses.

“This training and contextualization ultimately requires specific expertise that is often coming from veteran employees and highly trained professionals, not just from outsourcing firms that can provide scale-up capabilities for specific areas of AI training,” said Park.

Randall agreed that enterprise leaders must treat their data labeling decisions as part of a broader AI governance and operational strategy, not just a technical outsourcing choice. He said his firm’s research on vendor management indicates that organizations should treat labeling vendors as they would treat cloud providers.

That is: “diversify, insist on explicit contractual firewalls around staff mobility and data reuse, and build contingency plans so an acquisition doesn’t strand your model pipeline or expose proprietary data,” he said.