Chamath Palihapitiya on the AI Compute War: Nvidia, AMD, Energy, and Memory

AMD's High-Stakes Bet

• 🎯 Lisa Su has made a "bet the farm" move with the MI450 chip, offering 10% of the company if the compute gets deployed.
• 💡 This strategy aims to help AMD re-enter the AI compute market and compete with Nvidia's significant lead.

Nvidia's Dominance and Ecosystem

• 📈 Nvidia has seen a 10x revenue growth since 2022, capturing nearly 100% of incremental AI data center revenues.
• 🧠 Their success is attributed to a comprehensive ecosystem including software, networking, and co-design, rather than just individual chips.
• ⚡ The unit of compute is now the entire data center, with performance per watt being a critical factor due to power constraints.

The New AI Metrics: Gigawatts and HBM

• 📊 AI deals are increasingly measured in gigawatts, as power has become the primary constrained resource and a unifying metric across diverse chip generations.
• 🔑 High Bandwidth Memory (HBM) is identified as another crucial constraining factor for the expansion of AI capabilities.

Controlling the AI Supply Chain

• 🔗 Companies that control energy (electrons) and the HBM supply (like SK Hynix and Samsung) are gaining significant power and leverage within the AI industry.
• 🚀 OpenAI's strategic deal in Korea to secure HBM forward capacity gives them control over allocation, mirroring historical power dynamics.

Market Disbelief and Future Outlook

• ⚠️ Wall Street estimates for Nvidia's growth post-2027 are notably lower than the industry's projected 100-gigawatt buildout, indicating a significant disconnect.
• 🔮 Predicting future compute demand remains challenging, as it depends heavily on the development of new, yet-to-be-invented AI applications beyond current chatbots.