Why Building AI Data Centres Isn’t Working Anymore

[00:00:00] Speaker 1: Hi, welcome to another episode of ColdFusion. $650 billion, that's the amount four tech companies were expected to spend in 2026 alone for new data centres. To put $650 billion into perspective, stack that amount in $100 bills and the pile would rise 710 kilometres, clearing the orbit of the International Space Station by over 300 kilometres. And the plan is to get this number up to $9 trillion by 2030. And yet, with so much money allocated to this, nearly half of the data centres planned to open in the US this year have already been delayed or cancelled outright. According to a report by Bloomberg, around 140 data centre projects were supposed to open in the US this year. Together, they represent roughly 12 gigawatts of computing power, enough energy to power 9 million homes. But there's a problem. So far, a third of them are actually being built. The rest just sits entirely on paper, press releases and announcements. In another report by the Financial Times, the data centres that are actually being built will likely face major delays. The companies involved in building them may deny that there are any delays, but satellite imagery says otherwise. [00:01:11] Speaker 2: But if you go and look at Microsoft, for example, none of the data centres that they have announced have broken ground since 2023 have actually been finished. The Fairwater data centres, they've said that that was fully complete. It's not even half complete. It's one of the strangest things I've seen in history. [00:01:29] Speaker 1: But beyond just the completion issue, the general public hates data centres. Between infrasound noise making people sick, polluted water, higher energy costs and a myriad of other problems, people have had enough. Because who's actually paying the price? Not the shareholders. Not the companies collecting tax breaks. It's ordinary people. Homeowners. Families. Communities. Those who happen to live a few hundred metres from one of these facilities. They never got much say in the matter. So here's the question. If the world is spending so much money on AI infrastructure, how is it possible that that very infrastructure itself is falling apart before it even gets built? And is this finally the sign that many have been predicting? Let's get into it. [00:02:13] Speaker 3: You are watching Toll Fusion TV. [00:02:20] Speaker 4: I'd love to hear your perspective, not on the spend, but on the return needed to make those numbers work. [00:02:26] Speaker 5: They'll never get it. Yeah, they're just shitting away that money. Right, we're going to spend a trillion dollars because we need all this data centre capacity. Some of them are spending more cash than they have available. [00:02:41] Speaker 4: They're shitting away the money at scale. [00:02:43] Speaker 5: It's hard to do, right? But... [00:02:46] Speaker 1: When the AI boom exploded into public consciousness with ChatGPT in late 2022, the race to build the underlying infrastructure became almost frenzied. Data centres became the hottest asset class on the planet. Investors poured in and governments competed. Nowhere was the build-out more intense than the United States. Last year, it was estimated that 92% of US GDP growth came from data centre spending. If you remove AI infrastructure, the rest of the economy grew by 0.1%. That's why the stock market skyrocketing felt so out of touch. And to give you an idea of where we're at, the shoe company Allbirds recently announced that they're pivoting to focus on leasing AI data centre equipment. It made no sense, but their stock rallied 580% as a result. All the signs are here. Data centres are in vogue. The pitch was simple. AI gets smarter the more computing power you give it. Computing power comes from data centres. In the US alone, the number of AI data centres went from essentially zero a decade ago to somewhere between 4,000 and 5,400 today. The pace and scale of the build-out was so violent that more money has been committed to data centres in six years than was given to the Marshall Plan to rebuild Europe after World War II, the Manhattan Project to build the atomic bomb, the entire Apollo program, and the cost of the International Space Station combined with $120 billion extra left over. All of this spending for a business plan that isn't even solid yet, and besides, GPUs rapidly become outdated and they become expensive to keep replacing. I'm not saying that AI is never going to work, but is it going to be profitable in the time frame that's needed for funding not to completely collapse? It's looking less likely. It's perfectly possible that this is all a huge waste of money. [00:04:40] Speaker 5: So when you think about all these people spending all this money, look, if you're the best programming, you know, like CLOD is, and that's your niche, great. But I don't know what the niche of these other ones are. Do you? [00:04:53] Speaker 1: No. More than two of these facilities were being built every single week at the peak of the boom. States like Virginia, Texas, Georgia, and Arizona were rolling out the red carpet. The tax breaks made the data centre business even more attractive. Texas, for example, handed out over $1 billion in incentives for the Stargate project alone, the $500 billion OpenAI and Oracle AI campus in Abilene. Virginia, home to what's known as Data Centre Alley in Loudoun County, saw 56 data centre projects receive nearly a billion dollars in tax savings in a single fiscal year. The promises attached to all of this were enormous. Thousands of jobs, a new digital industrial revolution. One analyst compared it to the railroad boom of the 19th century. The reality, however, is very different. The companies bringing in these facilities often received property tax abatements for years, meaning the local community absorbs the infrastructure costs without the tax revenue in return. One report found Oregon schools lost $275 million in potential tax income due to these abatements. And those jobs that were promised, well, they rarely materialised at the scale suggested. Even the largest data centres typically employ fewer than 150 permanent workers. The construction jobs are real and well paid, but they're temporary and many are filled from people outside the state. The true cost on communities has been baffling and we'll expand on this further in the next chapter, but we'll also see where the big data centre build out starts to fall apart. Now, you may not be transferring hundreds of billions to buy a data centre, but if you're looking for a simple and secure way to manage your finances, then Revolut is a good option. Revolut is an all-in-one finance app designed to simplify spending at home and abroad. Over 75 million people are already on it. You can get $40 when you sign up to Revolut and make your first purchase. The offer is available in multiple countries, so check out the link in the description or the QR code on screen to learn more. One of their unique features is how easy it is to generate single-use virtual cards, helping protect your payments whenever you're shopping online. Because whenever you're making an online payment, it generates a one-time card number that can't be reused. It massively reduces your risk of your card details being exposed online. On top of that, you can send or request money, make payments and split bills on the spot, even across different countries and currencies. It just removes a lot of friction day to day. And then there's the card itself. You can customize it in the app and add it straight to Apple Pay or Google Pay and make payments instantly, or upgrade to their metal plan for a premium-looking clean stainless steel card. Revolut lets you organize all your finances in one place, instead of juggling multiple apps, accounts and expensive fees. So again, if you want to check it out, you can get $40 when you sign up and make your first purchase. Just hit the link in the description or scan the QR code on screen to learn more. Thanks to Revolut for supporting ColdFusion. And now, back to the story. [00:07:46] Speaker 3: The AI bubble is about to burst because the hyperscalers, as they're called, have used up all their own cash and are now borrowing money. They are effectively net debtors at a time when they're spending huge volumes of money on data centers without having any guarantee that they're going to earn profits from that. And it's that worry that actually there isn't money to be made in that business, which is gradually dawning on investors. And that's going to be why they are going to pull away. In fact, there's already signs of people withdrawing from the AI sector because they fear that this is a bubble. It seems to be utterly inevitable. The question is only when. [00:08:30] Speaker 1: So what went wrong? Well, the answer is a few things all at once. So bear with me here as I try to get through them. At the top of the list is supply. Power is the first bottleneck for these data centers. They're extraordinarily power hungry. One single large hyperscale facility can consume as much electricity as a city of 200,000 homes. And the modern AI chips driving modern data centers, the GPU racks running models like Claude, Gemini, ChatGPT and Grok use significantly more power than the servers that came before them. According to Sightline Climate, around 25% of planned 2026 projects haven't even disclosed how they plan to power themselves, which to me is insane. To add to this, all the main electrical components need to run these facilities, the transformers, switchgear, batteries. They're all in critically short supply and the majority of which are imported from China. As Bloomberg puts it, America's data center build out hinges on Chinese imports. High power transformers from China surged from fewer than 1,500 units in 2022 to over 8,000 units in 2025. But with geopolitical tensions and tariffs disrupting supply chains, that pipeline has become unreliable. As one analyst put it, if one piece of your supply chain is delayed, your whole project can't deliver. Even skilled labor is a big issue. As this tweet puts it, "Meta can't hire fiber technicians fast enough, so now they're training them for free." And that's because they need the people fast. So basically, these tech giants aren't only short on equipment, but the people required to install them. The second big reason is the true cost of these data centers. The community suffers. There's some crazy stats and stories we're going to cover here. In Virginia, it's estimated that residents are already breathing in the exhaust gases of 10,000 diesel generators. In Utah, there was an approval for a 40,000 acre AI data center that would expel the heat equivalent of 23 atomic bombs into the surrounding valley every day. Upon its announcement, people went bananas. The owner of the data center, Kevin O'Leary, decided to shrink it to half the size, citing that he had no choice. According to the UN, data centers will use as much water as 1.3 billion people by 2030. The scale of some of these projects is baffling. A lot of people don't realize the magnitude of what's going on. [00:10:56] Speaker 6: The kind of AI data center that it's constructing now in Louisiana is nearly 400 times the size in footprint of the first data center that it built to support Facebook, its social media platform. And that data center is on track to be one fifth the size of Mannheim. And it is on track to use around five gigawatts of power, which is nearly the average power demand of London. That's it's it's one facility that would use the average power demand of a city like London. Wow. Meta wants to build more than one of these facilities. And of course, every single one of its competitors. [00:11:32] Speaker 1: As I say, that's just Meta. [00:11:33] Speaker 6: It's just one facility for Meta. [00:11:37] Speaker 1: Here's that Meta data center in person, so you can really see the scale. [00:11:40] Speaker 7: You can start to see the lights from back here. I knew they were starting this, but I didn't know it was this impactful. Look at all the semi trucks, all the cement trucks you'll drive by. And it just keeps going and going continues on for more and more. This Meta AI data center that they are putting in here that's taking up all this beautiful farmland that we will never be able to replace. It's only going to employ one to 500 people. So all this for about the same amount of jobs as my local Walmart. [00:12:18] Speaker 1: While writing this episode, I had a researcher from North Carolina reach out to me, and they were working on a data center documentary. This next clip will give you an idea of just how much permanent damage can be done to water via the cooling systems. [00:12:29] Speaker 8: How much water is this thing going to use? And he said 2 million gallons a day. [00:12:35] Speaker 9: Well, what about the contaminants, the forever chemicals? [00:12:39] Speaker 8: We do know that they have to put additives, just like your car radiator. You have to put antifreeze in there to keep it from overheating or keep it from freezing. The additives generally be things like PFAS that won't stick to metals, but also can't be removed from the water when it goes to a treatment plant. [00:12:59] Speaker 1: If you want to watch the documentary, I'll leave a link for it below. The data centers that did get built left a trail of grievances across the country. Residents living near facilities in Virginia, Georgia and Texas report a consistent low frequency hum from the cooling system. Noise that vibrates through walls and disrupts sleep. One resident in Manassas, Virginia spent $200,000 on insulation and new windows and still can't escape the drone. In Georgia, families started finding heavy sediment in their taps after construction began nearby. And if somehow you can get through a good night's rest, you might wake up to an envelope with a massive electricity bill. And that's just to ruin the rest of your day. The Georgia Power Utilities Company raised rates six times. That's right, six times between 2023 and 2025. A 24% jump. And that's because of data center demand. During that period, how many times do you think the wages of the residents in that area had risen? I can safely bet it's not six. In Oregon, Pacific power consumers saw a 50% increase in their bills since 2020. [00:14:05] Speaker 10: I think that we all collectively should just stop paying our electric bill. Because this is getting out of hand. There has to be something that we collectively can do. $814 for one month of electric? This is ridiculous. How do they expect people to be able to keep their lights on when the bill is half of my mortgage? [00:14:22] Speaker 1: Sometimes it feels like a rush to see who can destroy the world the fastest. [00:14:25] Speaker 11: You know, I think AI will probably like most likely sort of lead to the end of the world. But in the meantime, there will be great companies created with serious machine learning. [00:14:35] Speaker 1: And then we all know about the RAM issue. Chip markets have been thrown into complete chaos. I've already done an episode on this. But AI data centers are absorbing an estimated 70% of all global DRAM production capacity in 2026. And consumers suffer as a result. A standard 64GB DDR5 memory kit went from $190 to over $700 in just three months. And Sam Altman was to thank for that one. He promised to purchase 40% of global DRAM output from two manufacturers simultaneously, apparently without either one knowing about each other. When Micron learnt the commitment wasn't legally binding, its stock dropped 22% in a single day. So naturally, all of this has led to backlash. Data center cancellations due to local opposition quadrupled in 2025. According to intelligence from the platform Heatmap Pro, at least 25 projects were canceled that year due to community backlash, up from just six in 2024. And that number is accelerating. The will of the people may not last because the US government is watching this backlash closely, and the FBI has classified anti-AI sentiment as an emerging terrorist threat. [00:15:43] Speaker 12: If you've criticized AI or data centers, you may now be considered a domestic terrorist. US law enforcement warns of anti-tech extremism. The chaotic atmosphere that may result from emergent AI technology in the next five years may fuel large-scale protests that devolve into civil unrest and anti-tech violent extremist activity. So if you have ever protested AI publicly, you are now on a list. [00:16:10] Speaker 1: But despite the potential of ending up on an FBI watch list, people are fighting back. [00:16:15] Speaker 13: We are going to be fighting against data centers everywhere in New Jersey. [00:16:31] Speaker 1: A recent QUNIPIAC survey found 65% of Americans now oppose data centers being built in their communities. Maine became the first US state to pass a statewide ban on data center construction, prohibiting new builds until late 2027. 13 other states are now considering similar measures. This has become a kitchen table issue. Rising electricity bills, water use, and a sense that decisions are being made behind closed doors without any community input. [00:16:59] Speaker 9: One small town voted to prevent an AI data center from being built, but construction went ahead anyway. The town was sued by the developer, and a small town like Saline Township wouldn't be able to defend itself in a lawsuit. [00:17:13] Speaker 14: Illinois just approved a data center the size of 600 football fields. It will consume more than half the amount of electricity as the entire city of Chicago. But the silver lining to this is that hundreds of community members came out to say no against this proposal. But unfortunately, the city just didn't give a f**k and approved it anyway. Remember, these are the same people that tell you to take shorter showers. [00:17:38] Speaker 1: And this all isn't just happening in the US. In Australia, where I live, for some reason, we're planning to be the second largest data center hub in the world. [00:17:46] Speaker 15: A quarter of Sydney's drinking water will soon be needed to power 270 new AI data centers. Australia is getting set up to become the second biggest data center location in the world. [00:18:00] Speaker 6: I'm in West Fitzgray, and this is just one of the many data centers that have popped up around the area. We call it mortal because it's so imposing over the whole neighborhood. You can literally see it from every corner. [00:18:10] Speaker 16: Australia's data centers are projected to use more new electricity than all of the country's homes and EVs over the next 15 years, and they have the potential to drive up power prices. [00:18:19] Speaker 1: It's turned opposition for data centers into a universal rallying cry, even in today's political landscape. Microsoft, meanwhile, has slowly cancelled or deferred up to two gigawatts of planned data center capacity globally. Analysts at TD Cowan described the pullbacks as pointing to, quote, "data center oversupply relative to current demand forecasts," end quote. The famous Oracle OpenAI Stargate campus, the one in Texas, reportedly quietly stalled its expansion amid ongoing supply and financial complications. Of course, they had to do it quietly because the speculation of the dreaded B word is just around the corner: the bubble. So is the bubble popping? Eh, kinda. Probably not in the dramatic 2008-style collapse sense, but it's definitely a strong reality check. I've touched on this a few times in previous episodes, but there's a structural problem here. These investments are being made on the assumption that AI demand will grow exponentially and indefinitely. It's still a big bet because the infrastructure spending is very real, but the returns? They're still speculative. For example, something as simple as open source models could derail the whole thing. Who would pay $20 or $200 a month to get something that's only 20% better than an open source model that's free? If there's much cheaper or free models that are 80% as good, companies spending $1 trillion to train frontier models seems ridiculous in comparison. [00:19:46] Speaker 2: We are talking about an industry that has absorbed over a trillion dollars in the last three years, the media attention from everyone, and these continual stories about the magic of AI. But when you say, "Hey, is it making any money?" They go, "Oh, no. Oh, we couldn't possibly. We don't do that yet, but don't worry. Uber lost a lot of money. Uber burned about $32 billion, which is less than half of what Anthropic has raised in the last six months." But all of this hasn't stopped the talk of building [00:20:13] Speaker 1: data centers in space, but that's a story for another day. In terms of financing, the situation has become uncomfortable. It does echo previous cycles. Around $34 billion in data center bonds were issued recently, with 84% of them being rated A. That's a very safe level, low enough risk for pension funds to buy. But there's something fishy here. Those A-rated bonds are paying eight, nine, or sometimes 12% interest. That is junk bond level rates, and that implies there's a huge amount of risk. It's risk that credit ratings haven't acknowledged yet. And now, where have we seen this before? Hint, it rhymes with blue thousand and eight crisis. But this time, though, it's not systematic in banks, but more so private investors and groups. The AI buildout assumed infinite political tolerance, infinite grid capacity, infinite supply chains, and infinite community patients. None of these assumptions held. At this point, this is what the AI data center build-out kind of feels like. [00:21:14] Speaker 13: And in tech news, OpenAI launched construction of a new data center in Greenville, Tennessee, on top of a sick child today. The artificial intelligence company announced the data center will be positioned in the heart of the small race-car-themed bedroom where eight-year-old Billy Treaker fights a rare kidney disease on a daily basis. Engineers and surveyors began laying the groundwork for the data center this week, which they say will unleash up to 10 gigawatts of power annually, once fully up and running, on top of Stan and Rebecca Treaker's only child. OpenAI executives say they plan on expanding out from Billy's sleeping area to his play and reading areas, as well as to the corner of his room where his mother sits in a rocking chair praying for her son's survival every night, all to guarantee ample space for their 10-ton processors to operate. [00:21:59] Speaker 1: To further understand how badly the buildout is going, here's a quick story. In June 2025, Fermi America, co-founded by former U.S. Energy Secretary Rick Perry, announced Project Matador, later rebranded as the President Donald J. Trump Advanced Energy and Intelligence Campus, a 17-gigawatt AI megaproject in the Texas panhandle. It was one of the most ambitious data center proposals ever announced. But in less than a year, the CEO had resigned, then the CFO followed two days later. The company still had no confirmed anchor tenant. That's the client who's going to be leasing the data centers. It's a basic prerequisite for any data center business. Its market cap collapsed from nearly $20 billion to $3.4 billion. The departing CEO admitted that he may have, quote, "misunderstood where the supply chain is," end quote, on cooling infrastructure. Quote, "I will accept that as a failure," he said. If the sitting President of the United States can't get his named data center off the ground, it tells you something for the broader state of the industry. If compute capacity keeps getting constrained, that can be cancellations, community bans or supply chain issues. The companies that depend on it, like OpenAI and Anthropic, face a very real growth ceiling that has nothing to do with the quality of their models. There's even speculation that that's why when a new model releases, it performs relatively well, but after a few months it gets throttled down to save on compute costs. Now, before we finish, this wouldn't be complete without the other side of the coin. Are there sustainable ways to build data centers? Or do we even need massive data centers for AI at all? Well, for the first question, there is such a thing as subsea data centers. Underwater data centers almost sounds like something in science fiction, but they already exist in China. It includes deploying container-like capsules on the ocean floor, using the naturally cold sea water as a passive cooling system. One operator reports that 99% of the electricity goes directly to computing, compared to roughly 50% in traditional air-cooled facilities. Do we even need big data centers? Recently, local AI models have become more common. Instead of using a billion dollar data center, just use smaller models that run locally on your device. If it's 80% as good, then most users will be satisfied for simple tasks. Apple's large unified memory is a leading example of the kind of local hardware that enables this. It could be that in three years, when most laptops can run decent models, the demand for centralized infrastructure drops. The data center doesn't disappear, but it gets smaller and less intrusive. And then there's the argument that in the future, all mainstream LLMs will become a commodity. Even Larry Ellison hints at this, although his solution is, well, typical of him. [00:24:43] Speaker 17: So if you look at ChatGPT or Anthropic Rock, Llama, what have you, they're all trained on all of the data on the internet. In other words, publicly available data. But for these models to reach their peak value, you need to train them not just on publicly available data, but you need to make private, [00:25:08] Speaker 1: privately owned data available. And critically, community engagement and transparency need to become foundational. It can't be an afterthought. The backlash happening right now, at its core, is a failure of trust. So what do we make of all of this? So I've been covering AI on this channel for over 10 years now. And despite the recent appearance in the public zeitgeist, my stance on it has always been the same. AI can be revolutionary if used correctly, simplifying maths and cutting-edge physics, recognizing patterns and data that humans might miss, making repetitive tasks easier. But it's certainly not a silver bullet in all use cases. So if we're raising electricity prices for hardworking communities, so someone on X can generate a meme by typing a few words, or other people can feel the internet with slop, well, call me crazy, but that doesn't sound like a great trade-off to me. Don't get me wrong, we've always needed data centers. Before AI, they were quietly running everything from our bank accounts to our streaming services. The truth is they're not going away, and they probably shouldn't. But the question isn't whether we should build them at all. It's how, and it's also a question that's a function of, is the amount being built really worth the payoff? It's something to think about. So what do you guys think? What are your views on data centers? And where do you see the future of all of this going? Anyway, that's about it from me. Thanks so much for watching all the way through to the end. Really appreciate it. So my name is Dagogo, and you've been watching ColdFusion, and I'll catch you again soon for the next episode. Cheers, guys. Have a good one. [00:27:06] Speaker 3: ColdFusion. It's new thinking.