One rack, one neighbourhood
A GB200 NVL72 rack draws 120-140 kW. That platform family is heading for three-quarters of global AI rack shipments. At that point the binding constraint on AI is not silicon — it is the substation.
GB300 is projected to take 70-80% of AI server racks shipped worldwide this year. Concentration at that level stops being a vendor's market share and becomes an industry-wide dependency.
Do the power arithmetic
One rack at 120-140 kW is roughly the continuous draw of a hundred homes. A modest hall of forty is comparable to a small industrial plant. This is why datacentre siting has quietly become an energy-policy question: the limiting resource is not GPU allocation, it is interconnection queues, transformer lead times, and whether a regional grid operator says yes.
Those constraints run on entirely different clocks from AI. Transformer lead times are measured in years. Grid interconnection studies are measured in years. Model generations are measured in months. An industry planning on eighteen-month cycles is now dependent on infrastructure planning on decade cycles, and nothing about that mismatch resolves itself.
Two generations, one building
Rubin is in full production with partner systems in the second half and CoreWeave integrating, which sets up the familiar overlap: two generations with materially different power and thermal profiles occupying facilities specified for the older one. Retrofitting power and cooling into a running datacentre is among the least pleasant projects in the industry, and it is about to be very common.
The strategic move was not the chip
Nvidia expanding outward from the accelerator into CPUs, networking and the rack specification is the decision that will matter in five years. Owning the accelerator is a strong component position. Owning the integrated stack converts it into a platform position, and platform positions are displaced far more rarely.
It is also the honest explanation for why credible alternatives keep failing to convert. AMD's MI300 line, Intel's Gaudi, and the custom silicon at Google, Amazon and Microsoft are all real products with real customers. What none of them replaces is the surrounding system, and the switching cost has migrated there.
The risk for buyers is not price — competition still disciplines that. It is optionality. Every generation shipping as an integrated rack is a generation in which the practical ability to mix vendors declines, and that decline does not show up on any invoice until the moment you need it to be false.
NVIDIA Newsroom — NVIDIA kicks off the next generation of AI with Rubin → · The Next Platform — Nvidia extends its grip on the AI datacenter outwards → · CNBC — Nvidia's new PC chips represent CEO Huang's bid to win at every layer of the AI stack →