AI Data Center Investment Mania Goes Exponential as Money Gets Thrown at Hurdles & Shortages

Exponential curves for investments burn out, often with a pop. But they can last longer than imagined. By Wolf Richter for WOLF STREET. The race to build and equip $1 trillion of AI data centers as fast as possible, no matter what the costs and hurdles, has run into revolts that have triggered local data-center construction moratoriums and bans across dozens of states, with a bunch of states considering data-center construction moratoriums and bans – New York already implemented a 1-year moratorium until it gets its regulations sorted out – amid concerns about soaring electricity costs, blackouts, water shortages, the issues caused by onsite gas-turbine or diesel power-generators, etc. Some of the planned data centers would consume multiple gigawatts of power, but the grid cannot supply that kind of power all of a sudden. At the same time, there are enormous unanswered questions about the commercial viability of these massive amounts of investments, amid doubts that AI will generate the trillions of dollars in revenues to make that investment worthwhile. Where are these trillions of dollars in new revenues supposed to come from? No one knows. But build it, and the revenues will come? Nevertheless, the race to build AI data centers continues unabated. The amount spent only on the construction of data centers spiked by 6.2% month-over-month, and by 57% year-over-year to a seasonally adjusted annual rate of $75 billion in July, according to construction data from the Census Bureau today. Since the beginning of 2021, monthly construction spending on data centers has spiked by 717%, along a near-exponential curve: Obviously, these kinds of the-sky-is-the-limit near-exponential curves eventually fizzle. But they can last longer than imagined. These amounts only reflect the construction costs of the buildings, the improvements around the buildings, and the equipment integrated into the buildings, such as HVAC systems. The amounts do not include the most expensive parts of a functioning data center: the servers, the racks, the electronic and optical equipment to connect the servers to the internet, the electrical equipment to supply power and cooling to the servers, the power generators, the transmission lines, etc. To accommodate this mad rush to build data centers as quickly as possible, construction companies and suppliers have developed technologies that speed up the work of building and equipping data centers to get them up and running faster. According to a report by the WSJ, they include: Custom concrete: “Cement manufacturer Amrize uses predictive modeling to design custom concrete mixes, a process traditionally done through lengthy trial and error; time savings: several weeks. Robotic concrete driller: “Stanley Black & Decker’s DeWalt brand and August Robotics have created a robot that drills thousands of holes to anchor server racks and other systems to the floor; time savings: six weeks.” Off-site construction of electrical and mechanical rooms: “Clayco and Turner subsidiary xPL Offsite make modular electrical and mechanical rooms at off-site factories, then truck them to data centers for installation; Time savings: several months.” Optical cable connectors: “3M makes components for fiber optic cables that allow servers to be connected in seconds, not minutes. A data center can have hundreds of thousands of connectors. Time savings: six months.” Bottlenecks and shortages have dogged the manufacturers of on-site power generation equipment, especially gas turbines. Companies have started repurposing retired jet engines for on-site power generators. Musk has jumped into the fray to alleviate the shortages for his own data centers. In July, he acquired APR Energy, which makes among other things gas-turbine power generator sets. But the biggest bottleneck for gas turbine manufacturers are the blades and vanes, so Musk confirmed over the weekend that SpaceX will start manufacturing turbine blades and vanes. Shortages of semiconductors, including memory chips for AI servers, have caused prices of semiconductors to soar, and they have started to spread to consumer electronics, and from there to inflation metrics. There are now shortages of specialized labor, such as electricians. This kind of sudden maniac spending boom, funded by corporate cash and massive debt and equity issuance, leaves its marks everywhere, including by helping to push up government bond yields as they all compete for the same pool of money. This drive to build and equip and power up gigantic data centers, no matter what the costs and hurdles, is pulling resources and labor from other projects, and costs are rising, and we’re already seeing it in the inflation data. For example, the Producer Price Index for construction materials – steel mill products, concrete, lumber, gypsum, etc. – spiked by 10.5% year-over-year, the biggest increase since June 2022. Drilling down into the product category level, the PPI for “Fabricated Structural Metal Bar Joists and Concrete Reinforcing Bars” – which includes steel joists and rebar – spiked by 17.7% year-over-year. Since January 2021, the PPI for construction materials has spiked by 46%; since January 2020, by 58%. This chart shows the price level of the index. For more, see our analysis: Construction Inflation for Nonresidential Buildings Soars amid AI Investment Mania Enjoy reading WOLF STREET and want to support it? You can donate. I appreciate it immensely. Click on the mug to find out how: Like the Space Race this is big. Control AI and AI related operations and you lead. China vs USA. Unlike the Space Race, the output of all the billions spent in training fancy LLM models is a just a data file full of mathematical weights which can be both distilled from other models or simple stolen as a data file. Once created or stolen it can then hosted in relatively small machine to provide inference models to use without limits and offline. That’s right, billions spent on data centers and training frontier models can be copied in seconds and then used privately by a machine that fits in a single server rack. Vastly easier to copy than funding and engineering rockets, training astronauts, building launch infrastructure and space communications systems and executing space missions. you nailed it Ben. there’s no impenetrable moat being built here. Good point. Just like old IBM 360s all the huge data center equipment will be miniaturized quickly into a laptop size Gizmo, and the huge data centers will just be another unused shopping center mall. The phone in your pocket has as much computing power as the super computers from the 1980’s. I think the bigger problem is that AI is now generating a large portion of webpages and documents. AI is now being trained on that output. Like a self licking ice cream cone. Easier said than done. We started hitting the limits of Moores Law a while back. Amen. Also, while certain gpus that use CUDA are useful because most AI projects can and are most easily run on them, inference can run on a potato. Even older CPUs like my AMD 5600h with a 6Gb vram gpu CAN RUN the best, smaller ~35b or smaller models with quantization plus 48Gb of ram. Two old computers can do a lot of work, if more slowly. Hence, the real value opportunity is getting robotics or video generation or AI-assisted creation of drugs/ materials, e.g., like Coscientist. The AI ARE incestuously buying each other to ensure they share in any breakthroughs. However, as most such work only requires INFERENCE, it can be done in cheaper ASIC or Chinese GPUs. Meanwhile, as Felix & Friends on YouTube just masterfully demonstrated there are few safe havens now. Those governments cannot have reasonable budgets without taxing the principal component that has had the highest US/EU/UK/Japan growth: Capital. Lead what? China seems to be doing just fine. China has achieved parity with US in R&D investments. And based on where the trajectories and government supports are, it will likely outstrip and push ahead. China’s open source models are competitive with US peers, their paid AI offerings are significantly cheaper. US companies are being propped up by protectionist policies against importing Chinese technology. If the Chinese EVs landed here – US consumers will have real choices at mass market prices for a change. The reason US was ahead in space sciences because it’s investments were supported by being an attractive hub for talent globally. Now, knowledge is more democratized and accessible than ever and the country with the best ability to attract, nurture and grow the brightest minds will pull ahead, like it has been the recipe in the past. I’m not sure what US is going to lead in 10 years from now, considering it’s already behind in significant areas that matter to the future, and actively sabotaging it’s relationships with reliable partners. I don’t know. have you seen the performance of Chinese Big Tech stocks. Smart money says they are way behind. BABA is down 60% from its peak in 2020 and at the same time NVDA is up 2000%. They are un investable companies. The stocks have been flat or trending down for several years during one of the biggest technological moments in history? China has 1 tech stock with over a 1 trillion market cap and that is tencent at 1.1 trillion. The US has 12. That global talent came from 1940’s Germany. “US companies are being propped up by protectionist policies against importing Chinese technology.” So what you’re suggesting is that we hand our tech & car industries over to China for lower prices? Wait a second, haven’t we tried that since before 2001, when China joined the WTO? The simple fact of the matter is that things / everything in the USA costs more due to labor costs. That’s a fact. So you’re arguing for untold numbers of people working in tech, car mfg, etc to be replaced with cheaper Chinese products. How many people in the US are going to go for this, when that means they lose the job? BENW wrote: “….. everything in the USA costs more due to labor costs..” Maybe it is time to look at why ? What makes labor in the USA more expensive ? What do other countries do to make their workers less expensive ? I can think of a few things to look at. 1. Health care. This is a sector screaming for reform. 2. Education. Getting rid of the private for-profit diploma mills and how they sell their service to gullible students with debt. 3. Housing. Why is housing unaffordable for so many working people ? I am sure there are other areas the posters here can add to this list. “Like the Space Race this is big” Right – Just imagine how poorly the US macroeconomy of the last 55 years would have fared without Tang. I suppose with all the hype, and insane amounts of money involve…. This AI thing better change the world and make boat loads of money doing it. I def ask it very complicated questions in a line of thinking. Before this you kind of had to run down the info yourself. And of course it’s still a good idea to double check the info if it’s important. I just use the Google one on their website, not any of the more sophisticated ones, because I really don’t have time to sit down and integrate them into life. Thinking of booting up the PC and messing with some of them. The costs associated are kind of annoying. They want quite a lot to access them. Depends upon what you are going to do with the AI. All the major players offer some useful levels of function for free. The first paid level expands considerably upon that, but doesn’t cost more than say a premium streaming TV service. It does take some experience to get best use out of them. Most people approach it initially like a super Google search. It can do this kind of okay. There are better ways to query them however. You’ll get much better results by assigning them theoretical characteristics and limiting scope prior to the AI answering. It will take a bit more thoughtfulness and effort on your part to get good results. When you get over the initial threshold however those results can be multiples more useful than just basic searching for answers to questions. I thought, “Self, AI is so darn great I should ask AI!” So I did: “is AI data center buildout a lot like lemmings?” AI Data Center Buildout: A Lemming-Like Rush The rapid expansion of AI data centers is often likened to a lemming-like rush due to the aggressive and uncoordinated nature of the buildout. This comparison highlights several key aspects of the current situation. Characteristics of the Buildout Aggressive Expansion: Companies are rapidly constructing data centers to meet the soaring demand for computational resources driven by advancements in AI. Lack of Coordination: The expansion is happening without a unified strategy, leading to potential oversights in sustainability and community impact. Implications for Communities Sustainability Concerns: The swift growth raises questions about the environmental impact, particularly regarding energy consumption and resource allocation. Local Impact: Communities may face challenges such as increased noise, pollution, and strain on local infrastructure as these large facilities are built. Conclusion The analogy of lemmings captures the urgency and potential recklessness of the current AI data center buildout, emphasizing the need for careful planning and consideration of both environmental and community impacts. I’m curious as to how much the construction spending/demand for construction labor for data centers has had an impact on residential construction material costs/labor costs and how that’s had an impact on housing costs. Does anyone know of any reports/studies that would provide some more clarification on that? Thanks for the article Wolf! Re: ‘To accommodate this mad rush to build data centers as quickly as possible, construction companies and suppliers have developed technologies that speed up the work of building and equipping data centers to get them up and running faster. ” I infer these efficiencies will facilitate the arrival of the consequences of the buildout sooner than expected. The data centers are consuming huge amounts of steel right now – to the point of nearing shortages. Beam is hard to come by and sheet is really tight. It’s driven prices up around 50% for us this year. Our primary steel supplier told us to shop around (as opposed to utilizing our purchase program) for the rest of the year because their primary mills are missing coil delivery dates by months, and they can’t get enough material. They cited data center demand as the elephant in the room with transportation equipment still being soft, and agricultural equipment suffering from a prolonged downturn. Under the cover of tariffs, mills are jacking prices to the moon, rivaling the wild numbers of 2021. Looking forward to that bubble popping. I would imagine steel scrap prices are soaring or will be about to also. MW: Bessent now says debt-reduction plan could be months away. It will face hurdles when it arrives. Bessent’s debt-reduction plan just months away meets the newly elected 2026-2028 Legislature. That should be entertaining. What, me worry? I leave worrying to my wife. She’s really good at it. I can’t wait to see what that plan looks like and the sparks that it causes. Not sure how the Treasury Secretary has any say so in reducing the debt. I would love for him & Warsh to be very vocal about the need to reduce the debt. Their best playbook is to get serious about squashing inflation by raising rates which will cause debt service to balloon faster. That might be the only thing that gets America’s / Congress’ collective attention. There is a lot of clever AI financing going on, much of it being made by entities who create the financing, then somehow manage to disengage themselves from the responsibility of many of those loans. So it looks to me like the average investor might need to be even more careful than usual where they put their money. The limitations of building data centers might be the one thing that saves companies a lot of money and maybe even get some good press. If everything was super cheap and power available there would be massive excessive compute. AI is like a normal calculator versus an expensive scientific. The majority of people will be fine with a normal AI model, can likely run it locally, and do it in the cheap. Happy to have slightly slower code production or slightly less quality if I don’t have to buy tokens. OpenAI is clueless as they are chasing better frontier models when others are pursuing products(some more successfully than others). Still unclear how any make it profitable. This is the most grotesque waste of natural resources in human history. By God somebody stop this greed. It is despicable beyond words. One thing about you DC, you really know how to sugarcoat it. I had lost whatever faith I might still have had in man’s morality until you cleared it up for me. You were being sarcastic, right? ;) Look at it as a new deal WPA program with all the jobs it Is creating, but just for one time…. Then someday all the overbuilt data centers can be repurposed for something else…..maybe. Rusting junkyards? I put this query to the free version of Chatgpt: Assuming your were an investor, how would you characterize the current level of investment going into AI, Data centers and related resources? Do you think a crash is going to happen or that a bubble is forming? The answer was far too detailed, large and far reaching to quote here. I’d encourage readers to copy this query and paste it in. Then peruse the results. I will quote one short striking passage from the result. Mind you if you paste the query in it won’t be word for word the same result, but should be very similar. Anyway it said as a historical analogy it would describe the current situation with AI as: I’d compare it to railroads + electrification + the dot-com boom, with a little bit of 2008-style financing layered on top. The content that you read in your AI results was stolen by ChatGPT from publishers around the internet, including wolfstreet.com, and ChatGPT then sold it to you, and you paid for this stolen material via your ChatGPT subscription. AI is killing publishers by stealing their content and their traffic and their ad revenues. AI will in some form of another destroy almost everything on earth. It may start out being gradual over time, but at some point, we fall off a cliff. Then, we’ve all got to be concerned about survival. It’s only a matter of time before some rogue or intentional cyber attack makes everyone think, “Oh shit. AI might need to be put back in a box.” One YouTuber made a video on how AI is like Soylent Green. It is the people, stolen, mushed up, and then sold back to you. Pretty hilarious take on it. What if…Just imagine for a moment. …Perhaps an even more dramatic seismic shift in AI occurs where data centers become obsolete… It is possible. The other side of the globe leapfrogs the US as “We” pour trillions of dollars at this only to be one-uped. I asked Gemni the following, and got the following answer: A typical smart phone in 2026 is equivalent to a supercomputer from what year? “A typical smartphone in 2026 is equivalent in raw computing power to a top-tier global supercomputer from the late 1990s to the early 2000s. While it completely dwarfs iconic 1980s machines like the Cray-2 by thousands of times, its raw processing throughput matches the absolute fastest multi-million dollar supercomputing clusters built at the turn of the millennium.” I put the question to Google about my Jitterbug phone. “A typical Jitterbug phone in 2026 is roughly equivalent in processing power to the absolute fastest world-champion supercomputers from 1993 to 1994. While it might seem surprising that a phone designed specifically for seniors with basic features—like large text, simplified menus, and an emergency response button—is a powerhouse, the evolution of mobile silicon has made computing power incredibly cheap and dense.” I wondered how a Tracphone would measure up ? To my surprise, it was even better than the more expensive Jitterbug. “A typical low-cost, budget smartphone (like a TracFone) in 2026 possesses raw computing power equivalent to a world-class supercomputer from the year 1997 to 1999.” Google should change their name to “HAL”. And HAL was right, it is “human Error”. Amazing article and comments. This article reminded me of the Gold Rush days when rumour and hype took over reality. Yukon….many heading for Dawson City. People went nuts with iconic pictures of long lines of hopefuls trudging up the Chilkoot Pass, then stopping to build rafts to get to the north. Except for a few lucksters at the beginning, you know who made the money? Suppliers, boozers, whores, etc. Some people left prospecting to stay in the support column, but most returned home utterly broke and defeated. I remembered working on my own gold rush event. 3 days a week I was hired to fly a ton of beer and pop to a mining camp in the NWT, 200 miles each way. They chartered aircraft to haul beer even though there was an available road (long nasty drive, but still). Several drill tent camps in the mountains, and 7 helicopters on site to move men and supplies around. Not one mine was ever built there, in fact….the company town of the highest paid miners in the World, Tungsten NWT, was shut down just a few years later as a new tungsten mine opened up in China and undercut them by 30%. The entire town was abandoned. A few years later I flew over the same long and nasty road and watched giant orange Allied Moving van trucks heading south. A 300 mile gravel road. All this recent mining hype was paid for with Investor Money. All of it. Data mining, gold mining…all the same. One more thing. In my twenties I was going to buy some miner stocks. A wealthy customer of ours…..multi millionaire, he shook his head NO when I asked him for his advice. His words, “It isn’t about the mining property, it isn’t the property, it’s the promoter that makes you the money. Wrong promoter, don’t buy it”. Same thing today, even more so I would imagine. Just saw an article headline that says AZ AG is calling for a moratorium on new data centers. I can’t remember when the last time I read about a NEW data center being approved. I’m sure there are some, but it seems to be about 10 or more are being shot down for each one that’s approved. GO VOTERS! More power to you. Make yourselves heard. Ford sales for Aug 170k this is small inc of july, but about the same for June last year aug sales were higher than june so does this imply sales in aug were “Terrible”? Last year was distorted by the end of the EV incentives on Sep 30, which caused a huge front-running surge in EV sales which boosted total sales. So the year-over-year comparisons are against this inflated base a year ago. Then after the EV incentives went off on Oct 1, sales plunged. So you will see HUGE year-over-year sales gains starting next month. The learning curve cost reductions on data center construction and on server power efficiency must be some of the steepest declines in human history. The pace of innovation in these areas is truly astonishing. It will be very interesting to see what things look like in 20 years, after enough time for new entrants based on these technologies to have replaced today’s incumbents. An example: 20 years ago, many small towns still had a functioning, subscription-based newspaper. Today, they’re almost all gone, replaced by “content creators” using online tools like websites and social media. Every time I see one of these articles, Global Crossing comes to mind. BOND SELL-OFF DEEPENS I work in the tech industry and my companies “token maxing” approach just flipped on a dime, with many coworkers abruptly losing access to critical AI tools as AI spending blew past budget expectations. Many other businesses are adopting the same approach as AI-driven increases in productivity and innovation are nowhere near expectations. As Wolf alludes to here, eventually these capex investments run out of steam and reach a sharp inflection point. Given broader industry trends, I think we’re 1-2 quarters away from this exponential increase in spending rapidly leveling off. As soon as that happens, AI bubble begins to pop. I think we’re in the 7th or 8th inning here. 100% correct. Biotech jumps of stuff like this all the time, quickly squeezes it for productivity/innovation gains, does an honest assessment and then moves on. The profitable companies one are already moving on. LOL! This will be the largest mis-allocation of capital and resources the human race has ever seen, exponentially larger. Hedge accordingly. Lotsa money being spent on something that will only benefit the jaded oligarch class. In two years we’ll be talking about a glut of unemployed electricians. Polticians who vote for this crap will be voted out. AI will have the same fate as flock cameras. “Lotsa money being spent on something that will only benefit the jaded oligarch class.” Waitaminute. It’s THEIR money that’s getting spent on this stuff. If it blows up, it’s THEIR money that’ll blow up. If it takes the stock market down with it, it will be biggest reduction in wealth inequality in history, I would say. Great article! Look forward to further coverage of the AI-mania. Another great data point to look into is the price of the computer hardware itself. Nvidia recently revised prices upwards (even for existing contracts as I understand). Each new build continues to generate pressure on the already strained HBM supply chain crisis, so each new build causes prices to build these kind of hyperscaler DC’s to skyrocket (and indeed all computer in general, including a lot of consumer electronics). A vicious cycle if I have seen one. The NY Fed President John Williams said that the increased yields in treasuries are due to a strong economy. How did this moron get to where he is? He was at least partly correct … just another way of saying: They’re letting the economy “run hot,” with higher inflation and higher nominal growth (+8.0% nominal GDP growth annualized in Q2), and therefore higher yields reflect that. It all goes together, it’s the official policy now to deal with the government debt. We’ve been talking about this for a while. The spending on AI related services will happen, just a matter of when. Will it be before the Chinese commodification of AI models makes it make more sense to have a free local AI doing the majority of the AI-ing on locally secure servers, or after. Not saying to use Chinese AI’s! Sometimes free is the most expensive option, lol