Exo is back after nearly 10 months, run your own DeepSeek v3.1 671B with RDMA at 32.5 t/s

https://exolabs.net
https://github.com/exo-explore/exo
https://www.jeffgeerling.com/blog/2025/15-tb-vram-on-mac-studio-rdma-over-thunderbolt-5

If you have about 40-50k USD for 4 Mac Studio with 512GB RAM each, you can run your full DeekSeek at 32.5 t/s. Probably the cheapest way to do it locally, Nvidia setup will cost you around 5-10x more.

Thought they abandoned the project. No update since Feb. But I guess that is the "curse" of working with secretive Apple tech. Have early access to cool tech like RDMA which is still in beta. Will Apple make a comeback to server market? Or is this just a good demo to enterprise users on how they can build desktop AI rack? Will cost around 50k for 4 top spec Mac Studio, but something that you can't easily do with any Nvidia rig at the same price.