Some believe that artificial intelligence will become the most important technology of the 21st century. Others argue that it is an obvious economic bubble. I believe both sides are right. Like 19th century railways and 20th century broadband internet buildouts, AI first rose, placed second, and ultimately changed the world.
Numbers have no meaning. Tech companies are projected to spend around $400 billion on infrastructure this year to train and operate AI models. In nominal amounts, it's more than any group of companies have ever spent doing anything. The Apollo program allocated approximately $300 billion inflation-adjusted dollars to bring America to the moon between the early 1960s and early 1970s. With AI build-outs, businesses must collectively fund new Apollo programs, not every decade. Every 10 months.
It's not clear that companies are ready to regain their investment, but by their own testimony, they're just trying to keep spending. Total US AI capital expenses expenditures are projected to exceed $500 billion in 2026 and 2027. This is Singapore's annual GDP. but, Wall Street Journal American consumers report that they spend just $12 billion a year on AI services. That's roughly Somali's GDP. Understanding the economic differences between Singapore and Somalia gives us a sense of economic cleavage between island vision and reality. Some reports show AI usage is actually declining in large companies that are still trying to figure out how big a language model can save.
Every financial bubble has a moment when you look back and think: How did the sensory person miss the sign? There are many signs of today. Thinking Machines is an AI startup created by former open AI executive Mira Murati, and has sourced the largest seed round in history. It raised $2 billion in funding at a $10 billion valuation. The company has not released any products and refused to tell investors what they were trying to build. “It was the most ridiculous pitch meeting,” one investor who met Murati said. I said. “She said, “We run an AI company with the best AI people, but we can't answer the questions,” while a recent analysis of stock market trends shows that There are no typical rules for wise investments You can explain what's going on with the stock price right now. Stock prices have historically followed the basis of revenue, but today's markets are driven overwhelmingly by momentum as retail investors believe others are loading into meme stocks and AI companies.
All economic bubbles also show signs of financial overengineering, such as secured debt and subprime mortgage support securities that exploded during the housing bubble in the mid-2000s. Ominously, AI appears to be in its own stage of financial magic. As economist AI Hyperschool, the biggest spender of AI, is Use accounting tricks It has the effect of inflating their profits to curb reported infrastructure spending. As an investor and author Paul Kedrosky told me on my podcast Plain EnglishLarge AI companies are also spending huge amounts of AI on SPVs or special purpose vehicles, disguising the cost of AI build-outs.
My interview with Kedrosky received the most enthusiastic and free feedback of the show I've had a while. His level of insight per minute was off the charts, mentioning:
-
How AI capital expenditures collapse
-
Why AI buildouts differ from past infrastructure projects, such as railways and dotcom buildouts
-
How AI spending is creating a black hole of capital sucking up resources from other parts of the economy
-
Ordinary investors may be able to feel the bubble pop just before it happens
-
Why is the whole financial system balanced by a large chip maker like Nvidia?
-
What amazing industry will face calculations if bubble pop
Below is a sophisticated transcript of conversation organized by topic area and decorated with charts and graphs to visualize points. I hope you learn as much from his commentary as I do. From a purely economic perspective, I don't think there is a more important story in the world.
Derek Thompson: How big is the buildout for AI infrastructure?
Paul Kedroski: There are huge amounts of money unfolding, making it a very small area with very small recipients and some very small areas like Northern Virginia. Therefore, this is a very concentrated capital pool and is large enough to affect GDP. I did math, and in the first half of this year I discovered data center-related spending, a huge building full of GPUs. [graphical processing units] The racks and servers used by large AI companies to generate responses and training models probably account for half of GDP growth in the first half. This is definitely a banana. This spending is huge.
Thompson: Where is this money going?
Kedrosky: For major companies like Meta, Google and Amazon, less than half of the data center costs are GPU chips. It's about 60%. The rest is a combination of cooling and energy. And the relatively small components are the frames of the building, concrete pads, and the actual structure of the real estate.
Thompson: Do you think AI spending is already distorting the economy in 2025?
Kedrosky: Looking back, this is the analogy I drew. Large capital expenditures on one narrow slice of the economy of the 1990s diverted capital from US manufacturing. This is a small manufacturer of hungry capital, making it difficult for them to raise money cheaply. The margin had to be high because of the increased cost of capital. Meanwhile, China entered the World Trade Organization, with tariffs falling. Mostly due to rising capital costs, it has made domestic manufacturers extremely difficult to compete with China. It was all sucked into this “Death Star” from Telecom.
So, in a strange way, we can trace some of the manufacturing job losses in the 1990s in communications. Because it was a loud suction sound that sucked all the capitals from every other part of the economy.
The exact same thing is happening. If I'm a large private equity company, I'm not paid to spending money outside of data centers. It's the same phenomenon. If I'm a small manufacturer and want to benefit from onshoring manufacturing as a result of tariffs, I'm trying to collect it as a paper. The hurdle rate has become much higher. This means that you have to generate much higher returns as you compare it to other parts of the economy that accept huge amounts. And when we look at what's going on with AI, and the massive intake of Openai, the returns seem to be incredible. So, I'm carelessly hungry for a huge slice of the economy again, just like what we did in the 1990s.
Thompson: That's very interesting. The story we're used to talking about manufacturing is that China has taken our position. As economists like David Autole call it, the “Chinese shock” essentially came to China's manufacturing industry, with the production of Shin Shenzhen replaced Ohio's production. It adds that Telecom has absorbed the capital.
And now you're fast forward to the 2020s. Trump is trying to reverse China's shock with tariffs. However, we are recreating AI as a new telecom. The new Telecom is a new Death Star taking capital that could go to the margins.
Kedrosky: It's even more insidious than that. Let's say you're Derek's huge private equity company and manage $500 billion. You don't want to allocate that money to many manufacturers a check of $5 million at a time. What I'm seeing is a nightmare of having to track down all of these small businesses doing what they know.
What I want to do is write separate $30 dollar checks for $50 billion. I want to write a few huge checks. And this is the dynamics of private equity that people don't understand. Capital can be allocated in a variety of ways, but partners in these companies don't want to write a large number of checks on a large number of checks, even when the hurdle rate is competitive. I'm a human, I don't want to sit on a board of 40. So, even if everything else is equal, it is not equal, this other perverse dynamic. Therefore, due to the internal dynamics of capital, it places manufacturers in a worse position that could otherwise benefit from the supervisory phenomenon.
Thompson:What about the energy part of this? Electricity prices will rise. Data centers are extremely energy-thirsty. I think consumers are against building local data centres, but data centres have a huge political power of their own. How does this unfold?
Kedrosky: So I think we'll quickly look at offshoring in our data centers. That will be the response. It will more and more be happening in India, it is happening in the Middle East where large allocations are taking place in new data centers. It's happening all over the world. The focus is on moving offshore for exactly this reason. Bloomberg recently spoke to us about the northern Virginia suburbs. This is currently essentially surrounded by data centers. This was previously a rural area, everything around it, all the farms, sold out, and the people in this area wait a minute, who will sue? I have never signed up for this. This is the beginning of the Nimby phenomenon. Because it's visceral and emotional for people. It's not just about the price. There is also: If you have a 6 acre building next to you, it's not something you signed up for.
