Meta bought 49% of Scale AI at a roughly $30 billion valuation. That is the opening move of a company that lost its open-weight model lead to DeepSeek and is now treating the gap as an existential problem. Zuckerberg has personally taken charge, identified talent and compute as the two critical deficits, and slashed stock buybacks to fund the offensive.
The talent strategy is blunt: $200 million over 4 years is the standard offer for researchers being recruited into the new Superintelligence team, roughly 100x peer compensation. Some offers targeting OpenAI leadership crossed $1 billion and were still declined. On the compute side, Meta has scrapped its traditional datacenter playbook entirely and is deploying multi-billion-dollar GPU clusters inside temporary tent structures to compress build timelines. SemiAnalysis projects Meta's training FLOP ramp will approach OpenAI's pace, moving the company from GPU-poor to GPU-rich on a per-researcher basis.
The full report traces exactly how Meta went from Llama 3 open-source dominance to the Llama 4 Behemoth failure, and what the infrastructure, org structure, and data strategy for the Superintelligence team actually look like. The numbers on capex versus buybacks alone justify reading it in full.
[READ ORIGINAL →]