Meta’s Infrastructure Evolution and the Advent of AI
Meta's 21-year infrastructure retrospective, from LAMP to a 129k-H100 cluster and gigawatt-scale builds (Prometheus, 5GW Hyperion by 2028). Context on where frontier training capacity is heading, not something to use.
**The gist** Meta traces 21 years of infrastructure, from the LAMP stack to AI scale: training grew from 128-GPU jobs to a single **129k-H100** cluster spanning five data center buildings, with **Prometheus (1 gigawatt)** and **Hyperion (5 gigawatts, targeted 2028)** next, plus custom **MTIA** silicon for ranking inference.
**Why it matters** The post maps what frontier-scale training actually requires — power, cooling, networking, orchestration — and Meta restates its open-hardware commitment via the **Open Compute Project**, where it accounts for roughly **25% of tech contributions**. Useful for calibrating how much capacity sits behind the hosted models your agents call.
**The gist** Meta traces 21 years of infrastructure, from the LAMP stack to AI scale: training grew from 128-GPU jobs to a single **129k-H100** cluster spanning five data center buildings, with **Prometheus (1 gigawatt)** and **Hyperion (5 gigawatts, targeted 2028)** next, plus custom **MTIA** silicon for ranking inference. **Why it matters** The post maps what frontier-scale training actually requires — power, cooling, networking, orchestration — and Meta restates its open-hardware commitment via the **Open Compute Project**, where it accounts for roughly **25% of tech contributions**. Useful for calibrating how much capacity sits behind the hosted models your agents call. **Watch out** It is a retrospective plus roadmap with **no cost figures**; the gigawatt clusters are **still under construction**, and Meta itself says no one can predict how AI workloads will evolve.