NVIDIA trims Vera Rubin HBM4 and NVLink bandwidth as supply commitments hit $279B

TechRadar reports that NVIDIA had to dial back key Vera Rubin specifications:

- HBM4 memory bandwidth drops from 22 TB/s to 19.2 TB/s per GPU.

- NVLink interconnect drops from 3.6 TB/s to 3 TB/s.

Those are meaningful cuts for a platform marketed on bandwidth leadership. Memory bandwidth often bottlenecks large-model inference more than raw compute does.

The likely cause is supply-chain reality. HBM4 at the originally targeted speeds is hard to produce in volume, and NVIDIA appears to be trading peak specs for shippable quantity. Meanwhile, supply commitments, NVIDIA's purchase obligations to suppliers, have more than doubled from $119B to $279B, per TokenPost. That is a massive bet on continued demand. The company guides $108B in revenue for fiscal Q3 2027, reporting November 17.

The consumer side shows the same strain: the RTX 5080 has left NVIDIA's store, with street prices at 2.2x MSRP. On the software side, NVIDIA kept shipping:

- Nemotron 3.5 Transcribe: open weights, 40 languages, 5.33% English WER.

- NemoClaw for building specialized agents on a stack the developer controls.

- A validated Kubernetes runtime for Dynamo distributed inference.

Competitively, AMD and custom silicon (AWS says its chips saved customers more than $2B last quarter) gain talking points when NVIDIA cuts specs. Investors already worry about circular, debt-financed AI capex. Watch the November 17 call for Rubin ramp timing, HBM supplier mix (Samsung, SK Hynix, Micron) and any comment on the DOJ's Hugging Face probe.