AI factory infrastructure still the choke point

NVIDIA expanded its Vera Rubin NVL72 server family and reinforced a shift toward continuous, pod‑scale inference with new Rubin and Grace Blackwell hardware highlighted in keynote and infrastructure sessions at GTC 2026. (datacenterfrontier.com)) Microsoft began formal validation work on NVIDIA’s Vera Rubin NVL72 this month and signed into vendor validation pipelines that major cloud providers are using to certify next‑gen AI hardware. (msn.com)) Nscale and Microsoft announced a letter of intent to deliver 1.35 gigawatts of Vera Rubin NVL72 capacity for a flagship AI factory campus in West Virginia, underscoring multi‑gigawatt campus planning by hyperscalers and colo operators. (tmcnet.com)) Wiwynn demonstrated NVL72‑based rack‑scale platforms at GTC on March 16, 2026, showing liquid‑cooled, board‑to‑rack integrations developed with Wistron and promoting higher density deployments at booth #1121. (wiwynn.com)) Vendor materials presented NVL family configurations that scale at rack level—examples cited 8 to 16 systems per rack and fully liquid‑cooled designs aimed at improving PUE for dense Rubin deployments. (tmcnet.com)) Fourteen ecosystem partners including Dell, HPE, Google, Microsoft, Lenovo, Supermicro and Vertiv were named as GTC collaborators providing hardware, networking and software components for enterprise AI factory reference designs. (crn.com)) Commercial system vendors and integrators announced production programs: Aleria described deployments of Blackwell Ultra and DGX Vera Rubin NVL72 with DDN for sovereign AI, while AMAX and other hosts unveiled NVL72‑based hosting and rack solutions for enterprise customers. (azcentral.com))