Sooth Labs, a stealth artificial intelligence startup founded by former Meta researchers, is raising about $50 million at a roughly $335 million post-money valuation to build AI models that forecast the likelihood of specific geopolitical and market events, according to a Bloomberg report published late on April 22, 2026. Felicis Ventures is set to lead the round, with Turing Award laureate Yann LeCun and Google chief scientist Jeff Dean among the backers. Meta CTO Andrew Bosworth is advising the company.
The round turns a quiet research experiment into one of the better-capitalized new entrants in what has become a crowded corner of AI: models built to reason over the messy, open-ended question of what will happen next in the world.
The team behind the round
Bloomberg names Chuck Hoover, Yaser Sheikh, and Ruslan Salakhutdinov as the co-founders. All three come out of Meta's AI and research organizations, and Salakhutdinov is a well-known machine-learning researcher whose career has straddled academia and industry. Bosworth's advisory role and LeCun's investment extend the Meta research-network fingerprint on the cap table, even as the company positions itself as independent from its founders' former employer.
Felicis's lead check is consistent with the firm's recent appetite for frontier-AI bets at the seed and Series A stages. The backing from Dean, who oversees Google's foundational AI work, is the more surprising signal: it gives Sooth Labs validation from one of the few people in the industry with a direct view of how top-tier event-prediction systems might interact with Google-scale infrastructure.
Forecasting as an AI frontier
Sooth Labs's pitch — models that forecast real-world events — lands at an inflection point for the category. Prediction markets have ballooned in the last year, and enterprises are under pressure to put structured forecasts on everything from elections and conflict risk to commodity shocks and regulatory changes. Most of today's tools either rely on expert panels or on thin-wrapper applications sitting on top of general-purpose frontier models.
A purpose-built forecasting model, if Sooth Labs can deliver one, would slot into risk, policy, and trading workflows in a way that a general chatbot cannot. The technical bet is that calibrated probability estimates over future events are a distinct capability from general reasoning — one that benefits from its own training data, evaluation stack, and feedback loops.
What to watch
Details on the architecture, training data, and go-to-market remain unpublished. The company has not yet announced a public product, and the funding itself has not been formally confirmed beyond Bloomberg's reporting. For enterprise buyers tracking the event-forecasting space, the names attached to Sooth Labs matter as much as the product: a Felicis-led round with LeCun and Dean on the cap table sets a high bar for credibility in a category where overpromising is the norm.
