Jensen Huang Declares: Let There Be Light

🔑 Enhanced Key Takeaways

• •At CES 2026, Nvidia announced the Vera Rubin platform, a rack-scale AI system integrating Rubin GPUs, Vera CPUs, NVLink 6, BlueField 4 DPUs, and Spectrum 6 switches, shifting from individual GPUs to complete AI factories[1].
• •Nvidia is skipping new consumer GPU releases in 2026, including RTX 50 Super, due to memory shortages and prioritization of AI data center chips[2].
• •The company participated in nearly 70 VC deals in 2025, including an acquihire of Enfabrica to bolster inference capabilities via BlueField-4 DPU and Dynamo software for up to 5x inference performance gains[3][4].

🛠️ Technical Deep Dive

• •Vera Rubin platform: Rack-scale system with Rubin GPUs (two per system), Vera CPUs, NVLink 6 interconnects, BlueField 4 DPUs for inference offload, Spectrum 6 Ethernet switches, and Elastic AI memory fabric (Emphasis) to address memory walls in AI inference[1][3].
• •BlueField-4 DPU integrates with Nvidia Dynamo software and partners like IBM for KV cache expansion, enabling up to 5x inference boost and efficient token generation[3].
• •Rubin CPX variant optimized for million-token context processing in reasoning tasks[4].

🔮 Future ImplicationsAI analysis grounded in cited sources

⏳ Timeline