Roadmap

Blackwell Ultra keeps the rack shape and changes what is in the trays: 288 GBHBM3e per Blackwell Ultra GPUSources disagree: 279 GB vs 288 GB SKU and ECC accounting. Note that NVIDIA’s own MLPerf Training v5.1 blog quotes 279 GB, so this is not simply a case of secondary sources getting it wrong. NVIDIA · NVIDIA Technical Blog of HBM3e per GPU against 192 GB, ~1,400 WPower per Blackwell Ultra GPU NVIDIA per GPU against roughly 1,200 W, and 1.5×GB300 FP4 uplift Relative to GB200. NVIDIA the FP4 throughput. Scale-out moves from ConnectX-7 at 400ConnectX-7 SuperNIC NVIDIA to ConnectX-8 at 800ConnectX-8 SuperNIC PCIe Gen6; shipped on GB300. NVIDIA .^{NVIDIA,NVIDIA Technical Blog}

The capacity increase is the consequential one. Fifty percent more HBM per GPU is fifty percent more KV cache and batch headroom on the decode pool, which is where memory-bound serving actually converts capacity into throughput. NVIDIA reports 4.2× vs Hopper, 1.9× vs GB200GB300 training speedup At 512-GPU scale. NVIDIA Technical Blog on MLPerf Training at 512-GPU scale.^{NVIDIA Technical Blog}

The naming deserves a note, because it breaks the pattern the previous two generations set. "NVL72" counted 72 GPU packages. "NVL144" counts 144 dies across 72 packages — the same package count as today, with the die-per-package doubling now surfaced in the product name. It is not a doubling of the NVLink domain's endpoint count, and reading it as one will produce wrong capacity planning.

Rubin Ultra NVL576 is projected for 2027 and Feynman beyond that. These come from keynote slides and supply-chain reporting rather than from datasheets, and this site does not publish specifications for them.

~120 kWRack powerSources disagree: 120 kW nominal · 125–135 kW operating (Supermicro) · 132 kW fully loaded (Schneider Electric) Supermicro’s datasheet states an operating power of 125–135 kW and 132 kW of installed power-shelf capacity. Steven Carlini, writing for Schneider Electric: "When fully loaded into a rack, the latest NVIDIA-based GPU servers require 132 kW of power." The commonly quoted ~120 kW is the nominal design figure, not a measured ceiling. Supermicro · Steven Carlini, Schneider Electric — Forbes Technology Council · ServeTheHome today, against a worldwide mean rack density of 7.6 kWWorldwide mean rack density Up from 6.8 kW the prior year; 8.4 kW if racks above 30 kW are excluded. An NVL72 is roughly sixteen average racks in one footprint. Uptime Institute .^{Uptime Institute} Schneider Electric's forward-looking figure for the next generation is 240 kWNext-generation rack power Announced, not measured Schneider Electric, forward-looking: "The next generation, expected in under a year, will require 240 kW per rack." Steven Carlini, Schneider Electric — Forbes Technology Council per rack.^{Steven Carlini, Schneider Electric — Forbes Technology Council} At that density the facility, not the silicon, becomes the thing being designed: power delivery, coolant volume, floor loading, and grid interaction all move onto the critical path.

Everything this site describes about NVL72 — the busbar, the copper spine, the mandatory liquid loop, the 1.36-tonne frame — is a set of answers to that curve at 120 kW. Each subsequent generation is the same question asked again at a larger number.