Keep pulling the thread on Dylan Patel.
To successfully compete with NVIDIA, a new entrant in the AI chip market must offer a product that is at least 5 times better in performance or efficiency.
Dylan Patel predicts OpenAI will monetize its free user base by using a router to direct high-value, transactional queries to advanced agentic models and taking a commission on resulting purchases.
OpenAI and Anthropic are collectively receiving 30% of the AI chips being produced this year from their cloud partners, including Microsoft, CoreWeave, Oracle, Google, and Amazon.
Approximately one-third of all AI chips are being utilized for advertising-related workloads by companies such as ByteDance and Meta.
Anthropic's Claude Code surpassed the annual recurring revenue (ARR) of GitHub Copilot within just three months of its launch.
Hyperscalers are projected to increase their capital expenditures by 20% to 30% next year.
The biggest threat to NVIDIA's market dominance comes from hyperscalers like Google, Amazon, and Meta massively increasing orders for their in-house custom silicon.
Google is currently producing millions of its custom TPU chips, which are operating at 100% utilization.
Google is internally discussing the possibility of selling its TPU chips directly to external customers, a move that would require a significant cultural and organizational restructuring.
Chinese technology companies like Alibaba are renting high-performance GPUs outside of China because they offer better performance per dollar than the restricted chips available domestically.
The build-out of AI data center capacity in the United States is heavily constrained by the availability of electrical power.
Google has a significant number of manufactured TPU chips that are currently sitting idle because the data centers intended for them are not yet powered.