The debate now may no longer be about whether LLM is useful. It can indeed solve a certain type of problem. What is really more like a foam is the "reinforced concrete+debt leverage" around AI. Bloomberg reported that Oracle's Stargate data center project for OpenAI may be delayed from 2027 to 2028 due to labor and material shortages. Oracle subsequently denied the delay.  What's interesting is not whether it will be delayed or not, but why such news can immediately make the market nervous: because behind all the advanced construction, there is a gamble on whether the utilization rate and computing power unit price can cover depreciation, electricity bills, and capital costs when the park is really electrified. The official statement of Stargate is to invest $500 billion in the next four years (with an initial investment of $100 billion), and in September 2025, OpenAI will expand it into a platform with a planned capacity of nearly 7GW and a three-year investment of over $400 billion.  Add another more realistic tension: NVIDIA's roadmap has moved towards a rhythm of 'one update per year'.  The annual change of chips does not equal the annual replacement of data centers, but it will accelerate the "pressure to catch up with new products", making it easier for computing power prices to stratify and decline. Once the unit price drops, writing the high price into Excel's payback period today will become very fragile. Therefore, I prefer to regard the "AI foam" as a specific gamble: it is not useless to bet on AI, but on the uncertain calculation price in 2028 and whether it can pay for the locked advance bill in 2025.