NVIDIA and Oracle Take Vera Rubin to Multi-Gigawatt Scale - Data Centre Magazine

The multi-gigawatt deployment is fundamentally an energy infrastructure repricing, not a compute race. Vera Rubin clusters pushing into GW territory force Oracle to transform data centers from IT facilities into quasi-power-plants—dedicated substations, closed-loop liquid cooling, and SMR nuclear integration become non-negotiable. Upstream value shifts accordingly: CoWoS-L packaging (TSMC, Taiwan, China) and HBM4 (SK Hynix) command escalating capacity premiums, while liquid-cooling and power-equipment suppliers see order visibility stretching to 2027. On risk, siting a multi-GW facility triggers 18-to-36-month permitting cycles for grid interconnection, water rights, and environmental review—far outpacing chip cadence. Vera Rubin's advanced-node dependence on Taiwan, China foundry capacity means any export-control tightening could fracture cluster delivery, compressing Oracle's pricing elasticity. Competitively, AMD's MI400 cannot displace the CUDA ecosystem near-term, but AWS Trainium and Google TPU's self-designed-plus-self-powered model erodes NVIDIA's pick-and-shovel monopoly. Oracle's capex lock-in strategy mirrors telco leverage ratios. 12-to-24-month call: the binding constraint shifts from chips to watts. Power cost becomes the primary variable in AI inference pricing. Compute-plus-power bundled pricing will define the next cloud paradigm, and nuclear-AI convergence will spawn a new energy-compute consortium model.