Summary of the Core Content in One Sentence
This article reveals the most counterintuitive change in the AI infrastructure industry recently: Texas, which previously relied heavily on low taxes, lax regulations, and abundant resources to attract AI projects, has suddenly started to reevaluate its approach. From ordinary residents taking to the streets to protest the establishment of data centers, to the governor directly changing the approval rules, requiring AI data centers to bear all the costs of expanding the power grid, reducing water consumption, and lowering noise levels, without allowing ordinary residents to pay for the increased electricity and water bills. This is not a sign that Texas is giving up on the AI industry. Instead, it highlights how the entire industry has been focused on the idea that “AI computing power is the core of national competition,” only considering the billions in investment reported by companies and completely ignoring the numerous hidden public costs. By taking this step, Texas is trying to deflate the bubble in the AI infrastructure sector. This lesson is also highly relevant for regions in China that are pursuing strategies like “East Data West Computing” and building intelligent computing centers.
---
Explanation in 5 Simple Dimensions
1. The Competition Isn’t About NVIDIA Chips Anymore, but About Access to High-Wattage Power Outlets
In the past, the strength of an AI company was determined by how many high-end NVIDIA GPUs it owned; it seemed that as long as there was enough money and the chips could be purchased, the company could scale its computing power as much as it wanted. However, people are now realizing that chips don’t work in the air. A medium-sized AI training data center consumes as much electricity in one hour as tens of thousands of households do, and the power supply must not be interrupted, as even a few minutes of disruption could result in the loss of millions of training data. Texas’s power grid was expanded gradually to meet the needs of residents and factories. Now, with a large number of super-large data centers arriving, the demand for power is so concentrated that even if the state’s total power generation capacity is sufficient, the local substations and power lines cannot handle it. It’s like having a large electricity meter capacity at home, but plugging in ten electric water heaters at once would cause the wires to burn out. No matter how much money an AI company has or how many chips it has, if it can’t find a place to connect to high-wattage power, the project will fail. The competition for computing power has shifted from “grabbing chips” to “getting access to the power grid.”
2. 80% of the 410,000 Megawatt Power Requests Are Fake Orders from AI Companies
Texas’s power grid has a mountain of applications for power connections, with the total requested power equivalent to 2.8 times the state’s highest historical load, and nearly 90% of these come from data centers. However, most of these applications are not for actual projects. Applying for power connection costs almost nothing, and developers act like real estate speculators, submitting applications to ten different locations for the same project and choosing the one with the most favorable policies and lowest costs. The remaining applications are simply canceled, resulting in wasted resources. Previously, the grid was approved on a first-come, first-served basis, but now fake applications have blocked the process, preventing genuine projects from getting approved. Texas has changed the rules to require a deposit to prove that the applicant has the land, funding, and confirmed customers; if they back out, the deposit will be used to cover the grid’s losses, and ordinary residents won’t have to pay for the speculative behavior.
3. Past Promotions for Data Centers Were Misleading
When regions competed for data centers, they only highlighted the positive aspects, such as “a $10 billion project that will generate several hundred million in tax revenue annually.” However, they never mentioned the other half of the story: 90% of the $10 billion went to buying servers and chips, benefiting tech companies in Silicon Valley. The local community only got temporary construction jobs during the construction phase. After the center started operating, it often required only a few dozen maintenance staff, providing little long-term employment. The hidden costs were significant: building new power lines, competing for agricultural and residential water resources for cooling, and maintaining extra capacity in the power grid all ended up being borne by the residents through higher electricity and water bills. Now, Texas requires all projects to disclose their annual water and electricity consumption, the number of long-term jobs created, and the tax subsidies received, avoiding the misconception that the AI industry benefits the public.
4. Texas Is Not Against AI, but Screening Out Unreliable Projects
There’s a common misconception that Texas is trying to drive the AI industry away. In the past, anyone could claim to build an AI data center and get land and subsidies with just a PPT. Now, only speculative projects that aim to speculate on land, use inefficient cooling technologies, or have no paying customers are rejected. By limiting access to power and resources, Texas is ensuring that only reliable projects with stable customers and willing to bear the costs of expanding the power grid are approved. This helps to eliminate the bubble of false demand and enables high-quality AI projects to succeed.
5. China Should Avoid Repeating Texas’s Mistakes
China has the advantage of centralized resources for large-scale projects, and the speed of building computing infrastructure could be faster than in the US. However, it’s important not to go to the other extreme of competing for the largest or most racks. If regions build large centers without actual demand, the billions in investment will be wasted. AI chips evolve rapidly, and without long-term customers, government subsidies won’t be enough to cover the costs of equipment depreciation and technological obsolescence. It’s unnecessary for all regions to build large computing centers. The east, close to users, is suitable for real-time AI tasks with high response times, while the west, with abundant wind and solar power and lower electricity costs, is better for less time-critical AI training and data storage. Avoid forcing unsuitable projects into these areas, as that will result in a situation where there’s computing power but no customers, or vice versa.
---
The Most Fundamental Logic
Whether in the US or China, the principles of building AI infrastructure are simple: Those who want to launch projects must clearly understand the costs and benefits. The benefits should not all go to companies and investors, while the costs should not be borne by the public and local governments. The location of computing power must be accompanied by a clear understanding of all associated costs, both explicit and implicit. It’s essential to avoid misleading practices and make informed decisions.