Nvidia $NVDA and Meta $META are expanding a multi-year collaboration, and for me this is mainly confirmation of one thing: while the big players publicly talk about "capex fatigue", in reality they continue to massively lock in capacity for several generations ahead. Meta is set to take millions of GPUs (Blackwell + future Rubin), along with CPUs and networking from Nvidia, and use it in its own data centers both for training and for running models. An interesting detail is that Meta plans to deploy part of the compute through Nvidia Cloud Partners (typically CoreWeave), so it’s not just about owning data halls but also about flexibly "renting" capacity when rapid scaling is needed.

At the same time, I find it significant that Nvidia is pushing into CPUs as well: Meta is rolling out Grace CPU-only servers and plans Vera CPU-only systems in 2027. That’s exactly the kind of shift that could eventually hurt traditional server vendors (Intel/AMD), because a hyperscaler that starts taking a CPU platform from Nvidia is no longer buying just "GPU accelerators" but a chunk of the entire data center architecture. And Meta will also deploy Confidential Computing on GPUs for WhatsApp, which signals that this is not just research but real production use cases.