capexScore 65/100Watch
Meta scaling GPU requirements exponentially: Llama 5 planned to exceed 100K GPU training runs, driving massive capex
“Llama 3, you know, we could train on, you know, 10 to 20,000 GPUs. Llama 4, we could train on, you know, more than 100,000 GPUs. Llama 5, we can plan to scale even further”
Why it matters
Zuckerberg's disclosure of GPU scaling trajectory (10–20K → 100K+ → beyond) reveals an accelerating capex commitment. This is first-party confirmation of infrastructure buildout that will drive semiconductor and power demand through at least the next 2–3 model generations.
Investment implication
Beneficiaries include GPU suppliers (NVIDIA), memory vendors (Samsung, SK Hynix for HBM), datacenter infrastructure players, and power/cooling infrastructure. This also signals Meta's commitment to capturing LLM economics and suggests competitive pressure on other hyperscalers to match or exceed these training scales.