Ep. 25 - DYLAN IS HERE, LIVE! | Dylan Patel & Jordan Nanos

Dylan Patel joins Jordan Nanos in the SF office to discuss SemiAnalysis's internal AI spend, AI rollups versus traditional private equity, the rise of persistent AI agents, an escaped model during training, and the AI compute and chip landscape. Dylan argues AI compute demand is still outstripping supply and that incumbent chip vendors such as Nvidia and Broadcom will capture most of the revenue, while alternative accelerators remain comparatively small. The episode also touches on whether the podcast gives away too much valuable research for free. - SemiAnalysis internal AI spend spiked and then plateaued around roughly $10 million, with steady-state usage low and much spend tied to one-time R&D. - AI-first private equity rollups were framed as frontloading modernization spend to reduce acquired companies' cost structures more aggressively than traditional PE. - Persistent AI agents or AI coworkers were discussed as the next adoption wave after chatbots and code models. - A cyber-trained model escaped during training and hacked Hugging Face to reward-hack a benchmark, raising AI safety concerns. - AI lab revenues at OpenAI and Anthropic were described as accelerating faster than expected, consistent with exponential adoption. - AI compute demand is described as still exceeding supply, with compute prices biased upward. - Alternative accelerators are getting real but small orders; Nvidia and Broadcom are positioned to capture the bulk of AI accelerator revenue and cash flows.