50% Of AI Data Centers Have Quietly Been Cancelled Or "Delayed"

0:00 In 2025, the world's largest companies reportedly spent around $400

0:04 billion on capital expenditures to support

0:06 the development of artificial intelligence.

0:08 Adjusted for inflation,

0:09 that would be nine Manhattan projects or two Apollo programs,

0:13 all within the space of just one year and just on the infrastructure alone.

0:17 Put another way, last year,

0:19 more money was set aside for constructing and fitting out data centers

0:23 than was spent on building single family residential homes over the same time.

0:27 And this number doesn't even include

0:29 non-public companies like Enthropic or OpenAI,

0:31 which are harder to get reliable financial data on.

0:34 This number also doesn't include any of the other costs

0:37 outside of just building and fitting out the facilities themselves,

0:40 like staffing, energy, security, and uh strategic acquisitions like podcasts.

0:45 These numbers are also only for 2025.

0:48 And of course, recent announcements suggest that spending

0:51 this year will once again break new records.

0:54 Now, the borderline comical numbers being thrown around in the AI

0:58 industry may not be that surprising to any of you anymore,

1:01 but it has also been almost 4 years since this technology has

1:04 really come onto the scene with the first public release of Chat GPT.

1:08 In that time, not a single one of these companies

1:10 has figured out how to turn a profit with this technology,

1:13 even when using generous financial projections and accounting tricks.

1:16 The exception to this, of course, has always been Nvidia alongside the other

1:21 hardware suppliers and chip manufacturers upstream of them.

1:24 The classic analogy that you are probably sick of hearing by now

1:27 is that all of this may very well be an unsustainable gold rush.

1:30 But the hardware companies are the ones making

1:33 reliable profits by selling the pickaxes and the shovels.

1:36 However, by following the numbers,

1:38 it has raised some questions about where these shovels are actually ending up.

1:42 At the same time, these companies are promising

1:45 record levels of new spending on data centers.

1:48 Reports have indicated that over half of the sites that were supposed

1:51 to be open this year have either been delayed or outright cancelled.

1:55 Logically, it's very difficult for both of those things

1:57 to be true at the same time.

1:59 Even if you completely ignore the question

2:01 of how their end customers are going to keep

2:03 on paying for this, there are some other

2:06 concerning logical paradoxes developing at the same time.

2:08 These companies can't keep up with demand for new chips.

2:11 And yet, their inventories are growing.

2:13 They are depreciating their hardware over six years while

2:15 claiming next year's models will render this year's completely obsolete.

2:18 And then there is a question of how they plan

2:21 to power all of this given the uh current state of everything.

2:26 Nearly half of the US data centers planned

2:29 for 2026 are reportedly expected to be delayed.

2:32 Push back from communities who say they

2:34 don't want data centers in their backyards.

2:37 10 natural gas plants serving this one single data center.

2:40 access demand kind of being soaked up by by data centers.

2:44 We never said we were going to invest a hundred billion dollars in one

2:48 realm that never will set once we've

2:50 built this sort of generally intelligent system.

2:52 Basically, we will ask it to figure out

2:54 a way to generate an investment return for you.

2:57 Okay.

2:57 So, whether we really like it or not,

2:59 Nvidia has become the company holding up the world thanks to its market cap,

3:03 profit, and reinvestment into feeding the industry that feeds it.

3:06 As potentially the single most analyzed

3:08 company in the history of financial markets, I hate to break it to you,

3:11 but a YouTube video isn't going to unear

3:14 something that everybody else was missing all along.

3:17 However, there are three big questions that are being asked about

3:20 Nvidia's place in the current market that are worth understanding because, well,

3:23 again, even if you aren't personally invested in this company,

3:27 any significant change in its price

3:29 will have implications for the entire economy.

3:32 So, the first big question is where are all of these chips actually going?

3:36 In an interview with CNBC last year,

3:38 Jensen Huang claimed that the company was shipping

3:41 around 10 GW of GPUs within 2025 alone.

3:43 Some quick math based on their current product lineup

3:46 and reported annual sales suggests this might be a slight overestimation,

3:50 but it is roughly in the right ballpark.

3:53 Now, this figure was actually taken from an interview

3:55 where Huang was announcing their partnership with OpenAI to invest

3:58 up to hund00 billion in the company in order

4:00 to help them build out over 10 GW of compute themselves,

4:03 which was again roughly the equivalent of Nvidia's entire annual output of GPUs.

4:08 This, of course, didn't help the accusations of circular dealing.

4:12 But that's actually not the problem here.

4:14 The problem was that according to estimates by Goldman Sachs,

4:17 there are only around 7.7 GW of AI

4:20 data centers currently in operation across the entire planet.

4:23 Now, of course, there are a lot of new data centers under

4:27 construction that will all need their own complement of fancy Nvidia GPUs,

4:31 but there aren't nearly as many as the big announcements might make you think.

4:35 On the ground research performed by the market intelligence firm

4:38 Sighteline Climate confirmed suspicions first raised by the business journalist

4:41 Ed Zitron that a lot of data center construction wasn't

4:44 nearly as far along as the press releases would suggest.

4:47 Of the 21.5 GW of announced capacity expected to come online before 2027.

4:53 Only 6.3 gawatt was actively under construction.

4:56 And even that makes it sound better than it really is.

4:59 under construction could be anything from a data

5:01 center that is getting its final fit out before going online to a site

5:04 that has had nothing more than a foundation poured.

5:07 Now, I will leave a link to all of that investigative research below,

5:10 as well as some interviews that Zitron has done here on YouTube.

5:13 Even as we were putting this video together,

5:16 Bloomberg reported that Oracle and OpenAI's

5:18 flagship Stargate Data Center in Abalene,

5:20 Texas shoved its expansion plans amongst ongoing issues

5:22 that we will get into later in this video.

5:25 Now, in the interest of full transparency, across the wider market,

5:28 some of these numbers are hard to verify because they are coming

5:31 from a mix of private companies that do not need to provide

5:34 public accounts and large public companies that can mix in their AI

5:37 related operations with the rest of the business on their financials.

5:40 But the point is that unless every data center

5:42 on Earth was replacing its graphics cards every 14 months,

5:45 Nvidia's current rate of production would be over supplying

5:48 the real market that actually exists right now, which sounds bad.

5:51 But it gets worse.

5:53 Even if we generously assume that every data

5:55 center verifiably under construction right now goes online

5:58 within the next 8 months and every existing

6:00 data center updates their current hardware this year,

6:03 it's not quite as simple as just adding these two

6:05 numbers together to make 14 GW of total demand.

6:08 There is no hard and fast industry standard.

6:10 But since tech companies have started using megawatts

6:13 and gigawatts to measure the size of their data centers,

6:15 that has typically included the entire input of the center,

6:19 not just energy going exclusively into the GPUs.

6:22 In addition to this, there is also the networking,

6:25 cooling, storage, and processing overhead.

6:27 According to the International Energy Agency,

6:29 a typical data center normally dedicates around 46

6:32 to 65% of its energy to just the compute.

6:36 So, even if we again take the generous high-end estimate and apply it

6:40 to the generous estimate of every data center

6:42 under construction going online this year and then

6:45 generously assume every existing AI data center

6:47 will upgrade its hardware and then generously assume that every single last one

6:51 of these facilities are going to use Nvidia GPUs,

6:54 there just isn't enough of a gap to cram 10 gawatt worth of GPUs into.

6:59 So either Jensen Huang was overestimating how many cards they actually produced

7:03 last year or they are ending up somewhere that analysts can't see.

7:07 And oh yeah, that's just the first problem.

7:09 So it's time to learn how money works to find out where all

7:13 of this money and all of these computer chips are actually ending up.

7:16 This video is sponsored by Monarch Money.

7:19 Monarch Money is a private ad-free personal finance platform

7:22 that brings all your accounts into one clean dashboard.

7:25 It automatically tracks your spending and net worth

7:28 and helps you stay financially aligned without juggling spreadsheets.

7:31 Instead of just showing where your money went,

7:33 Monarch makes it easy to see the full financial picture,

7:35 make smart adjustments, and make progress towards your goals with confidence.

7:39 Plus, with bank level encryption and readonly access,

7:42 it's secure and thousands of users have given it a 4.9 star rating.

7:46 I started using Monarch for my monthly

7:49 budget check-ins and it's been a gamecher.

7:51 All my accounts, cards, loans,

7:53 and investments sync from over 13,000 institutions

7:56 into one beautiful view on web or mobile, saving me time hopping between apps.

8:01 I love the flexible budgeting that separates fixed

8:04 and flexible spending with customizable rules and categories.

8:07 And the insights explain spending changes

8:09 or unusual transactions in plain English,

8:11 helping me stay aware without constant alerts.

8:14 Seeing my net worth trend and goal timelines for saving or paying

8:17 down debt makes progress feel real and motivating without the stress.

8:21 Try Monarch yourself.

8:23 Go to the link in the description and use

8:27 code HMW50 for 50% off your first year.

8:34 Okay, so one thing that may start to sound

8:37 a little bit confusing when seeing the headlines and the press

8:39 releases about these new data center developments is why

8:42 they are measured in gigawatts in the first place.

8:45 A watt is a measure of power.

8:46 And when it comes to computers, not all watts are created equal.

8:50 If you were building a gaming computer,

8:52 its power draw would probably not be the first spec you focus on because,

8:56 well, it just makes no sense.

8:58 My current Mac Mini desktop draws about 15 watts,

9:01 but in almost every computing task,

9:02 it would absolutely destroy my 150W 15-year-old gaming computer that I

9:06 just don't have the willpower to upgrade for um unrelated reasons.

9:10 Now, I am not bringing this up to rant about the cost of RAM,

9:15 but instead because it relates to the second major problem facing Nvidia

9:19 and the expectation that it will keep on delivering more chips every year,

9:23 and that is this power limitation.

9:25 According to industry analysis,

9:26 the biggest bottleneck facing new data centers today is not

9:30 necessarily in the advanced computer chips to run their models,

9:33 but rather in getting the electrical infrastructure to support it.

9:36 Power has become the realworld constraint.

9:38 So that's how these projects are getting compared

9:41 in the same way we might compare FPS in Cyberpunk 2077.

9:44 In a typical data center, the transformers, power supplies, generators,

9:48 and connecting wires typically take up less than 10% of the total build cost.

9:53 But nothing can happen until those components are in place.

9:56 The price of things like a Prolle transformer have

9:59 more than doubled over the last 4 years and supply

10:02 struggles to keep up with this new demand on top

10:04 of supplying regular energy grids for regular households and businesses.

10:07 It hasn't helped that a lot

10:09 of these technical components have typically come from China,

10:12 South Korea, Mexico, and Canada, making tariff disruptions harder to navigate.

10:17 What this has meant is that since the supply of everything

10:20 that goes into these data centers has been so limited,

10:22 companies are buying everything they can as soon as they can,

10:25 even if it takes a long time to actually be put into use.

10:29 If Nvidia says the latest batch of HB200's is ready to purchase,

10:33 these companies are compelled to buy them even if

10:35 they don't actually have space to put them yet.

10:37 Because if they don't, then they will have to go to the back of the line again.

10:41 The same goes for other components like cooling or energy supplies.

10:44 Even if they don't have anywhere to use them right now,

10:47 they are motivated to buy them just in case.

10:50 In industrial planning, this is called the bullwhip effect.

10:52 And it can help to explain the extreme levels of spending

10:55 at the same time as very few projects seem to be making much progress.

10:59 Now, this is a widespread issue across the world.

11:02 But to really highlight this problem, there are actually two new fully fit out

11:06 data centers just up the road from Nvidia's headquarters

11:09 in Santa Clara that are sitting completely offline because

11:11 they are waiting for local utilities to catch up.

11:14 At the moment, Nvidia is benefiting greatly from this.

11:17 They can almost guarantee that as soon as they get their chips from TSMC,

11:21 they can immediately sell them off again because

11:23 there will be a buyer ready for it,

11:25 even if it does end up sitting in a warehouse for a few months.

11:28 It's a dangerous game that only takes

11:30 a small demand correction to cause a massive overupp.

11:33 But for now, it's been incredibly lucrative.

11:36 Except that that assumption is starting to strain now.

11:39 Nvidia ended its financial year in January,

11:42 and when their annual reporting was released,

11:44 they had again had a record year in terms of sales

11:47 in the top line and profit in the bottom line.

11:49 But some other numbers started to catch people's eyes.

11:52 primarily that their inventory had more than doubled

11:54 from the year before and quadrupled from 2024.

11:57 If Nvidia really is struggling to keep up with demand,

12:01 there really shouldn't be any reason why they

12:03 are sitting on so much of their own product.

12:05 This is either suggesting that it's getting harder to move these chips or more

12:09 realistically that Nvidia themselves are struggling

12:11 with the same supply chain problems upstream

12:13 of them and they are just trying to get as much supply as possible

12:17 with the confidence that they will eventually be able to sell it to someone.

12:21 Ironically, when this anomaly was initially highlighted,

12:23 it was actually an LLM market

12:25 algorithm that found the disconnect first alongside

12:27 the trend of their customers taking a longer

12:30 time to actually pay for their receivables.

12:32 So, yeah, the power issue was already a major problem.

12:35 But today, there is also the energy issue on top of that.

12:39 When these facilities are fully operational, they draw a lot of power.

12:43 And power over time is energy.

12:45 And energy has already become very expensive over the last

12:48 few years thanks largely to these very same data centers.

12:52 But the war in Iran has well I would say poured fuel on the fire.

12:56 But it's kind of done the opposite of that.

12:59 Higher energy prices significantly cut into the viability of running

13:03 these centers and the cash burn rates associated with them.

13:06 So far, most data centers get their energy directly from the local energy grid,

13:09 which means now they are going to be paying more

13:12 money and waiting even longer for capacity to come online.

13:15 Other newer centers have gotten around

13:17 this by using their own natural gas turbine generators.

13:20 But well, natural gas has just doubled in price,

13:23 doubling the largest line item in their ongoing expenses.

13:26 It's not great, but it all gets much worse when you consider

13:30 these issues in the context of the third major problem facing Nvidia right now,

13:34 and that is how long these chips are expected to last.

13:37 Now, not to yank my own crank here,

13:39 but about 8 months ago in our video on Amazon's AI spending,

13:42 we highlighted the problem of capital

13:44 expenditure going towards cuttingedge chips,

13:46 which by their nature depreciate extremely quickly as newer,

13:49 better models come out.

13:51 Well, 4 months later, Michael Bur of Big Short Fame,

13:54 as well as a handful of other investors,

13:56 put out pieces effectively saying the same thing.

13:59 The industry standard amongst the big tech

14:01 companies is to depreciate these GPUs over 6 years when in reality they would be

14:05 lucky to stay operationally viable for 3 years.

14:08 This just basically means that they count 16

14:10 of the purchase price against income to offset taxes,

14:13 but also more accurately report the true annual profitability of the business.

14:17 By stretching out depreciation for longer than reasonable,

14:19 service life of these cards makes expenses look better than they really

14:23 should be for the companies that make up a majority of Nvidia's demand.

14:26 In a somewhat confusing rebuttal,

14:28 Nvidia themselves actually responded to this criticism

14:30 by defending their own accounting practices,

14:33 which wasn't the accusation in the first place.

14:36 Nvidia doesn't actually have that many hard assets at all.

14:39 It is just a chip designer with most

14:41 of the real hardware being manufactured by their suppliers, primarily TSMC.

14:45 So nobody really cared about their depreciation schedules.

14:48 The point was if companies like Microsoft, Oracle,

14:51 and Meta had more honest accounting, it would make their profits look worse,

14:55 which may reduce investor hype around AI investment and ultimately the demand

14:59 these companies have for even more GPUs coming down the pipe.

15:03 The reason this critique was leveled at Nvidia specifically is

15:06 because if Microsoft is pressured into re-evaluating this accounting standard,

15:09 they might lose a little bit of profitability on paper,

15:12 but they still have a major business outside of this AI stuff.

15:16 But if they order fewer GPUs, well, that is Nvidia's entire business.

15:21 Now, both the ongoing supply bottlenecks and higher

15:24 energy prices make this problem much worse.

15:26 If Nvidia is consistently releasing a new model of flagship AI

15:30 GPUs every year that make everything that's come before it obsolete,

15:33 then it could become much harder to justify purchasing pallets

15:36 of current-day chips in advance in the hope that they

15:39 will be ready when a data center eventually comes

15:41 online because by that time they might already be irrelevant.

15:44 On the other end of the spectrum, the depreciation of chips already in service

15:48 becomes worse the higher energy prices get.

15:51 If energy prices are very low, it can remain worth it to run older,

15:55 less efficient hardware,

15:56 even if it uses much more energy to achieve the same thing.

15:59 But as energy prices have increased,

16:00 those margins can be squeezed to the point where it

16:03 costs more to run a server in energy inputs alone,

16:05 than it could be rented out for.

16:07 At that point, multi-million dollar racks

16:09 of last year's cuttingedge hardware effectively become e-waste.

16:12 Now, the market can stay irrational longer than any naysayers can stay solvent.

16:17 And it is worth realizing that even if the business case makes no sense,

16:20 it can stay in motion for as long as investors are happy to put money into it.

16:24 But well, even that might be starting to change.

16:27 For the last four years,

16:29 private credit companies like Blue Owl and Black Rockck's Credit Arm have

16:32 been major financing partners on some of the biggest data center projects,

16:35 but they are now facing their own industry-wide problems,

16:38 which is going to make it much harder

16:40 for them to maintain the supply of easy financing.

16:42 Go and watch this video next to find out what private

16:44 credit actually is and why it has all gone so badly wrong.

16:47 And don't forget to like and subscribe to keep on learning how money works.