$SMCI is now shipping $NVDA Vera Rubin NVL72 racks, integrated with its Data Center Building Block Solutions and the DLC-2 direct liquid cooling stack. Per rack: 1. 72 $NVDA Rubin GPUs and 36 Vera CPUs, operating as one machine 2. 18 1U compute trays, each with 4 Rubin GPUs and 2 Vera CPUs 3. 9 sixth generation NVLink switch trays, 216 TB/s of scale-up bandwidth 4. 20.7 TB of HBM4, up to 54 TB of LPDDR5X 𝗢𝗻𝗲 𝗦𝗰𝗮𝗹𝗮𝗯𝗹𝗲 𝗨𝗻𝗶𝘁 𝗶𝘀 1,152 𝗥𝘂𝗯𝗶𝗻 𝗚𝗣𝗨𝘀 𝗮𝗻𝗱 331 𝗧𝗕 𝗼𝗳 𝗛𝗕𝗠4 𝗮𝗰𝗿𝗼𝘀𝘀 16 𝗰𝗼𝗺𝗽𝘂𝘁𝗲 𝗿𝗮𝗰𝗸𝘀. The blueprints size a balanced bill of materials from a 5 MW envelope up to gigawatt scale. Cooling is what Supermicro is really selling, and it comes from its own portfolio: 1. Cold plates, manifolds, hose kits 2. In-row CDUs rated 1.8 MW each, N+1 redundancy 3. Rear door heat exchangers for residual heat 4. Facility-side cooling towers Every rack is tested with the full stack before it leaves. "We have spent years building the liquid-cooling stack, the manufacturing capacity, and the deployment teams for exactly this moment," said Charles Liang, president and CEO, Supermicro. Supermicro also said on its own X account that the first systems went to SpaceX AI, built at the expanded San Jose facility. Shipping first is the win. The open question for $SMCI is the margin these racks carry, because this is a company that has taken thin ones before to hold volume.