capexScore 50/100Watch
Meta planning GPU scaling from 100K to beyond for Llama 5; hundreds of billions capex bet on continued scaling
“Lama 3, we could train on 10 to 20,000 GPUs. Lama 4, we could train on more than 100,000 GPUs. Lama 5, we can plan to scale even further... we're basically making these bets on how much infrastructure to build out for the future. And this is like hundreds of billions of dollars of infrastructure.”
Why it matters
Meta is explicitly committing to scaling GPU clusters beyond 100K units for Llama 5 and beyond, with infrastructure spend in the hundreds of billions. This signals sustained demand for high-end GPUs and validates the capex cycle thesis for semiconductor and infrastructure suppliers.
Investment implication
Direct beneficiaries include NVIDIA (GPU supply), advanced packaging providers, and data center infrastructure companies. This also supports the thesis that hyperscaler capex will remain elevated through the mid-2020s as foundation models continue scaling.