The author is a Reuters Breakingviews columnist. The opinions expressed are his own.

By Sebastian Pellejero

NEW YORK, April 9 (Reuters Breakingviews) - Mark Zuckerberg is not prone to hedging. The boss of social media giant Meta Platforms META.O has a mixed track record in that regard: an early bet on mobile worked spectacularly; a pivot to the virtual-reality metaverse, less so. He is now staking hundreds of billions of dollars on artificial intelligence. This week's debut of Muse Spark, the first model from Meta’s Superintelligence Labs division, narrows the technological gap with leading rivals. Unlike them, though, it will primarily power Facebook, Instagram, and WhatsApp. Fortunately, that’s a strong enough business to potentially justify the massive cost.

Muse Spark performs competitively against models made by OpenAI, Anthropic and Google on key benchmarks, a striking turnaround after a disappointing previous effort, Llama 4. For Superintelligence head Alexandr Wang, it vindicates the decision to rebuild the company’s AI efforts from scratch, staffing the group by poaching engineers from rival labs at eye-watering cost. An uptick in Meta’s share price on the news also helps to stem a nasty decline since it acquired his company, Scale AI, for $14.3 billion last June.

The stakes are considerable. Take ChatGPT’s end-of-2022 release as the start of the AI era. Since then, Meta has invested nearly $140 billion in capital expenditures and may spend as much as $135 billion more this year. To generate a 15% return, the company needs roughly $41 billion in incremental operating profit. At current margins of around 41% and holding all else equal, that means growing revenue to $304 billion, about 9% annually over five years.

For a company expected to grow 25% this year, this bar seems manageable. It raises the question, though, of how much growth would have happened either way: Meta is dominant in its industry. Still, there are multiple ways to glean value. AI-driven cost savings, from cheaper computing power to fewer content moderators, can improve profitability. Better models sharpen ad targeting, potentially allowing Meta to raise prices.

Risks remain. Pivoting to proprietary models abandons the open-source strategy that won initial credibility with developers. Wang has never run a research lab, and his expertise in data processing does not automatically translate into the scientific culture that produces genuine breakthroughs. Meanwhile, competitors like OpenAI and Anthropic are spending massively to further their lead. Meta may need to do likewise so long as the arms race continues.

Still, while standalone AI labs chase unproven sales models or try to mimic e-commerce and advertising success, Meta simply needs to make an existing business work better. Betting the house is less foolhardy when the foundation is already in place.

Follow Sebastian Pellejero on LinkedIn.

CONTEXT NEWS

On April 8, Meta Platforms unveiled Muse Spark, the first model from its newly formed Superintelligence Labs. In a departure from earlier releases, Muse Spark will be a closed model that will power Meta's chatbot and AI features. The company said it planned to release a private preview of the model to a few partners via an application programming interface.