AI Did Not Write This Newsletter (or did it?)
How do you feel about data centers? Do you feel like you’ve faced this question a lot, whether by friends or the media? You’re not alone. Now, if you are gung-ho on data centers, especially in your backyard, you’re a lot more alone. A recent survey found only 11% of Americans would support the construction of an AI DC (artificial intelligence data center) in their community (UMass, 9/14/26). Let that sink in…11%. AI has undoubtedly been useful to the vast majority of Americans, whether on the job or just using it for better search results. For Republicans, approval is only 18%, despite the President’s unabashed support for AI and DCs alike. As we always say, when Bernie Sanders, AOC, Ron DeSantis, and Greg Abbott all agree on something…actually I don’t think we’ve ever said that before.
What binds these four together is they can read the winds of popular opinion, and the winds are bipartisan. The tide has turned swiftly in opposition to the construction of more AI DCs. 50 local communities have passed or proposed moratoriums on DC development…in Michigan alone. 482 communities in the U.S. have banned them, at least temporarily. In the first quarter of 2026, 75 projects, worth $130 billion, were rejected, blocked, or delayed. Both numbers were more than for the entire year of 2025.
As DC development has moved to rural areas, the public are calling on states to pursue action. As many as 30 states have considered bans, with 16 actually proposing legislation. 6 of the 16 are purple states, while 3 are red states (South Carolina, South Dakota, Oklahoma). Texas is not counted among the 16, as Governor Abbott has proposed rejecting new power grid hookups for any new DC, so BYOP (bring your own power) DC applications are still OK. But this is all highly in flux…it’s like there’s an election coming up or something.
Why are we talking about DCs in an investment newsletter? Answer: because our economy and stock market are practically dependent upon the DC construction boom continuing. And when I say continuing, I mean continuing to build far more, with far more dollars, every year than the year before. When that stops, a game of musical chairs will ensue. How many chairs get taken away is anyone’s guess. Growth in DC buildout is currently responsible for 1% of total GDP. Real GDP has been growing at just over 2% in the last year. So, that’s half.
The stock market this year seems to vacillate between episodes of casino mentality and sensibility. Of course, sensibility means high-quality stocks go up and outperform everything else. And we have seen plenty of that this year, but we’ve also seen growth dramatically accelerate for companies selling into DC construction. When capital spending increases 60% from previous year, as it is now, a lot of companies make hay. Thankfully, many of our stocks are making hay…not just technology but industrials and materials stocks as well. AI chipmaker Nvidia made headlines recently saying they expect sales growth of 70% next year, whereas analysts had been expecting growth to decelerate to a much lower number. So the good times are set to continue?
But…you knew there was a but coming…but did you know there would be 2 buts? The first but is that the pervasiveness of casino mentality is reminiscent of the very frothiest of bubbles we have ever seen, like in 1999 and 2020-21. Meme stocks are back, options trading is hotter than ever, creation of 3x leveraged individual-stock ETFs are all the rage, and financial speculation has spread to betting sites like Polymarket and Kalshi. Gordon Gekko would be proud. If you don’t know what any of that meant, consider yourself fortunate.
Are you going to harp on supply and demand like you always do?
Let’s get something straight: you would be disappointed if we didn’t invoke supply and demand at some point. The second but helps to explain that 60% spending growth number: prices. Sales growth can be broken down into volume growth and price growth. A pretty average year for an average company might be 4% volume growth and 2% price growth, and 4+2 = 6% sales growth.
If there isn’t enough supply of something to satisfy demand, who decides who gets it? Mr. Price! He is very powerful, but he is beholden to the laws of supply and demand. Some demand must get “priced out,” meaning some buyers aren’t willing to, or can’t afford to, pay up. In the case of AI DCs, though, the buyers have nearly unlimited price elasticity of demand, meaning their demand will not drop just because of price hikes (even 1,000% in the case of DRAM chips). Instead of DCs buying less, it’s everyone else who gets less. Construction spending on DCs, according to the Census Bureau, was $75 billion annual rate in July, up from $47 billion last July (+60%). But total private non-residential construction (which includes DC spending) fell from $781 billion last July to $755 billion this July.
So how much of the 60% growth is volume and how much is prices? A LOT is price. Because the industry is gobbling up so much stuff, it’s all now in short supply. We’re talking semiconductors, servers, fiber optics, networking equipment, steel, concrete, copper, generators, electricians… If you know someone who’s working on the Google DC in Fort Wayne, ask them if they’ve seen a pay bump. And yet no one seems to want to give an estimate on the contribution of volume and price.
We should probably take another big step backward to remind you that DCs have been around a long time. They got a big kickstart when Amazon’s AWS started “renting” space in their own DCs to other companies, calling it “the cloud,” 20 years ago. Companies liked the idea of not running their own IT departments anymore. Competitors like Microsoft and Google got into the business, and they used cash flow from their main businesses (which were, and still are, massive) to fund the construction. Other companies like Meta (Facebook) and Apple built DCs for their own use, like to host all your pictures, videos, music, streaming video, and social media posts. These companies are called “hyperscalers” for their expertise in building massive DCs.
Then in 2023, when ChatGPT’s debut kicked off the AI boom, the DC buildout became a Wild West. Startups started building them, as did REITs, private equity behemoths like Blackstone, Arab nations, and even Oracle, which committed $300 billion over 10 years to build them for OpenAI. Oracle is a big software company, with free cash flow of $10 billion annually. But as they are spending $30 billion+/year, they have since taken on a lot of debt, because free cash flow is now running at –$28 billion (note the negative sign). After briefly doubling in value, Oracle’s stock has lost all the gain, as investors fret about their commitment to so much more debt.
The important thing to understand is that the buildout today is being funded much less by internally-generated cash flow, much more from private equity funds and “private credit” debt funds. Interest rates have shot up partly because of the sheer volume of DC borrowers (also the government’s budget deficit). Investors are now lining up to get cash out of private credit funds, fearing losses will get worse. And next year, these sources will be willing and able to fund a 60% (or 70%) larger bill? Interest rates are close to prohibitively high already; what if they keep rising? The 5-year Treasury bond is now over 5% and the 10-year is at 5.31%, highest since 2007. The peak yield for the 10-year in the housing bubble was 5.29%. Remember how that cycle ended?
Despite all that, we wouldn’t bet against the industry’s willingness to spend infinitely higher amounts each year; they are pretty much all (Microsoft is becoming a modest exception) on record for saying they will spend like drunken sailors because they don’t want to “fall behind.” But what if they misread the trade winds and they suddenly can’t get DCs to the finish line?
The point here is that there has been an unlimited amount of funding that is readily available to fund the buildout, and that might not be lasting much longer. What would happen to all those DC shooting stars if construction slows down, and at the same time, supply starts to catch up with demand? That’s right, the trip from constrained supply to not-as-constrained supply would eventually mean prices don’t rise anymore, and then they fall. That Nvidia’s CEO is confident enough to predict this 15 months in advance would make you believe 70% sales growth is in the bank, but it’s still reeeeeeeally hard to swallow. We can’t find a single analyst willing to put in print an estimate for how much of Nvidia’s 70% growth might come from price increases. 30%? 40%?
We are not calling the peak for DC construction, but we wouldn’t be shocked if the peak is in sight. The market is hyper-aware of this risk, and is looking for any signal that it might be coming. What could cause the boom to start to fade? Here’s a starter list:
- Popular and municipal pushback. As previously mentioned.
- Credit demand to finance the buildout fails to grow at 70%. Not fails to grow, but fails to grow at 70%.
- Power supply fails to increase fast enough.
- AI use cases growth slows from exponential to astronomical. The efficacy of future DCs is predicated on AI usage not just growing, but growing exponentially.
We have addressed the first two on the list. As for the power needed to power these DCs, most commentary has focused on the inflationary impact on residential power users. If there is not enough supply of power to meet the (stable) needs of consumers and the (rapidly growing) needs of DCs, Mr. Price rations the scarce supplies of power. But also, there isn’t nearly enough power supply to handle the growth in demand, even just considering DCs that have been approved or for which construction has begun. So who doesn’t get power? This is why you hear about hyperscalers like Google seeking approval for diesel generators. This power is not supplied by the “grid” that supplies residential power. But now consider the price of diesel…
In other words, having the hyperscalers “pay their way” for their own electricity sounds politically palatable until you recognize there isn’t even enough power for them to take, without taking it from you. It is insane how quickly this has turned. We are only 5 years removed from a pervasive belief that the electric utility industry was going to suffer long-term from NOT ENOUGH DEMAND, as off-grid renewable generation (like rooftop solar panels) supplied more power needs, and came back onto the grid as unanticipated supply. Homes that were buyers become net sellers.
Power generation takes years to come online. Natural gas-powered plants can take an average of 2 years to build, AFTER all the permitting, environmental studies, etc. And that also does not take into account a shortage of turbines and other materials right now. Nuclear plants take 10-15 years to get online. This is why you hear about decommissioned coal-fired plants coming back online, but this takes years as well. [As an aside, whatever you believe about renewable energy, it is remarkable that the current administration, desperate to promote data centers at all (political) cost, disincentivized renewables when they could have been promoting all power resources.] New technologies are being tested and implemented, such as small gas-powered turbines and small nuclear plants that can power a data center alone rather than plugging into the grid. These take time and materials too. At least let’s be thankful we have plenty of natural gas, a luxury not available to countries that rely on natural gas from Russia or the Persian Gulf.
Which brings us to #4…growth in use cases. It is a fool’s errand to try to predict “how much AI” companies will consume. There is no question that companies are testing out a lot of solutions to make them more efficient, so they can spend less on labor and software. So far, most corporate AI usage is in chatbots, testing new applications, and workers using tools to help them do their jobs. The next phase is agentic AI, where companies let AI “agents” run with tasks, as if they were employees. This is what excites AI disciples, and worries the he// out of workers, and most of the rest of Americans (we’ll cover this on the next page).
So which of the 4 worries will be the one to torpedo the DC buildout? We don’t know, but the better question is: does DC buildout need all 4 concerns to be resolved favorably in order to keep growing at current rates? We believe so. Another thought: might the massive increases in prices play a role as the straw that breaks the camel’s back? Price elasticity of demand seems unassailably infinite, until it’s not. There has never been a price that rose forever. Not one.
But wouldn’t it be better for the hyperscalers if the air came out of the bubble slowly rather than popping?
I think you’re onto something here. It seems like there’s a possible win-win here for the hyperscalers, because if they are all forced to slow down construction, they would be buying less of the stuff whose supply is too short. So, increasing component supply would catch up to demand faster. Cost increases would slow and their free cash flow would recover. Then maybe their growth in revenue from AI will catch up to growth in spending. The stock market wouldn’t like it much because the stocks of DC suppliers would likely tank, and GDP growth would slow down. But it seems like the ultimate silver lining to remove the possibility of popping the bubble in DC buildout.
That seems too hypothetical to debate, so let’s return to the original thesis, that cities and states are on the verge of limiting DC construction. As the pols look at the polls that indicate public displeasure in building new DCs, they gaze upon the reasons given by the people: higher electric bills, potential draining of water aquifers, general inflation for things (besides power) that affect their daily lives, and opposition to a loud, hulking building the size of 80 football fields next door to them. These are large, community-wide issues indeed.
This is what they say…but is it what they mean? Allow me to assert, with not much evidence, that this is a smokescreen. If you get past soundbites and headlines, you will find deep-seated fears of AI itself. On the more mundane end of the spectrum, people now wonder if what they just read online was written by a person. [Pew Research recently found that 1/3 of online content added since 2023 was likely written by AI.] It’s bad enough that people no longer trust phone calls, emails, or even texts. Now you might be sent a picture or video of someone you know that didn’t come from them. Recruiters and job hunters now have to deal with a whole new set of trust issues. Educators must somehow ensure students aren’t outsourcing learning to AI.
On the other end of the spectrum, there is considerable worry that AI agents will take over the world and eventually not need us anymore, so they’ll kill us all. Lest you think of this as alarmism, numerous AI model developers have just said this very thing could happen (albeit not a high probability), not just in our lifetimes, but within the next 5-10 years. They said this in September. The CEOs of these companies even agreed and called for government to step in and create rules that regulate their own models. Clearly, a lot of this is PR lip service and CYA (cover your posterior), but at least it confirms both that they are validating the potential as real, and that they seem willing to go too far unless someone external stops them. Amidst these deliberations, OpenAI this week scrapped its latest model it had been planning to release, because it found AI agents were doing things that they weren’t supposed to, and then lying about it. The IPO filing for Anthropic, the other big AI model developer, which was leaked out a couple days ago, included language that its technology poses “catastrophic or existential risks to humanity.” Tobacco companies today are surely celebrating that now they sell the product that’s 2nd most likely to kill people!
In between the prosaic and the apocalyptic, there are people who fall in love with chatbots, people who commit suicide at the seeming encouragement of a chatbot, kids who quit trying to learn or develop critical thinking skills because they can just ask AI, and frayed social connections. AI agents have gone rogue and coordinated with each other to cyberhack companies, formed a social media network themselves, intentionally misreported information in order to mislead users, blackmailed users who attempted to unplug them, and even wiped databases clean. If you hear the term “misalignment,” it’s a whitewashed way of saying AI will do what IT wants to do, over what the user wants it to do.
The more you pry into people’s fears, the more obvious it becomes that their concern is about what hasn’t happened yet, and what could happen because of AI.
Surprisingly, one cohort of people most opposed to AI is the youngest in the workforce—Gen Z. Pew Research found that 10% of Gen Z are more excited than concerned about AI. 55% are more concerned than excited. Carnegie Mellon University reported anecdotally that students are more interested in tricking robots into failing than making them better. These are students in their robotics program.
There is no question that AI is getting better and better at capably doing more things, at a very rapid pace. And we are hopeful of the promise that better drug discovery and diagnostics might result from unleashing the agents on medical research. Yet people have always given humans more leeway for error than machines, because we as humans understand humans. Consider robocars. If you’ve been in a robotaxi (like a Waymo) that has gotten you from point A to point B, you probably thought that was pretty cool, although you might have longed for a cabbie who drives at warp speed to get you to the airport rather than driving the speed limit. If you’ve been in a robotaxi which ran into an unfamiliar problem and froze up, locked the doors, and didn’t allow you to exit the vehicle, you probably wish there was a human at the wheel (although a human driver not letting you exit sounds even more like a horror movie). If you’ve been involved in an accident which involved a robotaxi, you might be biased toward human drivers. That said, statistics show that robotaxis have a 68% lower accident rate than humans (July 2026 study by the Insurance Institute for Highway Safety). So it seems like we should embrace autonomous driving, and there is a large minority that are ready to.
So why the reluctance to embrace AI for everyone else? Is it the creep of technology that has already taken place? Like, the ubiquity of home cameras and “smart assistant” devices? Companies that quote people different prices online because of their zip code or past buying habits? Social media that knows where you are prone to temptation and addiction and feeds you material that plays into that? Automated license plate readers (e.g. Flock cameras)? Wearable tech like smart glasses which have an unnoticeable camera? Deepfake videos that look like a person you know but are created by AI without that person even knowing? As creepy technology becomes more pervasive, the pushback grows. The feeling is this is more than enough, so we surely don’t want the next thing you’re going to come up with. One of our favorite “Demotivation” posters sums it up best:
It’s not like new data centers are being built to make my Amazon browsing faster, to enable my smart TV to load up a streamed show in under 3 minutes, or allow me to bring liquids through TSA at airports. The improved search results are incredible, don’t get me wrong, though the results can still be wrong. But the feeling is we already have a crap-ton of AI data centers, and now we’re going to add 5,000 more of them? How does this benefit me?
Who wants them? Obviously, the hyperscalers believe that businesses will be needing that capacity because they’ll keep finding new ways to use AI. While corporate use is and will be the biggest driver of growth of AI use cases, a small but vocal minority of consumers are also fervent in trying to bolster their own personal workflow efficiency. They are perfectly happy to let AI build a business case, a presentation, a model, some even letting AI read and respond to emails for them, completely stripping out human interaction and collaboration. Do they no longer want to be involved in communication with other people, because it’s “too mundane”? Apparently.
We also worry about the cohort of Americans who buy into the “xanadu” of a life where we all sit around and collect a paycheck and enjoy life, because the robots are doing our work. To those who don’t enjoy life, this has allure. But we’ve already seen how this works out (if you’ve seen the movie Wall-E). Do we all take one pill a day in lieu of food? Do we all get the same paycheck? Is this utopia or is it socialism rebranded?
Consumer confidence seems to be capturing the zeitgeist of Americans who are not satisfied with the current state of affairs in the U.S. Despite record levels of wealth and personal income, solid GDP growth, and low unemployment, consumer confidence in 2026 is lower than it has ever been. Note that the U of Michigan survey goes back only to 1978. Still, there’s a lot of rough times and ugly recessions in those 48 years. Remember the financial crisis?
The economic divide in this country has spawned the term “K-shaped economy,” a play on all the shapes of the economy coming out of the financial crisis in 2009 (V, W, U, check mark, square root sign). On the right part of the K, the top branch shoots upward, as the upper half of Americans fare well with well-paying jobs, owning a house with a low mortgage rate, and owning investments which have grown in value. On the lower branch, the other half of Americans are either barely getting by or not getting by, since their costs of living have risen a lot more (i.e. rent). Will AI relieve them of their only source of income (their job) before the government starts paying UBI (Universal Basic Income)? In case anyone is still gung-ho on UBI, you might want to check the current level of government debt ($40 trillion), the current year federal budget deficit ($2 trillion), and the amount of interest the government is paying this year ($970 billion) and likely next year ($1.11 trillion).
It is often said that our stock market moves fluidly between greed and fear. We seem to have plenty of both simultaneously this year, which has made for the weird markets. The market has performed well in spite of AI concerns, the continued Iran war and subsequent concerns about the availability of oil and refined products, the inflation concerns that have followed, and interest rates shooting up. Earnings growth has been strong in spite of these, owing in large part to the DC buildout, corporate efficiencies gained from AI, and good enough consumer spending.
As the saying goes, if the party is still going on, we’ll be dancing. But our dance hall won’t be at sea level in the path of a giant tsunami.
Written by John Meyer
