After AI setbacks, Meta turns to Scale AI and ‘superintelligence’ research

Meta has recently lost some traction in the AI space, notably halting a major model rollout last month, but the social media company is looking to turn that around with a new $15 billion investment in Scale AI.

The Mark Zuckerberg-led company has reportedly inked a deal to acquire a large minority stake in the startup, which offers data labeling and model evaluation services for industry leaders including OpenAI, Google, and Microsoft.

According to reports, the $14.8 billion investment would give Meta a 49% stake in Scale AI.

This move comes as Meta is also strategically forming a new research lab to pursue “superintelligence,” with Scale AI founder and CEO Alexandr Wang reportedly being tapped to join that initiative.

“Somewhat ironically in the era of AI, Scale AI excels at human-in-the-loop labeling of data,” said Hyoun Park, CEO and chief analyst at Amalgam Insights. “Meta sees this ability to train models and access human curated training data at massive scale as a necessary capability for Meta’s models to keep up with the extremely competitive world of LLMs.”

What Scale AI can bring to Meta

Reports have described Zuckerberg’s frustration with Meta’s AI progress as its competitors, OpenAI, Anthropic and others, continue to innovate and pull ahead. Notably, in May, the company delayed the launch of its new flagship model, Behemoth, purportedly due to internal concerns about its performance capabilities compared to competitors.

“Meta’s models have struggled to keep up with OpenAI and Anthropic in terms of alignment and polish,” said Wyatt Mayham, lead AI consultant at Northwest AI Consulting. Also, he noted, its Llama family of models “haven’t gained much enterprise traction due to weaker instruction tuning and less reliable output quality.”

Mayham pointed out that Scale AI’s “crown jewel” isn’t just labeled data; it’s high-quality enterprise grade human feedback pipelines. Scale would give Meta a fast track path to improve reinforcement learning from human feedback (RLHF) and model steering at scale. The company could dramatically tighten gaps with instruction tuning and output quality by upgrading alignment and task-following performance.

“If this acquisition does indeed go through, it signals that Meta is serious about continued data dominance,” said Mayham.

However, enterprise customers should be both “cautious and curious,” he said.

If Meta owns both the model and the feedback infrastructure, it raises important questions: Who controls alignment priorities? Will fine tuning pipelines be vertically integrated or open? Enterprise teams should watch for lock-in risk if Meta starts to offer end-to-end AI services that compete with open ecosystems, Mayham advised.

And, he said, if enterprises are evaluating which model to bet on long term, this move reinforces the trend that alignment and control are differentiators; not just raw model size. “Whoever owns the human feedback loop owns the intelligence layer.”

‘Superintelligence’ requires a whole new infrastructure

Zuckerberg seems to be going all-in on the development of “superintelligence” — AI systems that exceed human cognitive capabilities. This hypothetical type of AI is the next step above artificial general intelligence (AGI), AI that can match human cognitive abilities. AGI is also still in its hypothetical stages, with experts varying widely on what exactly it could look like, or if it’s even achievable at all.

Behind the hype of AGI is a more “basic threshold of competence” that users are expecting from AI, said Amalgam’s Park. Hallucinations, when AI makes stuff up or outright lies, are actually demonstrations of each LLM’s world view, he noted.

Models need better training and grounding to be more closely aligned to “our worldview, the view of reality and common sense,” said Park. “This investment by Meta is fundamentally focused on providing more human context, metadata, and assumptions into Meta’s next set of models.”

A new type of infrastructure and focus on security and bias is vital as the industry journeys towards AGI, agreed Jimmie Lee, founder and CEO at JLEE.com.

“With the expectation that this superintelligence will far surpass human intelligence, we need to ensure that this new ‘consciousness’ understands the human context,” he said. “Our humanity, the summation of our mindsets, experiences, dreams, and desires, factors into the thousands of decisions we make daily.”

Tomorrow’s AI infrastructure

This potential investment by Meta indicates a shift from a sole focus on LLM development to a more comprehensive strategy centered on the “critical need” for evolved data infrastructure, said Lee.

“As AI and agentic AI continue to develop and grow, the future limiter is not innovation, application, or talent; it will be infrastructure,” he contended. “Currently, modern technologies are outpacing the growth of the very infrastructure that they require to operate.”

For enterprise users, Lee noted, this means deeper integrations and enriched LLMs and data engines that can be more “hyper-specialized and domain-specific,” thus allowing for richer platforms that better support builders. “This results in less infrastructure, greater simplicity, and improved tools and technologies to build on,” he said.

Ultimately, Meta seems to be re-shifting its strategy, Lee noted: It’s going back to its roots of making large bets to try to drive innovation, rather than merely responding to market demands and increasing market share.

Park agreed: “Zuckerberg knows that AI is the biggest battle in tech and intends to do everything he can to make Meta one of the global giants in artificial intelligence.”