Dylan Patel's SemiAnalysis grades Meta's AI reset one year after the 'disastrous Llama 4' launch. Meta has spent to rebuild talent, $14.3B to bring in Scale AI's Alexandr Wang, $1B+ for Nat Friedman and Daniel Gross's fund, and a wave of ex-OpenAI researchers, and just stood up a 3,000-engineer org (70% new grads and seniors) working full-time on reinforcement-learning tasks and environments. The compute story is the headline: five 'titan' clusters of 1GW or more are under construction at once, Prometheus (Ohio), Hyperion (Louisiana), plus El Paso, Iowa and Indiana. Prometheus has ballooned from ~1GW to over 3GW in two years using a tent datacenter design; Hyperion is 1.5GW across six buildings. To stitch sites together Meta built 'AI-Backbone,' linking up to 27 datacenters across six campuses at 22 petabits/second, with titans reaching campuses up to 2,000km apart. SemiAnalysis projects Meta will have more AI compute than OpenAI and Anthropic by end of 2026. The caution: on models, Muse Spark 1.1 only tests roughly on par with Opus 4.6, and 'success is far from guaranteed.'
Why it matters · Five gigawatt-class clusters going up at once is a sustained capex signal for GPUs, optical networking and power, even as the article warns Meta's model race is far from won.
Worth asking · Meta is buying the most compute of any hyperscaler, but can it turn watts into a frontier model?