NVIDIA

Type: Company (Semiconductor / AI Compute)

NVIDIA is the dominant force in AI compute infrastructure, executing a strategy shift from selling discrete GPUs to selling fully integrated rack-scale systems. The Vera Rubin platform, announced at CES 2026, exemplifies this "extreme co-design" approach: six co-designed chips (GPU, CPU, NVLink switch, ConnectX NIC, BlueField DPU, Spectrum Ethernet switch) engineered as a single rack-level product.

The company maintains an annual architecture cadence — Hopper (2022), Blackwell (2024), Vera Rubin (2026) — that outpaces the multi-year development cycles of custom ASIC competitors. Each generation delivers 3-5x improvements across key metrics (FLOPS, memory bandwidth, interconnect speed), compounding NVIDIA's lead in cutting-edge training and inference.

NVIDIA's moat is evolving. While CUDA software lock-in remains significant, the deeper competitive advantage is now system-level integration. No custom ASIC vendor matches the full-stack co-design spanning compute, networking, security, cooling, and power delivery. SemiAnalysis projects NVIDIA's inference market share will fall from 90%+ to 20-30% by 2028, but the company captures the premium segment (training + cutting-edge inference) where system integration and generality matter most.

Key Contributions

• Vera Rubin platform: 50 PFLOPS FP4, 288GB HBM4, 336B transistors per GPU (NVIDIA Vera Rubin)
• NVL72 rack: 72 GPUs, 260 TB/s aggregate, cableless modular trays (NVIDIA Vera Rubin)
• Rack-as-product co-design strategy with 6 chip types (SemiAnalysis)
• Annual architecture cadence vs multi-year ASIC cycles (SemiAnalysis)

Mentioned In

• Rack-Scale AI Compute — Primary architect of rack-as-product paradigm
• HBM4 Memory Architecture — First to deploy HBM4 in production GPU
• Custom Silicon vs GPU — Central player in ASIC vs GPU competition
• Silicon Photonics — Spectrum-6 with co-packaged optics