The Quiet Collapse of AI Data Center Projects

[00:00:00] Speaker 1: It is January 2025. Microsoft has just announced that it will spend $80 billion on AI data centers this year alone. [00:00:07] Speaker 2: Microsoft is expected to spend some big money on artificial intelligence this year. The company said it's on track to invest approximately $80 billion to build out data centers. [00:00:16] Speaker 1: $80 billion in one year on servers and cooling systems, concrete and electrical infrastructure. The CEO says the company doesn't have enough capacity. The CFO says they're working from a capacity-constrained place. The message from Redmond is clear. Build more. Build faster. The demand is coming, and we cannot afford to miss it. Six weeks later, Microsoft starts canceling leases. Not trimming. Not slowing. Canceling. Walking away from agreements worth hundreds of megawatts of computing capacity with at least two private operators. Then more cancellations. Then expired letter of intent deals on sites totaling over a gigawatt. Then withdrawal from at least five land parcels across multiple major markets. By March, TD Cowan analysts confirmed that Microsoft had quietly walked away from data center projects totaling more than two gigawatts of planned capacity across the United States and Europe. For context, two gigawatts of electricity is enough to power 1.6 million homes. According to TD Cowan analysts, the pullback is attributed to an oversupply of computing capacity. Amazon followed in April. Not cancellations of signed deals, but a pause on leasing negotiations, particularly overseas, that Wells Fargo analysts described as a pullback similar to Microsoft's. The largest cloud company in the world, which had been on track to spend more than $100 billion on AI infrastructure in 2025, was suddenly being more cautious about what it actually committed to. And now, as of April 2026, Bloomberg reports that close to half of planned U.S. data center builds are delayed or canceled. Half. So what on earth is happening? Because three years ago, these same companies couldn't build fast enough. And now they're walking away. To understand this, you have to go back to November 30th, 2022, the day ChatGPT launched. ChatGPT, maybe you've heard of it. If you haven't, then get ready, because this promises to be the viral sensation that could completely reset how we do things. Before that date, AI was something that existed in academic papers, niche research labs, and breathless tech conference keynotes. After that date, it was something 100 million people were using within two months, and every major corporation in the world was asking the same urgent question. What is our AI strategy? Wall Street saw the panic and made it worse. The magnificent seven tech companies, Microsoft, Google, Meta, Amazon, NVIDIA, Tesla, and Apple, collectively added trillions in market capitalization on the bet that whoever built the most AI infrastructure fastest would own the next decade of technology. NVIDIA's stock went from $150 to over $900 in 18 months. A single chip company became the third most valuable business in human history. And the hyperscalers, the Microsofts and Googles and Amazons of the world, responded to the market pressure the way they always respond to market pressure. They committed capital, enormous amounts of capital, ahead of any realistic demand forecast. It is estimated that the four largest hyperscalers, Alphabet, Amazon, Meta, and Microsoft, had collectively committed to spending approximately $650 billion on AI infrastructure in 2026. And then DeepSeek happened. On January 27th, 2025, four weeks after Microsoft announced its $80 billion commitment, a Chinese AI lab released a model that performed comparably to the most advanced US AI systems. The cost to train it, $5.5 million. OpenAI reportedly spent $100 million to train GPT-4. NVIDIA lost $593 billion in market value in a single trading session, the largest single-day market cap loss in stock market history. [00:03:53] Speaker 3: DeepSeek appears to have found a more efficient way to power AI. Was that a moment of rethink for you? And are you doing anything differently now? [00:04:02] Speaker 4: I think the DeepSeek team is very talented and did a lot of good things. I don't think they figured out something like way more efficient than what we figured out. [00:04:07] Speaker 3: But do you think there is a more efficient way to build an AI? [00:04:11] Speaker 4: Oh, I'm sure. We have made incredible efficiency strides year over year, and I'm sure we'll keep doing that in the future. [00:04:18] Speaker 3: So if that's the case, why are you building all this? [00:04:22] Speaker 4: If we had an AI that we could offer at one-tenth of the price of current AI, I think people would use it 20 times as much. And we would still need twice as much compute to satisfy the then-current demand. [00:04:32] Speaker 1: What DeepSeek proved wasn't that AI was fake. It proved something more specific and more uncomfortable. That the assumption underpinning hundreds of billions in infrastructure investment might be wrong. The assumption was this. More computing power equals better AI. Train on bigger clusters, run on more chips, build more data centers, and get ahead in the race. That logic justified every lease, every land parcel, every gigawatt of planned capacity. It was the entire rationale for the build-out. DeepSeek suggested that smarter algorithms, not bigger machines, might be the actual differentiator. And if that's true, then the amount of raw computing power you need is substantially smaller than the hyperscalers we're building for. Well, that's a problem when you've already signed leases on two gigawatts you suddenly don't need. But here's the thing about the Microsoft pullback that gets lost in the drama of the cancellations. It's not just about DeepSeek. It's not just about oversupply. There are actually three separate problems hitting AI data center investments simultaneously. And they are all reinforcing each other. The first is the demand problem. The gap between AI infrastructure spending and AI-generated revenue continues to widen. It is estimated that 95% of AI projects funded by external investors do not generate revenue. The MIT study cited earlier found that only 5% of enterprise AI implementations positively impacted revenue. You can build the largest data centers in history, but if the applications running on them aren't generating enough income to justify the cost, the financial math eventually forces a reckoning. Microsoft's internal situation is illustrative. The company had been building capacity partly to handle incremental AI training workloads from OpenAI, the company it had invested $13 billion in. When Microsoft and OpenAI renegotiated their agreement in early 2025, allowing OpenAI to use cloud services from other providers, a significant chunk of the demand Microsoft had been building disappeared overnight. Leases signed against a demand forecast that no longer existed became obvious candidates for cancellation. The second problem is power, and this one is more structural than anything related to AI demand. It is estimated that by April 2026, close to half of all planned U.S. data center builds have been delayed or canceled, but not because companies change their minds about AI, because they can't get electricity. The components required to connect a data center to the power grid, transformers, switchgear, and battery systems, have lead times of two to four years. There is a global shortage of large power transformers. A significant portion of those components was manufactured in China, and the tariff regime now in place has made Chinese-sourced electrical infrastructure substantially more expensive and less available. You can build the building, you can buy the NVIDIA chips, but if you can't get a transformer to connect the facility to the grid, the whole investment sits idle. A single missing transformer can hold up a $2 billion data center project for a year or more. The irony of the AI build-out running into a shortage of century-old electrical infrastructure is genuinely hard to overstate. The third problem is the economic environment. When companies start worrying about a recession, they don't cancel core operations, but they do get more cautious about discretionary capital commitments. And a lease on a data center in Indonesia that won't come online for two years is exactly the kind of commitment that looks less urgent when the economic outlook turns uncertain. And here's where the story gets complicated, because the pullback is real, but it is not uniform. While Microsoft and Amazon were quietly walking away from leases, Google and Meta were moving in the opposite direction. When Microsoft abandoned data center leases in Europe, Google picked them up. When Amazon paused international negotiations, Meta raised its 2025 CAPEX forecast to $72 billion. Google reaffirmed $75 billion. Both cited growing AI demand and rising hardware costs from tariffs as reasons to spend more, not less. What this actually tells you is not that the AI infrastructure boom is over. It tells you something more precise. The hyperscalers that overcommitted fastest, Microsoft chief among them, are recalibrating. The ones that were more conservative in their initial commitments are now in a stronger position to expand. It is the oldest story in any investment cycle. The people who moved fastest made the biggest bets. Some of those bets were right, some were not. And the ones that were not are now the canceled leases. Satya Nadella, Microsoft's CEO, said it himself in February 2025, before the full scale of the cancellations was public, that there will be an overbuild of AI infrastructure. He said his company would be leasing a lot of capacity in 2027-2028, which is a very polished way of saying, we overbuilt and now will buy from whoever is sitting on the excess. Not exactly the confident AI is the most transformative technology in human history energy of the January announcement. But here is the honest caveat that goes with all of this. The fact that Microsoft walked away from 2 gigawatts of planned capacity does not mean AI is a bubble that is about to pop. Google and Meta stepping in to absorb the free leases suggest the underlying demand for AI infrastructure is real, just not as uniformly real as the investment cycle of 2022 to 2024 implied. What the pullback actually reveals is that the first phase of AI infrastructure investment was driven substantially by fear. Fear of being left behind. Fear of the competitor who built faster, getting a permanent moat. Fear of the market punishing any company that appeared to be under-investing in AI. That fear drove commitments that were disconnected from actual demand forecasts. And when the fear subsided, when DeepSeq showed that more compute is not necessarily better compute, and when earnings calls started requiring executives to justify $80 billion capital expenditures with actual revenue, the commitments that fear had created started getting walked back. It is estimated that the four largest hyperscalers will still spend more on AI infrastructure in 2026 than in any previous year. The buildout is not stopping. It is correcting. The difference between a correction and a collapse is whether the underlying demand is real. And the underlying demand for cloud computing, for AI inference at scale, for the infrastructure that runs the models already deployed across millions of enterprise applications, is real. What was not real was the assumption that you needed to build for every possible future scenario at the same time, on the same timeline, at any cost. That assumption is now sitting on a lot of canceled lease agreements in Ohio, Indonesia, and the UK. Thank you for watching. For more breakdowns like this, please consider subscribing, and let me know your thoughts in the comments.