"Treat inference as hard thinking, not cheap reading When the AI industry assumed inference chips would be 'little tiny chips' because training was the hard part, Huang pushed back with a simple framing: 'inference is thinking and thinking is hard — thinking is way harder than reading,' while 'pre-training is just memorization and generalization.' This reframe had immediate strategic consequences — Nvidia invested in large-scale inference infrastructure while competitors built for a commodity market that did not materialize. Anyone evaluating which phase of an AI or knowledge-work pipeline deserves the most resources should ask whether they are underestimating the compute cost of actual reasoning versus simple pattern retrieval. Source: Jensen Huang, Lex Fridman Podcast, Jensen Huang: Nvidia, AI, Robots, and the Future of Computing"