Keep pulling the thread on Dylan Patel.
Mark Zuckerberg is building a two-gigawatt data center in Louisiana for Meta.
Amazon, Google, and Microsoft are building multi-gigawatt data centers and spending billions of dollars on fiber to connect them.
Excluding Google's internal workloads, NVIDIA chips run over 98% of global AI workloads.
When including Google's internal workloads, NVIDIA chips run approximately 70% of global AI workloads.
NVIDIA's dominance is attributed to its superior capabilities in software, hardware, and networking, a combination other semiconductor companies lack.
NVIDIA is moving to an annual product release cadence, with Blackwell, Blackwell Ultra, Rubin, and Rubin Ultra in its roadmap.
NVIDIA's Blackwell architecture provides a 10 to 15 times performance improvement for inference on very large language models.
NVIDIA has decided to lower its margins on the Blackwell platform to better compete with custom chips from Amazon, Google, and AMD.
NVIDIA aims to improve performance per total cost of ownership (TCO) by approximately 5x every year, a significant acceleration from Moore's Law.
Microsoft is currently constrained by power and data center availability, not by the supply of AI chips.
The cost of a major AI training run is increasing exponentially, with the next generation potentially costing between $3 billion and $30 billion.
x.AI plans to build a 200,000 to 300,000 GPU cluster.