Don’t make long bets on AI Data Centers

Computing technology’s history is full of big, hulking machines.

The first computers were filled large warehouses and didn’t have traditional processing units, but specialized hand-wired cards that linked together to calculate large numbers and sort information from large data sets. Great for NASA and big corporations in the 1960s like General Motors and the like, but not something that was ever going to get into the homes of the average person.

Then came along a bunch of companies in Silicon Valley who wanted to have the power of mainframes on their desktops, not wanting to have to carve out time on mainframes (hence time-sharing) to compile and run programs on punch cards or magnetic tapes. Inventiveness eventually found a way from mainframe to mini computer (a misnomer, since they were the size of several modern server racks put together) eventually to the micro computer, and finally a form we understand: the desktop computer (or Personal Computer, abbreviated down to simply PC. Even the nomenclature has shrunk over time.)

We have shrunk down each decade into smaller form factors. The 1980s brought the PC to the desk, and by the end of it we had laptops and “luggables.” The 1990s brought processor size down and increased speed, and the decrease in footprint meant truly smaller and thinner devices could be made. Think the PDA and the Nokia brick phones everyone had.

The 2000s got us even smaller as phones went foldable, then turned into glass slabs with the introduction of truly reliable touchscreen technology. Tablets and the thinnest laptops one could imagine were introduced. Processors got more efficient and battery technology advanced.

The 2010s saw all this tech become thinner and more reliable as the decade wore on, and by the 2020s we now have foldable screens, thin batteries that last for days, and what most impressive: computers way more powerful than those mainframes constantly holding our attention in the palms of our hands. Wearables are possible these days – smart glasses with transparent heads-up or full displays, watches and rings that track our health, earbuds that can translate languages – all as things got smaller and smaller.

All this came about through what modern digital creators do best: we shrink things down for portability and efficiency.

Just in the world of computers alone, that trend toward efficiency drives more advancements than any other reason, because the smaller one’s code base is for a program equals a faster program. The smaller a transistor can be produced, the more that can be packed on a chip and increase the clock speed of a processor or memory. The packaging size decrease for chips means more features can fit on the same size circuit board inside of devices of all shapes and usefulness.

Everything inside of our modern technology ultimately gets better over time due to engineers seeing the necessity of reducing the footprint of their particular design – whether it be hardware or software.

It would seem logical to then jump to at least one particular conclusion when it comes to AI: the current form where we need giant data centers and huge amounts of energy and silicon to run the models? It will not be anything near the form it takes in the coming years.

AI faces the same problem that computers in their infancy had: their usefulness to certain sectors make it an attractive bet, but the technology needs to become more efficient while also increasing its capabilities until it can become mainstream, then a regular commodity.

The technology is in the same place that expensive mainframes were in during the early days of computing development. A lot of development got companies like OpenAI and Anthropic to where they are with ChatGPT and Claude to produce results, but they’ve got the problem of miniaturization to figure out still. The tech can’t truly be relevant in daily life, whether good or bad is ultimately yet-to-be decided.

Here’s the ultimate crux of the problem too: amid all of the moving parts that need to be made amid the AI gold rush, there just isn’t enough capacity as it stands to even make all of the CPUs, GPUs and memory chips in current configurations to fuel the requirements for advanced models.

According to figures provided by DSCMI, a typical cost for a data center is around $1,000 per square foot all in total. So a 250,000 square foot data center? That’s a $250 million investment.

That’s an average cost, not an actual cost per data center. Depending on what it is being used for, the data center might actually take additional funding for installing cooling units, electrical costs, various iterations of servers, software licenses and so forth.

Looking at other sources, those figures can vary and are usually measured in the megawatt scale or by square foot. Other sources pin the figures at $7 to $12 million per megawatt of “IT Load,” which can mean a lot of things. AI loads in their current iteration require higher processing power and memory usage, so the servers equipped to handle them are usually configured to the highest levels.

It also depends on who is bidding on the construction costs for the facility, and how those are sub-contracted out as well. Which vendors are chosen and what bulk discount prices they provide for such facilities.

The cost of a computer for the individual is much higher at these levels than the costs a business buying between 25,000 and 100,000+ servers running constantly. (Sources online vary in numbers, but that’s a good average.) These servers typically cost on a low end between $3,500 and $4,000 each, and they are rack mounted so depending on size and workload requirements (attached storage, for instance) those figures can get up much higher on a per-unit cost, upward say of $10,000 per unit.

A lot of companies charge much more, and figures from the likes of Dell and Nvidia weren’t available when I was researching how much investment this looks like on paper.

So scale up. XAi for instance is working on a 1 million square foot facility scale. For reference, that’s 22 acres of land. That’s $1 billion investment for the total buildout based on back-of-the-napkin calculations above.

Companies need way more than that at the current resources and capabilities of the tech involved to make a fully redundant, can-handle millions of users concurrently buildout of infrastructure. Dozens of these facilities.

TSMC, the world’s leading processor fabricator with their main production facilities in Taiwan, made 17 million wafers worth of 12-inch silicon wafers – which the processors are literally etched into with a variety of processes. These wafers can generate more than 300 processors with each wafer generated. So call that… 5.1 billion processors made in 2025. (Based on research from Waferpro.com posted in 2024. Yields might be better now with a smaller process node.)

Companies building processors like Nvidia, AMD and Intel pay – depending on the node used – between $25,000 and $45,000 per wafer generated. That’s $150 per chip produced, on the low end, for current technology costs. The server chips required to build out are much larger chips than consumer tech. The numbers can vary, and get more expensive with each advanced node to help cover R&D costs to TSMC. And server customers aren’t the only ones in demand for TSMC’s services, so not all of those chips produced are server chips, or graphics chips for Nvidia. How much capacity do they currently have, even with a facility build happening in Arizona. (Still years away, you need to build out giant clean rooms to do this work. Those take a long, long time.)

Back to the math. Let’s say in our 250,000 square foot example on the high end side, you have purchased the highest end server for each space in a rack, so you have 25,000 servers. That includes at least 50,000 processors, 1.6 million gigabytes of memory, whole petabytes of storage, and a really expensive cost of replacement. That doesn’t include how many graphics cards are in each server. Conservatively call it 100,000 processors in those as well.

I’m not including electricity costs, water costs, technical employee costs, maintenance of the facility… it all adds up ridiculously when you add it all together for just a single facility.

So even if we wanted to build out big data centers across the country, we couldn’t begin to build all the tech that will be required to stuff into server racks to make all of this work, and we can increase the production capacity but that will still take years before new factory expansion for building chips becomes available. The main company that manufactures the chips we use in daily life is located in Taiwan, of note. TSMC’s capacity is usually booked years ahead for companies like Apple to produce chips for new iPhones and Macs.

Plus, the technology fueling this drive toward AI infrastructure buildout CONSTANTLY changing as Nvidia, Intel, AMD, Samsung, TSMC and many other major companies driving the hardware capacity produce new innovations on a regular enough basis that it makes planning a pain for businesses to determine what tech they’ll invest in for as the data centers which haven’t even started construction in many of the proposed places being sought. This technology gets used for at least three years before it has to be upgraded, bringing on a cycle of replacements that will continue to increase in price at this rate.

This is with essentially a few companies that even have the capacity for chip manufacturing overall, as the likes of Intel and AMD shed themselves of their production units years ago. Now they are in the process of playing catch-up while also trying to make investors happy by cutting costs and increasing profits on the sales of their current products.

A lot of turning gears that have to work out just right for everything to go exactly to plan, one in which the owners of the data centers themselves are not on board with in the long run.

Data center builds do not do well with the kind of replacement turnover that is happening at the moment. Profitability comes with the reliable use of computing resources within these expensive warehouse facilities, without the need of server turnover for at least five to seven years selling to customers big and small for use for various online and internally-focused processing needs. I use a virtual server in a data center to host this and the other sites in the network together, and constantly look for ways to reduce that cost over the years. Plus you need another service for redundancy, meaning a second bill every month to cover that cost.

It builds up on businesses fast as the stack increases, so not only do the data centers have to figure out how to operate as cheaply as possible without major maintenance or replacement costs eating into them post-buildout, but those using the services they provide also have to reduce costs as well. They aren’t built as mechanisms for major job creation or to provide an ever-increasing value on tax bases locally. They’re built to run silent and cheap and print money for the owners. AI is not an efficient investment overall for companies that operate data centers because of the upfront costs of building, then the added tax of constant upgrades to keep up with ever-changing model advancements.

The ugly truth is also this: the AI firms and the companies that are building all the infrastructure know these facts, yet are pushing for a huge amount of capacity buildout that is predicted to be needed in the years ahead. No one actually knows whether all the data centers they want will actually be needed, thus why this might end up being a bad bet for the economy overall chasing a gold rush that might end up being an empty hole.

Yet to keep the ball rolling long enough to figure out the major problems with the technology and maintain their economic value for private equity and stock investors on the open market, they’re making bets with each other on paper that don’t ultimately produce any actual cash. NVIDIA promised billions of investment in OpenAI in 2025, but has since reversed course on that bet and promises MUCH LESS in the coming years for their partner on paper. Microsoft is pulling back from their OpenAI investment too – which was more about trading their current computing resources in house for access to models that haven’t really produced the results from customer use they wanted. Microsoft has thrust Copilot onto the masses in Windows, and users have in turn deemed the company more Microslop in its current form as a result.

All of this creates a marketplace where it seems as if there is scarcity and prices increase for everything involving chip manufacturing, which has rollover impacts on other industries. But has anyone not been able to buy a new phone because of memory shortages? A new computer? There’s plenty a computer that is several years old can still do in the current tech environment and not require upgrades to do so. We have for decades had “good enough” processing power for a lot of tasks, with specialized systems for gaming, audio and video work, and things like CAD and programming coming down in cost to make those systems mainstream through the first decades of the new millennium.

So the concept to me that we must build the data centers to fuel something that isn’t even in an initial form is, well… kind of dumb to me. Investors will soon come to the conclusion that this isn’t the way, just as they have in the past (think Dot Com bubble.)

The amount of money claimed to be involved is what is concerning – none of it in real value at the moment – and will lead to major pain when that bubble finally deflates or bursts. But maybe that doesn’t happen either; the bubble could simply shrink as the conditions for AI implementation shift from current priorities to new discoveries in the coming months and years.

Where will the real money be made in the long run? That bet is going to come up big with the company that figures out the efficiency problem first. Like all innovation, that is going to take time to figure out – maybe not on the current speed, but definitely in a long run where data centers turn out to be the wrong end of a long bet.