Adding a piece of data, computing power growth will outpace AI-related revenue growth in the next six months. The computing power shortage issue has significantly improved since October, leading to a price war among top AI companies, as they have already pre-ordered too much computing power. Global AI-related revenue is projected to grow from approximately $110 billion in 2025 to approximately $300 billion in 2026 and $550 billion in 2027. Annual increases are $190 billion and $250 billion respectively, representing growth rates of 170% and 83%, although the growth rate is expected to slow. New installations are projected at 32GW and 49GW in 2026 and 2027 respectively. At $10 billion per GW, the annualized cost increase is $320 billion and $490 billion respectively. Shipments are projected to increase by 54% and 27% respectively, but due to increased power consumption per GPU, the equivalent GW growth rate is only 67% and 44%. Although Anthropic's revenue is still growing rapidly, with ARR revenue multiplying several times in the first half of the year, it can still support the consumption of computing cards. However, starting in the second half of the year, the growth trend in July, August, and September basically flattened out, and the computing power shortage problem was significantly alleviated. But what's worrying is that card shipments are still growing at such a high speed, while revenue is not keeping up, and computing power efficiency is also improving. These three factors combined seem to indicate that a turning point in computing power is finally approaching. Inefficient computing cards will be idle, while new, efficient computing cards will still be in short supply, somewhat like the upgrade cycle of Bitcoin mining machines. If computing power and land are scarce, and the latest machines are installed as much as possible, then by the second half of 2027, with the promised installations gradually being completed, there will be idle computing power in US computing centers. Extrapolating from the data from the above dimensions, this is consistent with business logic. On the other hand, a detailed calculation of Anthropic's computing power growth, including how much computing power they pre-ordered, and how its growth rate compares to revenue growth, reveals that revenue growth is definitely slower than computing power growth, eventually reaching a balance. This is because Anthropic has signed computing power commitments; too much idle computing power will lead to consideration of price reductions. Anthropic's recent adjustments, including a slight relaxation of usage restrictions, and the decentralization of advanced AI, are beginning to threaten the cost-effectiveness of domestic models. Although electricity prices are cheaper in China, the lack of top-tier chips means that the unit token generation cost is likely still higher than in the US.