50% Of AI Data Centers Have Quietly Been Cancelled Or "Delayed"
How Money Works · 2026-04-16
💡 Quick Take
1. AI infrastructure spending is astronomical, reaching $400 billion in 2025 alone, dwarfing historical projects.
2. Despite massive spending, not a single AI company (except hardware suppliers like Nvidia) is profitable yet.
3. There's a major disconnect between announced data center capacity and actual construction progress.
4. Nvidia is shipping significantly more GPUs than the current operational AI data center capacity can utilize.
5. Power infrastructure is the biggest bottleneck for new data centers, not necessarily the chips themselves.
6. The bullwhip effect is driving companies to buy components like chips and power supplies preemptively, even without immediate use.
7. Nvidia's inventory has more than doubled, suggesting potential oversupply or upstream supply chain issues.
8. Soaring energy prices, exacerbated by geopolitical events, are making data centers increasingly expensive to operate.
9. Companies are depreciating AI chips over 6 years, while their actual operational viability might be closer to 3 years.
10. Stretched depreciation accounting can mask the true profitability of AI investments, potentially inflating demand for GPUs.
11. The financing landscape for data center projects is becoming more challenging due to issues in private credit markets.
📊 Detailed Explanation
1. AI infrastructure spending is astronomical, reaching $400 billion in 2025 alone, dwarfing historical projects. This is a mind-blowing figure! The transcript highlights that in 2025, the world's largest companies are sinking around $400 billion into capital expenditures just to build the infrastructure for AI. To put that into perspective, that's equivalent to nine Manhattan Projects or two Apollo Programs, all happening in a single year, and that's *just* for the physical setup like data centers. It doesn't even include the costs of staffing, energy, security, or strategic acquisitions. It's an unprecedented surge in investment focused purely on the foundational elements of AI.
2. Despite massive spending, not a single AI company (except hardware suppliers like Nvidia) is profitable yet. This is a huge point of concern! Even after nearly four years since ChatGPT burst onto the scene, the companies developing and deploying AI haven't figured out how to make money from it. The transcript explicitly states that even with optimistic financial projections and accounting maneuvers, profitability remains elusive. The only ones consistently cashing in are the upstream players, the hardware suppliers and chip manufacturers, most notably Nvidia. They're the ones selling the "pickaxes and shovels" in this gold rush.
3. There's a major disconnect between announced data center capacity and actual construction progress. This is where things get really interesting and a bit fishy. While companies are announcing massive amounts of new data center capacity coming online (measured in gigawatts), on-the-ground research shows that a significant portion of these projects are either delayed or outright canceled. The transcript points to reports indicating that over half of the sites planned for this year have faced delays or cancellations. Even when projects are "under construction," it can mean anything from a foundation being poured to final fit-out, making the actual progress much slower than the press releases suggest.
4. Nvidia is shipping significantly more GPUs than the current operational AI data center capacity can utilize. This is a critical paradox. Nvidia's CEO, Jensen Huang, claimed they were shipping around 10 GW of GPUs in 2025. However, estimates show that the total operational AI data center capacity on the planet is only around 7.7 GW. Even if you factor in all the data centers under construction and assume every existing AI data center upgrades its hardware, there simply isn't enough demand to absorb Nvidia's current production rate. This suggests Nvidia is either overestimating production or the GPUs are ending up somewhere that's not easily visible to analysts.
5. Power infrastructure is the biggest bottleneck for new data centers, not necessarily the chips themselves. This is a crucial shift in understanding the AI infrastructure challenge. The transcript explains that the primary constraint for building new data centers isn't the availability of advanced computer chips, but rather the electrical infrastructure needed to support them. Power has become the real-world limitation. Components like transformers and power supplies are experiencing doubled prices and supply struggles, making them a critical choke point.
6. The bullwhip effect is driving companies to buy components like chips and power supplies preemptively, even without immediate use. This explains the seemingly contradictory situation of high spending alongside slow progress. Because supply chains for essential components like power infrastructure are so limited, companies are compelled to buy whatever they can, as soon as they can, even if they don't have immediate use for it. They fear going to the back of the line if they wait. This hoarding behavior, known as the bullwhip effect in industrial planning, inflates demand and spending, even as actual project completion lags.
7. Nvidia's inventory has more than doubled, suggesting potential oversupply or upstream supply chain issues. This is a red flag for Nvidia. The transcript notes that Nvidia's inventory more than doubled from the previous year and quadrupled from 2024. If there was truly insatiable demand, their inventory shouldn't be growing so rapidly. This suggests it's becoming harder to move chips, or more likely, Nvidia itself is facing upstream supply chain problems and is stockpiling components with the confidence they'll eventually sell them.
8. Soaring energy prices, exacerbated by geopolitical events, are making data centers increasingly expensive to operate. The cost of powering these massive AI facilities is becoming a major hurdle. The transcript points out that higher energy prices, partly driven by events like the war in Iran, are significantly cutting into the viability of running data centers. Whether they're pulling from the local grid or using their own natural gas generators, the cost of energy has doubled, making it a huge operational expense and potentially turning last year's cutting-edge hardware into e-waste.
9. Companies are depreciating AI chips over 6 years, while their actual operational viability might be closer to 3 years. This is a sneaky accounting practice that distorts the financial picture. The transcript highlights that the industry standard for depreciating GPUs is six years, but in reality, they might only be operationally viable for about three years before newer, better models render them obsolete. This practice allows companies to offset taxes and make their current profitability look better than it truly is.
10. Stretched depreciation accounting can mask the true profitability of AI investments, potentially inflating demand for GPUs. This directly ties into the previous point. By stretching out the depreciation period for AI chips, companies can artificially boost their reported profits. This creates a more favorable narrative around AI investment, which in turn fuels investor hype and, consequently, the demand for more GPUs from companies like Nvidia. If these companies were forced to use more honest accounting, their profits would look worse, potentially dampening the AI investment frenzy.
11. The financing landscape for data center projects is becoming more challenging due to issues in private credit markets. The money tap might be starting to tighten. The transcript mentions that private credit companies, which have been crucial financiers for major data center projects, are now facing their own industry-wide problems. This is likely to make it harder for them to provide the easy financing that has fueled so much of the current AI infrastructure build-out.
🎯 Expert Opinion
Wow, this transcript lays out a fascinating, and frankly, a bit concerning picture of the AI infrastructure boom! From an expert's standpoint, what we're seeing here is a classic case of a market getting ahead of itself, fueled by immense hype and a dash of speculative frenzy. The sheer scale of capital expenditure on AI infrastructure is unprecedented, and while it signifies a massive belief in the future of AI, the lack of immediate profitability for most companies is a huge flashing red light. It's like everyone's building the fanciest hotels in a desert without a clear plan for how to attract enough guests to make them pay the bills.
The disconnect between announced data center capacity and actual construction is a huge red flag. This isn't just about delays; it points to fundamental issues in execution and potentially over-promising. The "bullwhip effect" is a real phenomenon, but when it's driven by fear of missing out on scarce resources like power infrastructure, it can create artificial demand that masks underlying weaknesses. We're seeing companies essentially hoarding components, which distorts the true market demand and makes it harder to assess where the actual growth is happening.
Nvidia's position is, of course, central to all of this. They're the kingpins of the current AI hardware landscape, but the rising inventory levels and the questions about chip utilization are critical. It suggests that the "picks and shovels" analogy might be getting a bit strained. While they're still incredibly profitable, the sustainability of their growth hinges on the actual adoption and monetization of AI by their customers, not just the construction of data centers. The fact that their inventory is ballooning while they're still claiming massive shipments hints at a potential oversupply scenario brewing, or at the very least, a significant lag between production and actual deployment and use.
The power and energy constraints are perhaps the most grounded and enduring challenges. These aren't going to disappear overnight. The electrical grid simply wasn't designed for this level of concentrated, high-demand computing. This means that the pace of AI development will, to a significant extent, be dictated by our ability to upgrade and expand our energy infrastructure, which is a slow and complex process. The rising energy costs are a direct consequence, and they are going to force a reckoning for data center operators. We'll likely see a greater push for energy efficiency, alternative power sources, and perhaps even a more distributed approach to AI computing.
The accounting practices around chip depreciation are particularly concerning. This is a classic example of how financial reporting can be manipulated to create a rosier picture. If companies are significantly overstating the useful life of their AI hardware, it means their true profitability is lower, and the demand for new hardware might be artificially inflated. This can lead to a painful correction when the market realizes the actual economics. We're likely to see increased scrutiny from regulators and investors on these depreciation schedules. The implication is that the "AI revolution" might be more capital-intensive and less immediately profitable than the current market narrative suggests.
Finally, the tightening in private credit markets is a significant development. For years, easy money has flowed into these capital-intensive projects. If that flow slows down, it will undoubtedly put a brake on the pace of new data center construction and AI infrastructure build-out. This could lead to a consolidation in the market, with only the most robust and well-capitalized players surviving. We might also see a shift in investment strategies, with a greater emphasis on operational efficiency and proven business models rather than pure speculative growth.
In essence, the AI infrastructure boom is a high-stakes game of chicken. The players are betting big on a future that's still being built, and the current financial and logistical realities are starting to catch up. While the long-term potential of AI is undeniable, the path to profitability and sustainable growth for the companies building the infrastructure is fraught with challenges. We're likely to see a period of significant recalibration and a more pragmatic approach emerge as the initial hype subsides and the hard economics of AI infrastructure come into sharper focus.
⚠️ This content is not investment advice.
Kanal: How Money Works