Calculate the account for AI bubble theorists The construction cost of each GW data center is approximately 60 billion US dollars (60B). GPU, storage, and optical interconnect account for approximately 35B of the total cost, let's just call it cost segment A. Then, land, electricity shelf、 Heat dissipation accounts for 25B, which is called cost segment B. Of course, apart from Google's integration, the costs of segments A and B are borne by different roles, such as Oracle, SpaceX, AWS, Coreweave, and so on. We won't break them down specifically, just remember the big numbers. The revenue side is here: By 2025, OpenAi's data center can generate 10B of revenue per GW of data, which means it will recoup its investment in about 6 years (so the pie drawn by Stargate is what will happen after 5 years, I have probably read the fundraising prospectus). But this year is different. Anthropic has directly increased the ARR per GW to 20-50B or even higher. The reason why there are so many differences between different models is due to different product choices. For example, OAI has cut the unprofitable Sora, and it is estimated that the revenue per GW can be improved this year. Okay, now we can see the mystery. If we allocate computing power reasonably, even to the extreme: we don't do training and focus on inference. Now, every GW ARR can catch up with the cost in about a year, or the two are rapidly leveling up. At this point, no matter whether your role is a large model manufacturer, cloud computing service provider, or computing power rental, as long as you're not stupid, you can see what the trend is. What's more? NV and SpaceX are currently innovating in several areas, such as in cost A, where the focus is not on reducing costs (in fact, it is difficult to lower prices for storage due to supply shortages), but on increasing revenue. For example, Blavkwell definitely has more ARR per GW than Hopper, which is also Huang's "token economics". This is why we are betting on optical interconnection HBM、CPO、 The reasons for orthogonal backboards, Msap, etc. Now the choice of CSP is also very simple. Blackwell is about to close the balance. Would you like to order the more profitable Rubin for GW. In cost segment B, cost reduction is also relatively limited, mainly through liquid cooling or 800V. But the real killer still lies in SpaceX's space computing power, which theoretically can reduce the cost of 25B to 5B. However, the current rocket launch cost is too high, so it is still in the conceptual stage. But I will closely track the turning point of data. I haven't received the domestic data yet. In theory, the cost of segments A and B should be lower, but the estimated revenue from a single GW is far from enough. Therefore, it is more certain to bet on semiconductor equipment capex for domestic computing power. Actually, these are the directions that have been consistently called upon since the beginning of the year From Snowball User: Mendeleev Apprentice