Can Scale AI and Alexandr Wang reignite Meta’s AI efforts?

Summary

Meta is reportedly investing nearly $15 billion dollars in the data labeling firm Scale AI and taking a 49% stake in the startup, while also bringing on CEO Alexandr Wang to help lead a new “superintelligence” lab within the company. The deal is reminiscent of Meta’s previous large and risky bets, such as its $19 billion acquisition of WhatsApp and its $1 billion Instagram purchase. At the time those mergers closed, many people suggested Meta had grossly overpaid for the platforms — and today’s discourse is no different. There’s no shortage of investors and founders who were left scratching their heads this weekend over Meta’s latest tie-up. In the end, WhatsApp and Instagram became an integral part of CEO Mark Zuckerberg’s empire. The question is whether the Scale AI deal will similarly work in Meta’s favor, proving Zuckerberg’s prescient strategy yet again, or whether the company is grasping at straws in a misguided effort to catch up to rivals like OpenAI, Google, and Anthropic. In this case, Meta isn’t betting on an emerging social media app, but on the data used to train top AI models. For the last several years, leading AI labs such as OpenAI have relied on Scale AI to produce and label data that’s used to train models. In recent months, Scale AI and its data annotation competitors have started hiring highly skilled people, such as PhD scientists and senior software engineers, to generate high-quality data for frontier AI labs. It may benefit Meta to have a close relationship with a data provider like Scale. Meta’s leaders have complained about a lack of innovation around data in the company’s leading AI teams, according to a person familiar with the matter. Earlier this year, Meta’s GenAI unit launched Llama 4, a family of AI models that failed to match the capabilities of Chinese AI lab DeepSeek’s models, and which was broadly seen as a disappointment. Not helping matters, Meta is trying to combat an attrition problem. According to data compiled by SignalFir...