Let's Talk About the AI Bubble

What is discussed, and when

1. 0:00 – 5:00: Amazon, Anthropic, cape ratio, Michael Bur
2. 5:00 – 10:00: Amazon, Project Stargate
3. 10:00 – 15:00: cape ratio, Amazon
4. 15:00 – 20:00: Anthropic

Notable quotes from "Let's Talk About the AI Bubble"

• Throwing money into a new promising technological revolution pushing valuations namely the cape ratio to levels not seen since the docom bubble despite the space not at the…

  about cape ratio · at 1:11

• The compute of these data centers to AI companies themselves. These businesses include the likes of Amazon, Microsoft, Oracle, and a number of other companies with at times…

  about Amazon · at 3:31

• Committed to around $1.5 trillion in AI deals. This includes its involvement in Project Stargate, a $500 billion initiative to build out data centers in the US, a $300 billion…

  about Project Stargate · at 5:16

Who and what is mentioned in this episode?

People

• Michael Bur — Person · 1 mention · first at 0:48
  And Bank of England have warned about the growing risk around soaring valuations. Michael Bur, the investor who famously shorted the housing market before the 2008 financial…

Companies & orgs

• Amazon — Company · 3 mentions · first at 3:36
• Anthropic — Company · 2 mentions · first at 4:16
  Or companies with valuations of over $1 billion with notable examples including Anthropic and Elon Musk's XAI. And then of course we have the leader of the space Open AI…

Concepts & topics

• cape ratio — Concept · 3 mentions · first at 1:15
• Project Stargate — Concept · 1 mention · first at 5:22

Full transcript with timestamps

0:00 Hey everyone, it's Richard. You're watching the plane bagel. It's been about three years since chatbt first launched. And since that time, we've seen generative AI develop at a remarkable pace. AI images and videos have moved from horrific to nearperfect. And large language models have gotten so good that companies have started laying off employees to replace certain workflows with chat bots. But while it certainly feels like we've entered this golden age of artificial intelligence, with some companies even teasing at the prospect of artificial general intelligence or AI with humanlike cognitive abilities, there's been this growing sense that maybe, just maybe, we're actually in a bubble. One Bank of America survey found that 54% of global fund managers believed that we were in a bubble. The International Monetary Fund and Bank of England have warned about the growing risk around soaring

0:48 valuations. Michael Bur, the investor who famously shorted the housing market before the 2008 financial crisis, just this week announced a short position against key AI companies. And even Sam Alman, the CEO of Open AAI himself has admitted that yeah, we're we're probably in a bubble. And many believe that we're in for a repeat of the dotcom crisis of 2000. given that we have investors throwing money into a new promising technological revolution pushing valuations namely the cape ratio to levels not seen since the docom bubble despite the space not at the moment anyway really making any money openai who's behind chat PT and the videogenerating Sora platform was recently valued at $500 billion despite seeing just over $10 billion in annual revenue and losing even more than that to its expenses and then there's this monstrosity which you might have seen circulated in videos like that from Hank

1:43 Green that demonstrates the incestuous financial relationships we have in the AI space with companies like Nvidia effectively giving money to companies so that they can buy their chips. Now, that's a gross oversimplification, but you get the point. The optics aren't great. We're not really beating the bubble allegations. So, what's going on? Are we truly in for a do bubble 2.0? Well, to find this answer, we have to go to the source.

2:19 It's uh almost as inconclusive as me. Joking aside, at the risk of throwing yet another video into the sea of content on the supposed AI bubble, I did want to address the situation, explain what's going on, why things appear to be frothy, but also to highlight the key differences between the current AI situation and the dotcom bubble of 2000 without falling to the sensationalization that often befalls these type of topics. And to be clear, this is not to discredit the bubble claims or people like hankering or highlighting real issues in the space.

2:49 is just to give more details and context so we walk away with a better understanding of what exactly is going on. And to start off, it's probably worth highlighting the key players in the AI space. The first group are the AI chip companies, the businesses who sell the hardware needed to run these AI models. This includes AMD and of course Nvidia who's absolutely dominated and been incredibly profitable with this AI buildout which has helped skyrocket their market capitalization to a record $5 trillion making it the largest publicly traded stock. Then there's the infrastructure providers, the data centers, cloud providers, whatever you want to call them, uh the businesses who buy or rent these chips, uh build data centers to plug them into, supply electricity, and then ultimately sell the compute of these data centers to AI companies themselves. These businesses include the likes of Amazon, Microsoft,

3:38 Oracle, and a number of other companies with uh at times mixed financials, but a number of key players here do have very strong balance sheets. Then there's the AI companies themselves who buy the compute from these data centers to build, train, and ultimately run their AI models. And here we have a whole mix of companies from a financial standpoint. On the one end of the spectrum, you have firms like Meta, uh, which is a massive company with very good financials that's trying to build its own AI models. There are then hordes of AI startups with there being over 1,300 companies with valuations over $100 million and nearly 500 AI unicorns or companies with valuations of over $1 billion with notable examples including Anthropic and Elon Musk's XAI. And then of course we have the leader of the space Open AI themselves. by far the most successful company in artificial intelligence and yet one of the biggest

4:28 money furnaces with a company targeting revenues of 13 billion and losses of $8.5 billion for 2025. And the company's own estimates, according to one source, see it burning through 115 billion through 2029 with the company even losing money on their highest $200 a month subscription given that people are using the service more than expected. Now, because OpenAI is a private company, meaning they aren't a publicly traded stock, we have very limited insight into their financial health, but definitely a questionable business model. Most users are melting Open AAI's GPUs as free subscribers, after all.

5:05 Yet, this hasn't stopped the space from building out aggressively. Open AAI alone has made plans to build out up to an additional 26 gawatt in data center capacity and to date the company has committed to around $1.5 trillion in AI deals. This includes its involvement in Project Stargate, a $500 billion initiative to build out data centers in the US, a $300 billion deal with Oracle to purchase compute power over roughly 5 years, and deals to buy chips from Nvidia and AMD, which the Financial Times forecasts could cost 500 and $300 billion, respectively, with these just being some of the many arrangements Open AI has entered. Now, numbers like these can easily lose their meaning once they get high enough, but it's important to put them into context here. A gigawatt of power, for example, roughly equates the output of a single nuclear power plant and would provide enough electricity to power just shy of

5:58 900,000 households, according to the carbon collective. OpenAI is planning to consume the power of 26 of these simply to run its software. And that's just Open AI's own investments. McKenzie estimates that data centers themselves and the infrastructure for these companies will need to spend nearly $7 trillion in capital expenditures to meet the demand for AI and traditional IT applications. And while this is a global figure over a 5-year period, to put it into perspective, America's total capital expenditures for 2024 was around $ 8.3 trillion across all industries.

6:34 So, in effect, we're seeing the AI industry spend a fifth of America's total capital expenditures just in one industry. Needless to say, it's a lot, especially for companies that are losing billions of dollars annually. And so, because these companies are playing to spend well beyond their current means, they've naturally needed to call on investors to help fill the gap. And so far, they've risen to the challenge. Venture capitalists have poured a record of nearly $200 billion into AI startups so far this year alone with over half of total VC investments year to date going into the space. But the cost of buildout is only expected to increase. And that's where this uh concerning kind of circular chart comes into play. You see the diagram lays out a number of key AI financing deals and it demonstrates that many are a bit circular in nature.

7:25 Nvidia, for example, has pledged to invest $100 billion in OpenAI in return for the company buying its own chips. NVIDIA also invested in Core Weeave, an AI data center that sells compute to OpenAI, and agreed to backs stop the company's services if it failed to find customers. AMD supplied OpenAI with warrants for it shares in return for OpenAI buying its AMD chips with that announcement alone sending AMD's stock higher, which actually helps to pay OpenAI for the whole arrangement. and OpenAI announced a deal to pay $300 billion to Oracle for data center capacity who naturally is going to turn around and buy chips from Nvidia for those data centers who is invested in OpenAI. So you can see it kind of leads to this circle which uh has drawn some concerns.

8:13 And even since this chart has been published, we've seen even further deals announced, including OpenAI's $250 billion deal with Microsoft's Azour, with Microsoft being a major shareholder in OpenAI and the company's $ 38 billion deal with Amazon's AWS. And these are just some examples. OpenAI isn't even the only company involved in these sort of circular deals. But at the center of a lot of these deals has been Nvidia, the one who so far has profited the most from the AI gold rush. It seems like they're investing that dragon's horde in any AI company that pops its head up.

8:45 With there being recent announcements around investments in Intel, Nscale, Wave, and even Nokia, which I'm guessing they're looking to make their chips indestructible, which actually could be a smart move, but it's led to concerns that the so-called pick and shovel seller is actually just paying gold miners to buy its picks and shovels and effectively propping up the space. Now, to be clear, the arrows in this chart aren't dollar for-dollar arrangements, so the circle analogy isn't a perfect one, but it could mean that Nvidia is overstating its profits. After all, if you're investing in your customers to give them the cash they need to buy your chips, then you're effectively selling your products at a discount without recognizing the hit to your profit margin since investments aren't factored into the margin calculation. And some have pointed out that this type of vendor financing became very popular

9:31 with companies like Nortell towards the end of the dotcom bubble. Not to mention that with us seeing AI companies starting to borrow billions of dollars to buy these chips and even putting their current chips up for collateral for some of these loans, it's led to again this concern of this propping up of demand beyond what's sustainable. And amid this massive buildout and the deals being made in the space, there's a question about whether there will actually be enough demand from end users to justify all these investments. Bane & Company, for example, estimates that AI companies will need $2 trillion in annual revenue by 2030 to be profitable, which as noted by the Wall Street Journal is more than the 2024 revenue of Microsoft Meta Alphabet Amazon Apple, and Nvidia combined and five times more than the entire software as a service market. And there just doesn't seem to be a path towards that type of

10:23 profitability. Yes, 88% of companies are using AI in their workflows to some capacity and OpenAI itself has seen remarkable user growth. You might remember that Chat GBT broke records in terms of being the fastest growing app and the company now boasts 800 million weekly active users. But only up to 6% of these users are estimated to be paying for this service. And the majority of companies using AI, around 61% according to McKenzie, currently don't see an enterprise impact on earnings before interest in taxes. And Croup earlier this year estimated that AI revenue would only reach $43 billion in 2025 and $780 billion by 2030, which still represents pretty remarkable growth at 80% a year, but less than half of what Bane estimated would be needed for the industry to break even. And that's all assuming the AI companies are able to deliver on what they're promising. There's also the question of

11:19 the bottlenecks we're seeing form in the space with the key one being electricity. Building out grid infrastructure takes years of planning and development and that's assuming you get through all the regulatory hurdles of bringing online more fossil fuel or even renewable plants which don't always provide the consistent electricity needed for data centers. And while nuclear power plants have been a popular choice for data center powering in the US, the NRC application review process can take 5 years alone to complete with construction often taking another five or more years according to the US EIA's own website, which might be longer than these companies are able to woo investors. It's even led to some companies like XAI skirting regulations to set up on-site generators to try and meet their electricity demand. Then there's the consideration around the lifetime of the data center hardware

12:07 itself. Over the last few years, large tech firms have been expanding the assumed lifetime of its servers, spreading out their accounting costs over a 4 to sixyear period. Whereas some have argued that with Nvidia releasing updated, faster chips every year, the amount of time before these chips have to be replaced is actually much shorter. Which means that many of these companies are going to run into higher replacement costs than they're disclosing to investors. And when you throw in the sheer scale and concentration of the AI industry with 35 to 36 companies accounting for 99% of all AI token spending, two companies buy nearly 40% of Nvidia's chips and 10 companies, most of which are involved with AI, accounting for 40% of the S&P 500. The failure of any one of these companies could lead to fairly big shock waves, not just in the space, but even the markets as a whole. Not to mention the

12:57 real estate, utility companies, and financing firms all involved in the industry. And OpenAI alone has a lot of the market's weight seemingly on its shoulders at the moment. OpenAI's deal announcements have added hundreds of billions of dollars to the market value of companies like Broadcom, Nvidia, and AMD. If we see a pullback in that company, it's going to hurt valuations for many of these businesses, especially when you consider how expensive many of these companies have become with investors pricing these stocks for perfection, leaving very little room for disappointing results. So yeah, it's easy to see where the comparison with the dotcom bubble comes from. In fact, if you lump in software, computers, and peripheral equipment, and other information processing equipment, the current private domestic investment in IT as a percentage of GDP is back where it was during the dotcom bubble. And

13:46 what's important is that you can still see losses in the market, even if the technology proves to be promising. The internet did of course revolutionize our lives, but also led to an 80% collapse of the NASDAQ over two and a half years as thousands of shaky.com startups convinced investors that they were on the frontier of a new highly profitable business that ultimately would only actually benefit a few players. So, brings us back to the question, are we truly in for a do 2.0? Well, while there are of course parallels uh with, you know, a new technology, investor enthusiasm, and concerns that things are overvalued, there are also a number of key differences that make for a very different setup here. For one, there's valuations. Yes, the cape ratio, a commonly cited measure of stock market valuations, is fast approaching the highs we saw during the dotcom bubble, with a cycllically adjusted PE ratio

14:36 measuring the difference between a stock's price and how much money the company is actually earning. So how much investors are willing to pay per dollar of earnings per share effectively. However, it's important to note that with the cape ratio specifically, it does include smoothing, meaning that the earnings put less weight on any sort of surge in profits we see. If you use a more basic trailing PE ratio for the S&P 500, which only considers earnings over the last 12 months, you can see that the current ratio sits at roughly 30 times, which yes, it's still high on a historical basis, but it compares to the dotcom bubble where we reached a high of 46 times. So, we aren't yet at the same extended level of valuation, which brings us to the second point, which is that compared to the dot bubble, there's actually strong financials and earnings to back up many of these valuations.

15:23 Morgan Stanley, for example, has highlighted that S&P 500 companies are generating three times the cash flow as a share of their valuations versus what they were generating leading up to 2000. And on a relative basis, there are nearly half as many unprofitable technology companies when compared to 2000 when 36% of tech companies were losing money. Now, this doesn't account for private companies like OpenAI or Anthropic, which are highly unprofitable. But these startups have the backing of companies that are highly profitable and are generally viewed as financial strongholds. Many of these big tech companies have very little debt and hordes of cash on hand that they've accumulated over the decade. So while there is some evidence that leverage is slowly creeping up in the space, which could lead to some problems if that continues. Currently, a lot of the spending has been backed by cash on hand

16:09 or cash flows from these large companies, which does limit the fallout if we see one of these investments not ultimately pan out. I also want to highlight a few considerations when it comes to the circular dealings in the AI space. And this is not to really defend these companies or argue that there's not something, you know, kind of prop upy going on here. Uh but when it comes to Nvidia specifically, it kind of makes sense that this is the approach they're taking. Companies like Nvidia are making so much money that it can be difficult to find places to park that capital without dragging down the company's performance. So, it kind of makes sense that Nvidia would invest in the ecosystem that will ultimately benefit its future demand. If the companies it invests in are able to succeed with that investment, then that means that they are effectively incubating a future customer. It's also not where most of

16:56 their cash flows are coming from. The company over the last 12 months made $77 billion in operating cash flow and paid out $24 billion in investing cash flow, of which a fraction would be this sort of vendor financing. And while there are larger financing deals set up with OpenAI for the future, for example, it's important to note that many of these arrangements are performance-based, which makes sense given that there is this perception that OpenAI may be faking it till it makes it that it's kind of talking out of its ass with with these large deals that it's making, given again just the scale of these deals relative to its current revenues.

17:29 But the point is, Nvidia isn't just writing blank checks with, according to Bank of America, vendor financing only representing a fraction of current AI spending. It's also worth noting that bubbles don't tend to just pop on their own and part of the reason we saw the dot bubble deflate was actually fairly rampant fraud in the space. Many large companies were discovered to have been cooking their books to meet higher and higher investor expectations. And while yes, fraud is certainly possible today. Uh but we have seen reporting standards improve since the 2000.com bubble because of these crises. We're also in a very different economic environment for better or for worse mind you with rising interest rates often being cited as a factor that killed the dotcom bubble whereas currently we have the Federal Reserve seemingly putting rates on a downwards trajectory. And with all this, it's important to acknowledge that we

18:15 could very well see Open AI's profits scale from here. Again, as a private company, we don't get much of an insight into the firm's financials. But with the company launching new services and many businesses looking to lay off employees to integrate more AI into their workflows, there is a chance the company is able to build its revenue. In fact, just as I was finishing up this video, it came out that Sam Alman expects the company to reach $20 billion in run rate revenue for 2025 and hundreds of billions by 2030. Albeit, they're still expected to be unprofitable in the near term. Now, I don't know if that's how things are going to play out, but companies who invest in firms like OpenAI do get more insight than we have from the outside. And yes, companies make bad investments sometimes. We aren't today living in the metaverse that Mark Zuckerberg desperately hoped we would be. But assuming that all these

19:04 players are being perfectly irrational is just as dangerous as assuming that they're all being perfectly rational. Now, with all that being said, I I do think there's a bubble of some sort. To be clear, this is not to downplay the the very lofty valuations we have uh the risk of a correction here. It's just to try and distinguish between a crisis and markets being disappointed with how certain investments play out. It is true that with high valuations, there's little room for error and that leads to lower expected returns for investors.

19:34 It's also true that the current investment activity is just blatantly unsustainable when you consider all the AI activity going on. We're also seeing many big tech companies lay off employees, which some are pointing to as evidence that capital spending and slowing economic performance are impacting operations, and we've seen signs of strain on market liquidity, which just means that investors might be less willing to lend that money to these AI startups in the future. It is also likely that we'll see a consolidation of the space, similar to what we saw with the do- bubble. The truth is, we just don't have enough demand for 500 AI unicorns. And in fact, there's a lot of good research that highlights that technological revolutions have historically been bad investments on average, given that they tend to lead to this boom and bust in market prices. And the attrition experienced by all the

20:19 companies throwing their names in the hat has a negative impact on market returns. Even if the technology goes on to be successfully adopted as companies like Netscape perfectly highlighted during the dotcom bubble, being the market leader during the technologies rollout doesn't guarantee that leadership spot in the future. And if you've never heard of Netscape before, kind of perfectly illustrates the point. But the point is, we just don't know yet. AI is a new technology that could prove very promising and could lead to a lot of profits for these companies. Or demand just might never meet the high expectations that companies like Open AI ultimately have. We could be in for a metaverse 2.0. I would think this is probably a little more promising than that, but you get my point. Not to mention that at any point we could have a geopolitical event, a headline, a new development that completely changes what

21:04 we were expecting to happen. And even if we are ultimately in a bubble, timing such a thing is near impossible. Alan Greenspan, for example, the federal chair at the time, called out the irrational exuberance of companies back in December of 1996, over 3 years before the bubble would ultimately collapse. And with companies like Meta viewing the risk of underinvesting as higher than the risk of overinvesting, these rallies can ultimately last as long as investors are willing to pile money into the space. Yes, at some point these companies will have to transition from surviving on investor funds to uh earning customer revenues to fund their operations. But how long of a timeline we face until that has to happen remains to be seen. So hopefully you found this video helpful. As always, it's important to take a balanced view. There's just a flurry of variables that make it impossible to predict how things are

21:53 going to play out. So, while you certainly don't want to get caught up in the AI hype and, you know, throwing all your money in Nvidia because you have Chad GBT making Studio Gibbly versions of yourself, it's worth hearing all the arguments. Thanks for watching. If you like this video, please do make sure to like, subscribe, all that good stuff. It does help the channel tremendously. Let me know your thoughts on the AI bubble. Uh, do you think things are overhyped at this point or do you think concerns are overblown? Any and all thoughts welcomed in the comments down below. Uh, thanks again for joining me and as always, be safe out

Questions about this episode

What is discussed in "Let's Talk About the AI Bubble"?

In order: 0:00 Amazon, Anthropic, cape ratio, Michael Bur; 5:00 Amazon, Project Stargate; 10:00 cape ratio, Amazon; 15:00 Anthropic.

What is mentioned in this The Plain Bagel episode?

5 people, works and ideas are mentioned, including cape ratio, Amazon, Anthropic, Michael Bur, Project Stargate. Each one on this page links to its own page with every episode that mentions it.

How long is "Let's Talk About the AI Bubble"?

The episode runs 22 min; the transcript is about 4,181 words in 30 timestamped sections.

More from The Plain Bagel

• All analyzed episodes
• Books mentioned on The Plain Bagel
• People mentioned on The Plain Bagel
• Topics mentioned on The Plain Bagel