We were the first cloud provider to bring up and validate @​NVIDIA Vera Rubin NVL72. The underlying challenge is making several racks behave as one cluster. That means automated rack lifecycle (Racky, Valvey, Rack LifeCycle Controller), straggler detection down to a single slow GPU, and a two-tier, non-blocking RoCE fabric at 1.6 Tb/s per GPU, designed to reach 128,000 GPUs. How we did it: 📷 📷 📷 📷 ↧ Multi-Rack NVIDIA Vera Rubin NVL72 Bring-Up CoreWeave Go inside multi-rack NVIDIA Vera Rubin NVL72 bring-up at CoreWeave: rack automation, GPU straggler detection, a non-blocking RoCE fabric, continuous validation.