Late Night Good Article Sharing: How Many Data Centers Do Humans Need At present, the expected completion of all data centers in North America by 2026 is 20GW, and even considering delays, it is expected to reach around 30GW next year. This includes CSP self built data centers such as Meta, Xai, AWS, Google, as well as new cloud and Oracle. Most of the popular Crispy fried chicken Anthropic is rented from others. So for the second question, which is the title of this article, I have synthesized the latest statements from the most cutting-edge AI experts in Silicon Valley, such as Dylan, Dario, Gavin Baker, Sam, Mu Jie, and others (be sure to track the latest, as AI's views will become outdated in three months). The general conclusion is that in the medium term (before 2030), Anthropic and Oai alone will need 100GW of data centers, and considering other CSPs such as G, even if they fall behind others, they will still need at least 50GW. In the medium to long term (the next 5-10 years), AI's reshaping of global GDP will be 5-15%. Let me translate it for you. 100GW means that Anthropic and Oai need to reach a revenue of $2 trillion by 2030. Currently, the consensus in North America for 27 years is that the ARR of these two companies is $500 billion, so $2 trillion in 30 years is not an exaggeration. The second half of the conclusion is even more interesting. Currently, the global GDP is $100 trillion, and 10% is $10 trillion. Even without considering inflation and growth, it is equivalent to a 500GW data center. Of course, experts calculate crudely. For example, Dylan directly stated that a 1TW or 1000GW data center is needed. Note that there is a difference here. Anthropic, a coding product, belongs to AI's native revenue, but when we comment on the value of AI data centers, we cannot just look at this. Taking Meta, which is currently at the forefront, as an example, although the big model has collapsed, we take the Q1 financial report of 26 as an example. Its advertising revenue was 55 billion US dollars, a year-on-year increase of 33%. The financial report clearly expresses that this is Ai driven. I have broken down the natural growth of Q1 in recent years, and the best year is just over 20%. This means that Meta has achieved more accurate push and price increase of AI advertising through multimodal big models to analyze video and image data. This alone accounts for at least 10%, which is 5 billion US dollars, and the total for the whole year is 20-30 billion US dollars. And this part of the value will not be included in AI's native income, so it is naturally not as eye-catching as Anthropic (which means that the meaning of AI for GDP reshaping cannot be solely based on AI's native income). Google is no exception, but this is also the reason why the G and M models fall behind, because their business form determines that multimodality must be pursued, so they do not strategically focus on the most profitable coding, so Wall Street analysts show no mercy. But I think there is a gap in expectations in this part, and sometimes the capital market is still short-sighted. So we can come to a preliminary conclusion: based on the benchmark of 30GW to be completed in 27 years, even if the most conservative calculation is about 10-20% of the progress bar for data centers, optimistically speaking, it will be less than 5% next year. Considering that all the data center addresses for Beimei have been selected in 27 years, and with more physical restrictions in 28 years, it is likely to be postponed, so computing power will be in short supply for a long time. Of course, some friends may wonder if these predictions are all fooling around and if there will be a day when demand suddenly disappears. Good question, actually if you want to say whether the final data center completion scale will be 100GW or 1000GW, I don't know (maybe it's a certain number in the middle), but I will give two core indicators to judge whether Ai needs to retire or retire in stages: 1. Is the ability of cutting-edge models still improving. 2. Will the cost of AI hardware, or more precisely, a single token, continue to decrease exponentially. As long as one of these two flywheels is rotating, capex will not stop, and the rest are all interference terms. 2026 is the best year, and both flywheels are spinning rapidly. Firstly, this is the first time in thousands of years that humans have commodified intelligence, and the pursuit of intelligence by humans is unlimited. In theory, all the idols we worship in various fields are actually more intelligent and conscious than us. The current cutting-edge models take away 80% of the revenue, which is the reason behind it. Imagine if you were a factory owner, could you refuse Jack Welch to manage it for you? If you are conducting physics experiments, can you refuse to replace 10 Einstein with 10 dull students? I suggest everyone really use agents like Fable to see if they become addicted to paying. In addition, there are always people who like to use stock analysis to look at new things. For example, if Ai Lab earns one more yuan, it means that other companies will have one less yuan in their pockets: token consumption is unsustainable! In fact, new things often bring about inflation and an increase in total social wealth. For example, if you compare your income from 20 years ago to housing prices, there is almost no room for discussion. However, how to distribute the huge amount of social wealth that has been added this time is indeed another global problem (which is also the root cause of the current K-type). The solution I see is that both China and the United States are encouraging AI leaders to move from the primary market to the secondary market (don't think that they are just trying to make money, there are always such superficial arguments, is it possible that they are trying to split money), and Lao Huang has been investing in new cloud and open source models recently to avoid excessive monopoly. In fact, the bond and emotion between NV and China's open source big models are better than those in the United States. It is easier to explain 2 in this way. First, let me ask you a question: If the cost of electricity can be reduced to one tenth or even lower than the current level, what changes will the world have? Perhaps the imagination of ordinary people is nothing but that the air conditioning in my home can be turned on 24 hours a day in summer or the lights in the office building don't need to be turned off at night. In fact, the truly revolutionary changes are in the industrial sector. For example, the biggest cost of converting seawater into freshwater is the cost of electricity, which means that seawater around the world can be converted into freshwater. This means that it can quickly flow into deserts and become oases, and the world's ecosystems can support the current population to multiply several times. There are also applications in biosynthesis, electrolytic aluminum, steel, mineral extraction, and so on. So if the cost of electricity really drops so much, guess if Govs in various countries will crazily increase capex to build power infrastructure. Ai is the same, the real impact is still on the B-end. We say that even if it stays at the level of Fable, as long as AI Infra engineers around the world continue to reduce costs, capex will improve exponentially. Say it again, everything else is interference! Just these two core indicators. Think back to a certain period in 24-25 when there were reports of pre training crashing into the wall and Blackwell's shipment not going smoothly, which was also the time when XYS was hit by a market value of 50 billion and 5 times PE. In terms of industry, apart from North America continuing to make money, there is a significant expectation that domestically produced semiconductor equipment, chips, components, and rental calculations will be profitable. Just by conducting research, you will know that semiconductor equipment is now a completely seller's market, especially in advanced packaging. The H-series will make great efforts in the next two years. Although I experienced the largest loss in history last Thursday, I am extremely optimistic about semiconductor equipment this year, especially various testing equipment. In terms of the North American region, upstream optical chips and equipment are still the most promising for light, especially silicon optical cw. From the current perspective, NPO has the strongest sharpness (considering that the H-series in the domestic region will also be upgraded), and the driving force of NPO on silicon optical cw is also exponential (so don't expect the V of those silicon optical pointers to be very deep) ）， CPO may be a better buying point around Q4. In addition, the midstream of PCB may be an entry point in Q3, while Q1 and Q2 are more optimistic about the upstream. Of course, the upstream is not bad at present. Also, storing these, Micron is bound to reach a new high. At present, the biggest bottleneck of AI hardware is not just calculation, but storage and transmission, corresponding to storage and optical communication From Snowball: Mendeleev Apprentice