Fin, a semiconductor expert, holds all high cognitive views on AI. AI summary: -* * AI foam property** -Core view: different from the Internet foam: AI is an application driven infrastructure, and infrastructure is in short supply rather than surplus; Foam may be on the App end (demand fulfillment is slow). -Supporting arguments: 97% of optical fibers are idle due to the decoupling of Internet infrastructure; The CRWV backlog in AI has increased from 30B to 55B, with demand transmitted from App → Cloud → DC → Chip. -Risk warning: App end foam comes from slow AI/Agent iteration; When the penetration rate is close to 50%, the growth rate slows down, which may trigger a trend of single cutting (similar to the Cisco foam). -GPU depreciation and lifespan** -Core viewpoint: Depreciation is not a big issue: training GPU lifespan is 2-3 years (AFR<6%), inference is 5-6 years; Overall, it is reasonable to take 5-6 years. -Supporting evidence: Meta Llama3 report: annualized AFR of 9%; CRWV financial report: Renewal price of old H100 only drops by 5%, A100 sold out. -Risk warning: If depreciated for 6 years but with a short actual lifespan, it may lead to inflated profits (grey rhino); Technological iteration leads to TCO disadvantage, but short-term supply exceeds demand. -Infrastructure Investment and Capex** -Core viewpoint: The company is aggressively shifting towards Capex (Meta 38%, MSFT 33%), squeezing Opex and leading to layoffs; Semiconductor profits exceed the Internet and become the new normal. -Supporting evidence: Amazon cuts 30000 to save 60B on GPU purchases, leading to an increase in cloud growth from 18% to 24%; Nvidia investment partners drive Capex. -Risk Warning: Power/Grid Connection Bottleneck (Mining to AI such as CORZ); Debt risk (such as ORCL); The foam burst on the app side rather than infrastructure. -* * Company Strategy** -Core viewpoint: OpenAI balance: 10GW NV, 10GW ASIC, 6GW AMD; AMD's aggressive PIM technology; Meta all in, but often with the wrong rhythm. -Supporting evidence: AMD ROCm has made progress, but the software stack is slow; Oracle may enter the ASIC market (rumor). -Risk Warning: Zuck is lucky (Reels GPU to LLM), but his style is unpredictable (Libra, VR); AMD needs mature ASICs to compete for inference market share. -Employment and Social Impact** -Core viewpoint: AI efficiency improvement is limited (15-20% for large companies); The employment pressure comes more from Capex squeeze rather than pure efficiency improvement. -Supporting evidence: The Agent operates aggressively but aims for only 20% efficiency; SDE buys Nvidia to hedge against being squeezed out of the value chain by GPUs. -Risk Warning: Structural Unemployment Wave (Non Large Scale); AI deflationary (suppressing inflation); End side AI requires 7-10 years of evolution. -* * Technological Trends** -Core viewpoint: Early storage cycle (HDD/SSD shortage); The cost of AI inference remains unchanged (experience first, such as mobile phone power consumption). -Supporting evidence: WDC/Hynix financial report: increased inference ratio; The fastest growth stage for penetration rate is between 10% and 50%. -Risk warning: Token consumption becomes a civilized symbol; Human+AI>AI (open questions require human simplification).