AI data centers have gained traction in the last few years, and all the tech companies are heavily investing in them.
Though SpaceX is already making headway for orbital computing infrastructure, and Starcloud has already demonstrated that GPUs are capable of running in space, plans for scaling orbital compute will still take many years to fructify. Till then, we will need to have data centers on Earth and need to find novel ways that take care of the needs of the environment, society and of scaling the compute.
Divided House
The world seems to be quite divided today over those data centers. People like me who are on the X-verse are quite bullish and optimistic about the abundance and prosperity that AI and data centers are bringing and going to bring in the world.
Like any technology, there is a lot of confusion regarding data centers and the kind of impact they have on us. In countries like China, these doubts have been deftly handled by their government and most of the Chinese population seems to be utterly pro-AI, whereas in countries like the USA, both the regulatory regime and the posture of some leaders in AI such as Anthropic have led people to believe that data centers and AI will take away their jobs and are a net negative for humanity. The situation is pretty different in countries like India where despite having so much population, AI literacy is extremely poor and most of the folks have no idea what’s happening in the realm of AI and data centers. Either most of them are doomers or are oblivious to what AI can do.
Myths and Reality
What is the truth? The answer lies somewhere in the middle, and it is important to understand that most of the myths surrounding AI and the data centers are very far from the reality. Of course, data centers consume power, space and resources. They do have a footprint, but what doesn’t. Every technology has some pros and cons, and what is important to see is whether the technology is offering more than what it is taking away from the environment or not. Some of the concerns surrounding AI and data centers such as mismanaged power grids, dependency on a few factories in Taiwan and South Korea, and inefficiency of old chips, are real and valid. These concerns need to be managed and alleviated in a proper manner if we want to continue building data centers. However, as I said, most of the myths are simply rubbish. Let us dispel some of the common myths that are formed about the data centers and AI.
| Myth | Reality |
| A single ChatGPT query uses as much energy as boiling a kettle, running a load of laundry, or driving a car. | Those viral claims mixed the huge cost of training a model with the tiny cost of one question. A typical ChatGPT-style text query uses about 0.3 watt-hours. Boiling a kettle of water takes about 100 watt-hours, so one kettle boil equals roughly 300 questions. And that is also becoming more efficient with time. |
| Data centers already consume a huge share of the world’s electricity. | Global data-center electricity including cooling consumes around 1.5% of world generation (around 485 TWh out of a total of about 31,300 TWh). Lighting in buildings consumes four times the electricity of data centers. |
| Data centers use more water than cities, farms, or other major human activities. | Agriculture consumes around 72% of global freshwater withdrawals. US data centers consumed around 17.4 billion gallons a year in 2023, whereas lawn irrigation in the US alone consumed 3.3 trillion gallons a year and golf courses consumed around 530 billion gallons a year. |
| Data centers occupy vast land compared with solar farms, farmland, or suburbs. | All US data centers will occupy a mere 1,400 square miles by 2028, and it will be just 0.3% of US farmland. These data centers also pay massive taxes in comparison to other industries. Also, data centers are compact electricity users. Solar is a spread-out electricity collector. You should not compare the two as if they were the same kind of land use. Further, since 2014, converting land into suburbs has used several times more U.S. prime farmland than solar farms have. |
| Data centers raise every household’s electricity bill wherever they appear. | If the power grid already has extra unused plants, a big new customer can actually help by using those plants more, so the cost per unit can stay flat or even fall. Further, governments can change the rules so that a data center pays for the extra plants and power lines it needs, instead of passing that cost to households. In some counties the same centers pay enough local tax that home property taxes fall. The honest picture is that a household power bill can rise in a tight region, but it is not true that data centers raise every bill wherever they appear. |
| Noise and waste-heat islands make surrounding neighbourhoods unliveable. | A data center dumps heat and some fan or generator noise at the fence, but that is small next to a city’s air conditioners and is not enough to make a neighbourhood unliveable. In cold places, piping the heat into homes can handle most of it. |
| AI data centers destroy more jobs than they create, so the energy is socially wasted. | Jobs are not a fixed pile. AI can wipe out some office tasks and still raise living standards if each remaining hour of work produces more of what people value, just as electric lights ended lamplighter jobs and wages still rose. Building and running the centers also adds local trades work, and using electricity to lift output per hour is the opposite of wasting energy. |
| Building more compute dangerously centralizes political and economic power in a few firms. | A few firms do hold a lot of the best chips and cloud, and that is a real competition problem. Cutting their power does not hand the models to everyone else. It just makes intelligence more expensive for universities and smaller companies too. |
| AI is a financial bubble, therefore the electricity used to train models is wasted. | A bubble means investors overpaid for the stock. It does not mean the electricity was thrown away. If a firm collapses, the chips and trained models can still run useful work for years. The same kind of computing is already used to fold proteins for medicine, to forecast weather, and to run ordinary tools like search, maps, and office software. Those are real uses, so the power was not poured down the drain. |
Optimism
The abovementioned table provides a high-level overview of the misconceptions surrounding the data centers. There is still a non-zero probability that some catastrophic event like WW3 happens and all future growth of humanity is stalled. However, I think the likelihood of the same happening is quite low, and humanity is on its way to make a climb towards Kardashev 2. What are your thoughts?