Meta Bets $15bn On Scale AI

Meta plans to invest $15bn in Scale AI, the data-labelling start-up best known for supplying training materials to some of Silicon Valley’s most advanced models. The deal, one of the biggest of its kind, would give Meta a 49 per cent stake and value Scale at $28bn, according to people familiar with the talks.

The move underlines Meta’s urgency in trying to regain lost ground in artificial intelligence after its latest language model, Llama 4, failed to match the breakthroughs announced by rivals. In benchmark tests, Llama 4 struggled on reasoning and coding tasks, falling short of new-generation offerings from OpenAI, Anthropic and Google.

By tying up with Scale AI and hiring its co-founder Alexandr Wang alongside key members of his research team, Meta aims to accelerate its shift towards building a “superintelligence” lab—one intended to compete directly with the most ambitious projects in the field.

The agreement is expected to be announced within days, with Wang set to take a senior role in Meta’s AI efforts. The 28-year-old founded Scale in 2016 and built it into a cornerstone of the AI ecosystem, providing meticulously labelled data sets essential for training large models. While he is widely credited for his talent in promotion and strategy, multiple former colleagues suggest he played a limited role in technical innovation or team management.

Meta declined to comment on the transaction. Scale AI also refused to respond to questions.

The deal comes at a time when the race for artificial general intelligence—machines that can outperform humans across a wide range of tasks—has grown more competitive. Google and OpenAI have both unveiled new models capable of step-by-step reasoning. Anthropic’s Claude series has been praised for reliability and contextual accuracy. Meanwhile, smaller challengers like China’s DeepSeek have built surprisingly powerful systems at lower cost.

Meta, valued at nearly $2tn, has committed substantial resources to generative AI. But progress has been patchy, and insiders say Mark Zuckerberg has reshuffled leadership several times. In April, Joelle Pineau, vice-president of AI research, left the company, another signal of internal disruption.

The Scale tie-up is Meta’s most ambitious AI move since OpenAI’s ChatGPT triggered a wave of renewed investor and developer interest in large language models in late 2022.

Wang’s expected move to Meta echoes a pattern seen elsewhere in Big Tech. Last year, Microsoft paid $650mn to hire Mustafa Suleyman and other top staff from Inflection, while also licensing the firm’s technology. Google struck a $2.7bn deal with Character AI in a similar fashion. Both companies structured their acquisitions in ways designed to minimise scrutiny from regulators. Yet these deals have not escaped antitrust enforcers entirely.

Should Meta proceed with the Scale transaction as planned, it would surpass those earlier deals in scale and scope, while likely attracting greater attention from regulators in Washington and Brussels.

Inside Scale, the implications are less clear. Jason Droege, who joined the firm from Uber Eats in 2024, is expected to succeed Wang as chief executive. But the fate of other employees is uncertain. Wang had recently spoken about listing the company, though this deal may shelve those ambitions.

Scale has been trying to diversify its income sources, amid investor concern that it relies too heavily on a handful of high-spending AI clients. The company has moved into building custom applications for enterprise clients and is actively bidding for defence and intelligence contracts in the US.

Its core business, labelling and curating training data—remains critical to the AI pipeline. Poor-quality data can result in flawed or biased models. Scale’s output is used by most leading AI firms, and the business has deep ties across the ecosystem, including with OpenAI, Google, and various autonomous vehicle firms.

Yet that position is also a risk. Meta’s decision to invest heavily may force other clients to reconsider their reliance on Scale’s services, particularly if they believe the firm’s allegiance has shifted.

The Llama 4 underperformance appears to have sharpened Meta’s sense of urgency. Unlike OpenAI’s GPT-4o or Google’s Gemini, Llama 4 failed to impress in rigorous third-party evaluations. Critics say it lacks the multi-step reasoning and domain-specific fluency now expected in state-of-the-art models.

Meanwhile, open-source competitors continue to gain ground. DeepSeek, backed by Chinese state-linked funds, has produced competitive models on a fraction of the budget. Hugging Face and Mistral have also made strides with lightweight, performant architectures.

Zuckerberg’s goal is now to create a model that surpasses all current benchmarks, particularly in logic, problem solving and contextual interpretation. The new “superintelligence” lab is meant to focus on this next frontier. The acquisition of Scale, and its leadership, is intended to fast-track that agenda.

Scale’s valuation has now doubled two years in a row, and the $28bn figure being discussed would put it among the top-tier of AI companies globally. It also raises questions about the bubble-like conditions forming around the sector, where start-ups are routinely valued in the tens of billions despite still-developing revenue models.

What distinguishes this latest deal is not just its size, but what it signals: Meta is no longer content to play catch-up. After a period of relative caution and incremental releases, it is now taking bold, expensive bets, reminiscent of its metaverse pivot in 2021. Then, too, Zuckerberg moved decisively, though with mixed results.

Investors may welcome the clarity of intent. But they are also wary. Llama 4’s launch was seen as a missed opportunity, and Meta’s broader AI strategy has so far lacked the cohesion of its peers. Buying into Scale gives Meta access to a crucial input, better-data, and a chance to build a more competitive pipeline.

But as with all things in AI today, execution is what matters. For Meta to rejoin the front rank of AI leaders, this deal will need to deliver not just optics but results. The question now is whether one of the most ambitious AI acquisitions to date will be the turning point Meta hopes it will be, or just another expensive misfire.