Nvidia CEO Jensen Huang on Blackwell, Vera Rubin, and China Market

Blackwell GPU Sales and Supply Chain

• 🚀 Blackwell GPUs are experiencing "off the charts" sales, with Nvidia's cloud offerings currently sold out.
• 💡 Despite high demand, Nvidia has planned its supply chain incredibly well, working with partners like TSMC, SK Hynix, Micron, and Samsung to ensure sufficient supply.
• ✅ The company is confident in its ability to meet demand for both Blackwell and the upcoming Vera Rubin platform.

Vera Rubin Platform Rollout

• 🧠 Approximately 20,000 people are working around the clock to bring the Vera Rubin platform from silicon to systems and software.
• 🗓️ Vera Rubin is on track for delivery around Q3 of next year, continuing Nvidia's annual product cycle.
• 🛠️ The rack scale architecture, including the MVLink 72, is revolutionary and will seamlessly transition from Grace Blackwell to Vera Rubin, ensuring a smooth ramp-up.

China Market Strategy

• ⚠️ Nvidia's forecast for the China market is currently zero, despite its importance.
• 🤝 The company aims to re-engage the Chinese market with excellent products and compete globally, emphasizing the benefits for both the US and China.
• 📈 Nvidia is committed to engaging with both the US and Chinese governments to allow participation in the open market.

Energy and Infrastructure Constraints

• ⚡ Nvidia's rapid growth (e.g., $10 billion quarter-over-quarter) presents challenges across the board, including energy availability.
• 🌐 The company leverages its extensive network of cloud service providers, OEMs, and partners worldwide to find available power sources.
• 💡 Nvidia's architecture is designed to run every model and accelerate all phases of AI, making its GPUs highly versatile and valuable for extended periods.

Chip Depreciation and Lifespan

• ⏳ While A100 GPUs shipped six years ago are still in use, Nvidia's CUDA software continuously updates and adds value to older hardware.
• 🧩 The versatility of Nvidia's GPUs across different AI phases (pre-training, post-training, inference) and diverse model types prevents their value from falling off a cliff, unlike more specialized accelerators.
• 💰 Nvidia's contribution to a 1-gigawatt data center is estimated at $35 billion for Vera Rubin systems, with a full data center costing around $50-55 billion.