Dylan Patel 深度访谈

Dylan Patel breaks down the staggering AI infrastructure spending: Big Tech's combined $600B CapEx forecast, what's being spent now vs. future years, turbine deposits for 2028-2029, and data center construction timelines. Much of current CapEx is setup for future capacity.

OpenAI signed crazy deals and has way more compute access than Anthropic by end of year. Anthropic needs 5+ gigawatts but may struggle. Dario was conservative about compute spending; OpenAI went all-in. Anthropic may have to use lower-quality providers.

The geopolitical race: if AI capabilities arrive fast, the US and West win because of massive infrastructure investments. If timelines are longer, China can build a fully vertical, indigenized supply chain and eventually scale past the US.

Deep dive into the HBM memory crunch. Could accelerators use commodity DRAM instead? The answer is nuanced — bandwidth per wafer matters more than bits per wafer, and even for agentic tasks, no one wants a slow model. The economics of inference throughput favor speed.

Comprehensive analysis of the biggest bottlenecks to scaling AI compute: EUV tool shortages, memory constraints, power delivery, and data center construction timelines. The semiconductor supply chain is global and fragile.