GB300 is reported on track for 70-80% of global AI server rack shipments
Supply-chain reporting puts Nvidia's GB300 platform at an estimated 70-80% of AI server racks shipped worldwide in 2026, with the Vera Rubin 200 platform expected to broaden after the third quarter. A single GB200 NVL72 rack already draws 120-140 kW.
Concentration at that level makes the platform an industry-wide dependency rather than a vendor choice. When three-quarters of the world's AI racks are one design, its supply schedule, thermal envelope and power draw stop being Nvidia's engineering constraints and become everyone's planning constraints — including the utilities that have to serve them.
The power figure is where it gets physical. A rack pulling 120-140 kW is not a datacentre line item, it is a substation conversation. Multiply by a hall and the binding constraint stops being GPU allocation and becomes interconnection queues, transformer lead times, and the willingness of a regional grid operator to say yes.
Rubin entering full production with CoreWeave integrating in the second half sets up the familiar overlap problem: two generations of very different power profiles occupying the same buildings, with facilities specified for the older one.
Finviz — Nvidia GB300 chips become backbone of global AI data centers for 2026 → · Intellectia — NVIDIA market outlook, August 2026 → · IES VE — From data centres to AI factories: what NVIDIA GTC 2026 signalled →