Meta bought 49% of Scale AI at a roughly $30 billion valuation. That number matters less than what it signals: a company with $100 billion in annual ad cashflow has decided it is losing the AI race and is spending accordingly. DeepSeek took Meta's open-weight model lead. Llama 4 Behemoth failed. Those two facts are the proximate cause of everything that follows.
Zuckerberg is personally recruiting AI researchers at $200 million per person over four years, 100x standard compensation. Some OpenAI leadership received offers exceeding $1 billion. None accepted. He has also scrapped Meta's conventional datacenter playbook entirely, now deploying multi-billion-dollar GPU clusters inside temporary structures, trading construction timelines for raw speed. A new internal Superintelligence team is being built from scratch, not reorganized from existing headcount.
The full piece tracks Meta's trajectory from Llama 3 dominance to Llama 4's stumble, then maps out the compute ramp SemiAnalysis believes will rival OpenAI's training FLOP growth. The capex versus buyback reallocation charts and the datacenter model specifics are where the real argument lives. Read it to understand whether this is a credible recovery or an expensive panic.
[READ ORIGINAL →]