Still bullish on AI: Quoting insights from Tianfeng Overseas Technology 1. GPT-6 exceeds expectations: It all comes down to whether the model progresses and iterates. Looking back at June and July, the surface reason was deleveraging, but the underlying reason was that the model was in a transitional phase without significant progress. This time, Astra's Computer use essentially unlocks new scenarios. The software operated by the model eliminates the need to rewrite code to call APIs. The significance lies in entering the third phase of AI—replacing white-collar jobs. This TAM is $4T (global public companies' SG&A expenses are $10T, with G&A expenses at $4T). 2. Training volume far exceeds expectations: From two 10,000-card clusters -> 100,000-card clusters -> 400,000-card clusters. Ultimately, training is the biggest variable exceeding expectations next year. This industry is still led by scientists, and it may end up maintaining over 50% training levels for years. ARR is just a financing tool for scientists to borrow money. AI For Science is brewing (essentially unlocking new growth and revenue logic, ranging from anti-aging drugs to hand-drawn PCB circuit boards—things that didn’t exist before but do now). Although there are differing opinions on distillation prevention, the fact is that it’s becoming increasingly difficult. The gap between domestic and international models is widening. The 1GW revenue ceiling is being raised. OAI and Anthropic are holding big moves in reserve, with expectations that the 1GW revenue ceiling will continue to rise. Returning to the narrative from the first half of the year: deleveraging is complete, the supply chain hasn’t found any downward data, B300 prices are still at record highs, and enterprise adoption rates are climbing steadily.