Nvidia Unveils 120B-parameter Nemotron Super VL - Dataconomy

Nvidia's 120B-parameter multimodal release is a vertical-integration play mirroring ARM's 2010s trajectory from ISA to SoC to IP licensing—compressed into a single product cycle. The strategic signal outweighs the parameter count: by owning the model layer, Nvidia converts its GPU moat into software lock-in, forcing inference silicon to optimize for bandwidth-precision co-design rather than raw FLOPS. HBM3E interconnect and NVLink topology become the binding constraints for the next architecture generation. On compliance, the EU AI Act's transparency mandates for multimodal systems, layered atop US export restrictions, fracture Nvidia's "one global software stack" thesis into geographically partitioned deployments. The model's effective radius is now a geopolitical variable, not an engineering one. Competitively, AMD's MI300X paired with Meta's open-weight Llama 4 creates a credible hardware-plus-model bundle that undercuts Nvidia's software premium. Microsoft and Google sit in an awkward dual role—both customers and direct model competitors—making their procurement leverage a structural threat. The 18-month watch item is not parameter scaling but the distillation-quantization-edge-deployment pipeline. Whoever owns inference-side energy efficiency owns pricing power. The model layer is the new GPU.