Nvidia Confident in $1T Revenue from AI Chips

🔑 Enhanced Key Takeaways

• •Rubin platform entered full production following Jensen Huang's announcement at CES 2026[1][4].
• •Rubin GPU features 336 billion transistors on TSMC 3nm process, 1.6x more than Blackwell's 208 billion[1][3][4].
• •Rubin delivers 50 PFLOPS FP4 inference per GPU, 2.5x Blackwell's 20 PFLOPS, with 35 PFLOPS for training[1][3][6].
• •Rubin Ultra planned for 2027 with ~500B transistors, 384GB HBM4E, and 600 kW rack power[1][5].

🛠️ Technical Deep Dive

• •Rubin GPU: 336B transistors, TSMC N3 process, 288GB HBM4 across 8 stacks (2,048-bit interface per stack), 22 TB/s memory bandwidth (2.8x Blackwell)[1][3][4].
• •224 Streaming Multiprocessors with 6th-gen Tensor Cores supporting FP4/FP6/FP8/FP16/BF16/TF32/FP32/FP64; 3rd-gen Transformer Engine with adaptive compression for NVFP4[3].
• •NVLink 6 at 3.6 TB/s bidirectional per GPU (50% over NVLink 5); TDP 1,800-2,300W requiring liquid cooling[1][3].
• •Rack-scale: Rubin NVL144 offers 3.6 EFLOPS FP4 dense compute (vs. Blackwell Ultra B300 NVL72 at 1.1 EFLOPS); NVLink7 up to 260 TB/s total[5][7].

🔮 Future ImplicationsAI analysis grounded in cited sources

⏳ Timeline