The market has been scared by the rise in memory prices, which has led to Apple's price increase. They are starting to worry about whether cloud vendors can sustain their spending and whether AI infrastructure is being excessively built? This ghost story is not told for the first time, but it is necessary to analyze it again: The core issue is not price increases, but whether price increases will affect demand. So you can ask two questions: Will the price increase of apples affect your purchase? Most people still buy when they want to, right? When you buy a Mac book for $1499, will you stop buying it because of $1699? The more affected are the mid to low end Android devices, and I don't expect the sales of high-end Apple devices to be significantly affected. 2. Will the cost of AI tokens, that is, the ARR of large models, decrease? From the inability to grab GLM, which is constantly raising prices, to the hesitant Mythos/Fable, we cannot see a trend of decreasing demand. It is worth noting that some large companies have started to limit and regulate the unlimited use of tokens, which can only affect the extreme usage of individuals, and the total amount will definitely increase. So I don't think there's a problem with the long-term logic. 3. There are still a few scary pictures today, showing a huge increase in capex and a sharp decrease in cash flow for a large company's hyperscaler. But this is just scary at first glance. In fact, it depends on the ratio of cash flow to capex. Let's take a look at the ratio of OCF operating cash flow to Capex, as shown in Figure 1. Except for Oracle and Amazon, which are relatively dangerous, the other major companies are still relatively stable. Companies like Apple, which have very conservative investments, even have great investment space. From the perspective of debt/OCF ratio, as shown in Figure 2, besides Oracle and Amazon, there is still room for upward growth in the borrowing scale of major companies. 4. The cash flow situation of large factories will improve in 27-28. Firstly, the expansion of memory production capacity will begin to increase, which will inevitably reduce the upward slope of capital expenditures and alleviate the pressure on capital expenditures; The second is that the ROI of large companies is expected to improve, and both their Agent and AI products, as well as the profitability of cloud services, will see returns from their previous investments. Another easily overlooked point is that the cost advantage of self-developed chips (TPU, Trainium, MTIA) by major manufacturers will gradually begin to emerge in 2026-2027. Google estimates that the cost per token of its TPU v5 is 30-50% lower than NVIDIA H100, and Amazon's Trainium 2 also claims to offer similar cost advantages. As the proportion of self-developed chips in AI workloads increases, the unit computing cost will continue to decrease. Conclusion: The logic of investing in upstream (storage, semiconductor industry chain) remains unchanged this year, and starting next year, we will consider returning our positions to strong companies in Mag7