NVIDIA’s Ian Buck Says CoreWeave’s V100 GPUs Still Run Nearly A Decade Later, Even As Vera Rubin NVL72 Goes Live With a 4.8x Token Generation Speedup - Wccftech

V100s "still running" at CoreWeave is not a hardware durability story—it is a stranded-asset warning. Vera Rubin NVL72's 4.8x token-generation speedup delivers near-5x output per compute dollar, shattering the depreciation schedules neoclouds baked into their capex models. The 4.8x leap is not a single-chip event: HBM3e bandwidth, NVLink interconnect, and 800V DC power delivery must scale in lockstep. Data centers designed around 250W-per-GPU budgets face a 5-to-10x power-density wall with Rubin-class systems—this is a physical rebuild, not a socket swap. On compliance, V100 longevity paradoxically deepens the structural damage of export controls: large installed bases in restricted markets become permanently frozen at an older performance tier, cementing a bifurcated global AI stack. Competitively, AMD's MI350 and Intel's Gaudi 3 cannot close a 4.8x generational gap within one product cycle. Hyperscaler custom silicon—TPU v6, Trainium 2—remains the only credible path to escape NVIDIA's generational lock-in. 12-to-24-month outlook: expect CoreWeave and peer neoclouds to face economic obsolescence well before physical end-of-life. Legacy GPUs will retreat into edge inference and niche vertical workloads, fragmenting into a secondary compute market that no longer competes on headline benchmarks.