LessWrong GPU estimates
Summary
A detailed breakdown of expected GPU and compute availability across major tech companies like Microsoft, Meta, Google, Amazon, and XAI. Estimates are based on publicly available data and Nvidia revenue information.
Review
The document provides a nuanced exploration of AI computing infrastructure, focusing on GPU availability and compute capacity across leading technology companies. By analyzing Nvidia's revenue, chip production estimates, and company-specific purchases, the author constructs a detailed projection of computational resources for key AI players in 2024 and 2025. The methodology relies on multiple sources including earnings reports, industry estimates, and revenue breakdowns, acknowledging inherent uncertainties in the estimates. The analysis goes beyond simple chip counts, considering factors like custom chips (TPUs, Trainium), training compute requirements, and the evolving landscape of AI infrastructure. Key insights include significant compute expansion plans for companies like Microsoft, Google, and Meta, with emerging players like XAI also making substantial investments in AI computational capacity.
Key Points
- Microsoft, Meta, and Google expected to have 1-3.1 million H100 equivalent chips by end of 2024
- Blackwell chips offer approximately 2.2x training performance compared to H100s
- Total AI infrastructure spending is projected to grow significantly in 2024-2025