How China Just Overtook America In AI Traffic

Key Takeaway

Don’t let AI hype or doomscrolling make your decisions for you. Today, pick one repetitive task—summarizing notes, organizing data, drafting a first pass, or researching options—and test an AI tool against a clear quality standard. Keep what saves time, verify what matters, and learn where it fails. The durable advantage is not blind faith in AI; it is becoming the person who can direct, evaluate, and scale it responsibly.

Episode Overview

The episode frames AI as a strategic intelligence race between the United States and China while examining whether massive investment in chips and data centers resembles a dangerous financial bubble. It argues that AI infrastructure is already generating revenue and usage, but warns that debt growth, cheaper Chinese open models, and uncertain hardware depreciation could pressure leading AI companies. The host ultimately advocates a nuanced view: track the economics and use the technology directly rather than adopting simplistic hype or panic narratives.

Full Transcript

Transcript of How China Just Overtook America In AI Traffic from Impact Theory. Auto-generated from episode audio; may contain minor errors.

The reality right now is that people have very, uh, well- founded fears about AI, and also very unfounded ones. This video will help us answer the question: is AI living out its final days? Is there such a big systemic risk here that everyone should be afraid? On July 28, 2026, an AI startup called Orchid posted a presentation video that went viral on Twitter. In it, a couple celebrates their anniversary, but the guy forgot about it . The woman, clearly irritated, complains to her AI assistant Orchid and asks for help.

The assistant says he has foreseen everything and sends a message to the guy, reminding him of the date. Then he books a table for him and orders her the flowers she loves. He writes in response: " Okay, okay, I'll handle it." And the chatbot replies: “You won't be able to handle it. That's why I'm here." Orchid tells the woman that he is pretending to have planned everything himself, as he praises the AI for saving their relationship. And they live happily ever after. How wonderful. Instead of becoming a better partner, you can just stay a piece of [ __ ] and let AI " pick up" your girlfriend.

What a wonderful world, isn't it? It's no wonder this video has sparked a wave of outrage, and it's just another example of how out of touch with reality the developers of this nonsense are, which will likely lead to their downfall. One of the most important things to understand when considering the cultural aspect of a tool like AI is that some people will use it horribly, while others will use it effectively. But just because something can be used in a way that's pointless, or a company advertises it in a way that you see second- or third-order consequences— if you get stuck at that level of argument, you'll hit a dead end.

When you understand that AI is an arms race between the US and China for intelligence itself, then you will be thinking in the right direction . And the level of discussion around AI is driving me crazy. You need to understand it as a weapons system. You need to understand why intelligence is the most important battle we have ever fought as a nation before we can have any real conversation. Now he's going to get into the heart of it all and into cyclical financing, and we're going to talk about that, but also, on a cultural level, I want to make sure people are thinking about it in the right way.

There is only one reason why humans have been able to have such a huge impact on the planet, and no other. Our brains are organized for higher- level intelligence. I know people want to say that dolphins are smarter than us, or that octopuses are so smart that we could consider them to have an alien intelligence that rivals our own. Kids, if that were true, they would figure out how to breathe out of water, just like we figured out how to breathe in water. The reality is that we are the most dominant species the world has ever seen precisely because we have higher- level cognitive abilities: the ability to model the psyche, plan for the future, manipulate objects in the imagination, and then embody them.

Intelligence, flexible intelligence—that's the ultimate goal if we want to move forward as a species. So when you start thinking about what artificial intelligence actually is , everyone gets so hung up on the word "artificial" that they think of " Terminator" and forget about " intelligence." They forget that if we could scale intelligence, it would fundamentally change our lives, it would fundamentally change the world. We look at all these demographic problems and complain that it takes 18 years to prepare an adult for work, but that's no longer the case.

So when you argue on the level of Sam Altman being a scoundrel, or you don't like Elon Musk, you're missing the point of why this technology is so important. So please keep in mind: don't fall for this cultural frame. Thanks to Derek Thompson for this great article on Substack that helped create this video. At this point, most of us know that AI is in a “bubble” state. A financial bubble is when asset prices in a particular industry rise to unreasonably high, false levels that do not reflect the real value of what is being created.

I like his approach, I understand it very much, and today I will go a little beyond my own vision. I will, of course, give you some guidance on where I think we are in relation to this bubble. But now I want to make the case for an optimistic AI scenario. People like Raoul Pal, who I just spent time with yesterday, believe that we are nowhere near a bubble. Um , at least not in the way others think. And if you look at historical examples, there are reasons for this .

So, I'll walk us through this by letting him present a pessimistic view . I'll give him a chance to make the arguments that you're probably used to hearing from me, and I'll come at it from a different angle to give an alternative perspective, because I think the "bubble" continues to inflate precisely because people disagree about our current assessment. Have we completely lost touch with reality, and the cost estimates no longer reflect the real value of what is being built, or are we looking at something where we have n't even begun to grasp the first steps of the benefits it will bring to society?

So it is this clash of views that we will examine. To understand how big this “bubble” is getting, we need to know what a hyperscaler is and how they spend their funds. Hyperscalers are the largest cloud computing and data center providers that have built most of the AI ​​infrastructure. The four largest hyperscalers are Meta, Microsoft, Alphabet, or Google, and Amazon. Since 2021, these four companies have been pouring unfathomable amounts of money into AI development and related infrastructure. A good reason for this is that, first, AI is a race to create the next big product, and companies are trying to be first, but also because these companies can afford it.

Hyperscalers’ free cash flow— the money left over after paying for day-to-day operations and long-term expenses—has remained positive since the AI ​​arms race began. Just 2 years ago, the four largest hyperscalers—Microsoft, Alphabet, Meta, and Amazon— generated approximately $210 billion in free cash flow each year. But in 2026, their free cash flow fell below zero. Of course, not every public company is suffering as wealth flows from hyperscalers to semiconductor manufacturers. Before we get into that, it's important to understand that there is a big difference between cash flow and free cash flow.

So what you see are companies that are still optionally very valuable. That is, with a very positive cash flow. They are viable by default. When a company is " dead by default ," it means that it cannot become profitable through its core product or service, and is forced to attract external capital because it is not even breaking even. However, when you have a company like this that spends money on non-essentials, its core business remains the same. Their main business is still incredible " baby cows." So at any point they can say, “ Look, AI was a failed experiment.

We will stop spending money on this." And then they will again become as profitable as before. So it's important to understand what they're doing now, they're not investing in anything disruptive. They are making a big bet, and if it doesn't play out, it will certainly cause instability in the perception of them as a company. But if you look at what Meta did by changing the name, it was their bet that the metaverse would be the future. It was the peak of Web3 and everything that was happening there, and they had a vision of what it should be like.

They end up buying Oculus. This is becoming one of the company's key divisions . They spend billions of dollars every year focusing on this, changing the name, reorienting the entire company, and then deciding, "No, this isn't really going to be the future." They stop spending money and go back to being the " baby cow" they always were, although it was less noticeable due to investments in the future, but the core business continued to generate profits day after day. The same goes for these hyperscalers. So yes, they are now in a shocking position.

And you've heard me say many times, " It's amazing that Google, which has never had negative cash flow since its founding, is now losing money on that metric for the first time." This should definitely spark curiosity, and if you invest in them, you should ask, "Will this work?" Because they spend a huge amount of capital. But keep in mind, they can stop it, like Meta did with their metaverse initiative. They can stop these expenses and that money will stay put. So, it's important to make sure you don't perceive unnecessary capital expenditures as a fundamental business flaw.

These hyperscalers need to constantly spend, spend, and spend. So what do they do? They take on more debt. Here's what it looks like for the "Big Four." To raise more money for spending, public companies issue bonds —essentially a way for investors to lend money to the company for a certain period of time. The company pays the investor interest at fixed intervals , and the original loan amount is returned to the investor when the bond matures. From 2024 to 2025, the annual volume of bonds issued by hyperscalers increased from approximately $20 billion to $110 billion.

And in 2026, this figure reached $150 billion . Many are now comparing this to the dot-com bubble of the late 90s and early 2000s. In an article for Business Insider, Jennifer Sohr writes about J.P. analyst Morgan, who compares the divergence in AI trade to communications equipment in 1999. Quote: "Companies supplying communications equipment began to show parabolic growth in 1999, while companies making large investments in this area fell from their peaks." So, it's important to understand when we're talking about what the actual size of a bubble is.

Before talking to Raul Pal, I had the feeling that these companies were much further ahead than they actually were. So, if you look at debt as a percentage of cash flow, or more precisely, as a percentage of their value. They now have 4% in AI, which is a very manageable figure compared to 30% at the peak of the dot-com mania. So right now, in terms of scale, we're still very far from that level , which was actually quite shocking to me. I didn't realize we were so far from the peak of the dot-com madness.

So, yes, these guys are taking on an extraordinary amount of debt, especially when you look at it in " pure" dollars. But if you look at it as a percentage of their value, we're still in a much, much, much more prudent state than we were during the dot-com bubble. So, thinking about it on my own, I lean towards a position somewhere in the middle between Raul and the other experts. Raul, like everyone who believes in the rise of AI , has a feeling that everything will develop something like this .

Um, this is the most important technology the world has ever seen. The US is engaged in an arms race with China. The US simply cannot help but push this industry forward. So, even if one of these individual companies were to do something stupid and get into financial trouble, the US government would step in and either support that company or help sell off its assets to other players. So any funds raised to build the physical infrastructure of data centers will be used. They may not be used by Anthropic or OpenAI, but they will definitely be used.

And so, when your mental model says, "Oh yeah, I've seen this scenario before. I know the US is going to print an absolutely obscene amount of money to prop up either the banks or the housing market." But they will print an obscene amount of money to protect an industry, especially if they consider it systemically important, which is exactly what it is now. America is not just making the only bet on AI in the stock market, it can be argued that the entire world is making the same bet.

So the chances that a country like China or the US, which has the ability to print money to solve problems, will not do so are practically zero. So , if problems arise in the industry , all of us, the taxpayers, will insure it . Therefore, we can expect this to continue. Add to that the fact that our debt-to- value ratio is 4% versus 30%, and suddenly we may not have anywhere near the level of risk we had when the dot-com bubble finally burst. And when you ask the question not about whether it's a bubble, but about where we are in this bubble ?

Are we in the early stages? Are we in the late stages? I think we're in the late stages in terms of stock prices, but very early in terms of the real debt load on these companies, which are real " cash cows." So when you look at the amount of debt that these companies have taken on, it's not necessarily scary. But to me, when you look at the Capgemini 40 ratio in the stock market, that ratio is meant to show, "Hey, how much have these companies made in the last 10 years relative to the value of the stock market?" And when the market value becomes too high compared to the actual profits of these companies, it is like an alarm going off.

For me, this is a warning sign. The traditional indicator is 16-17. So you would expect the price people are willing to pay to be 16-17 times the 10- year average earnings of the S&P 500 companies . It's at 40 now. And the last time we were at 40 was right before the dot-com bubble. So this is the case where you say, "Ah, I see. We're starting to get into crazy territory ." But the companies are probably doing well. The dot-com bubble finally burst in early 2000, less than a year after this gap was noticed.

Another indicator of what lies ahead is that Microsoft ended June with its biggest monthly stock loss since the dot-com bubble in 2000. Back to the show, but for now let's talk about one rule in travel. The beef sticks should be in the bag. Simple and clear. Always. Airports, hotels, a long day without any chance to eat—I don't leave it to chance. Paleo Valley Beef Sticks end up in my bag before I go anywhere. I literally don't leave the house without them. When the day goes awry and I urgently need to eat, the right choice is always with me.

I eat them because they are real. Grass- fed beef, organic spices, natural fermentation. There's no soy, gluten, or any junk here. 6 grams of real protein in one stick , and they really taste amazing. No chalky aftertaste or chemical undertone, just food I can happily take. Most meat sticks are garbage wrapped in a healthy product. But not these. That's why they've settled in my bag forever. Get a 15% discount on your entire order. Click the link in the show description to save now. We'll be back to the show soon, but first let's talk about emails that are hunting you down.

Scams don't look like scams anymore. They are just crazy. They look like parcel notifications or bank alerts. The message usually comes from a name you know , and then just one click is enough. And the people who do this are not careless or stupid. They are simply busy or distracted. This happened to me. I'm sorry that this is true. We see something urgent and we react. This is where Surfshark comes to the rescue. Their phishing scanner is powered by AI designed specifically to detect such tricks.

It analyzes the sender , the link, the tactics, and tells you if something is wrong, even before you click. This is the pause you don't always give yourself . Go to surfshark.com/tomb or use the code tomb to get four extra months. Go to surfshark.com/tomb right now and use the code tomb to get four extra months. Now let's get back to the show. These companies are essentially counting on what other reporters rightly call a miracle. The five largest hyperscalers are trying to promise that their revenue will double over the next three years while reducing operating costs by about $80.

As depreciation expenses will increase rapidly over the next few years due to the construction of new facilities and the purchase of equipment, selling, general and administrative expenses should decrease at an unprecedented rate. As Purdue University accounting professor Kevin Koharki notes, analysts seem to assume that any increase in depreciation will be directly offset by selling, general, and administrative expenses, in order to keep operating margins stable or slightly higher. I can't think of any scenario where this has happened before. Okay, one thing I think they're missing is what people are actually looking at.

And that's what Jensen Huang, CEO of Nvidia, is trying to get across to people. There is currently a fierce debate about what the real value of these data centers is. Depending on how you look at it , we're either starting to get on shaky ground or we're doing fine. Jensen Huang's argument is: " Listen, guys. You keep looking at the depreciation schedule, acting like Michael Bury, panicking and thinking we have the wrong depreciation schedule. But the reality is that we have H100s that came out 6 years ago and are still in use." So, it's clear that they don't depreciate in 3 years like everyone says.

Yes, faster chips will come on the market, and yes, those chips will be better, but you're acting as if the old chips are simply going to cease to exist, which is not true. And that's why he's offering a financial package that would allow people to come in and, let's say, view a long-term data center lease as something that generates profit in the long term, rather than as a technical product that quickly depreciates. And he is so confident in this that he is ready to cover 25% of the risks on each individual deal.

But he is ready to act as a guarantor, in essence, to insure against a drop in the value of these chips. It does not guarantee loans, but the product itself. And he says that this product... I'm so confident that it will maintain its value or even potentially increase in value that I'm willing to cover up to 25% of the losses. It's based on the fact that the cost...I don't remember if it's H100 or H200, but the cost of that computing hour hasn't gone down, it's gone up.

So now, instead of being a depreciating asset—which it is because it's technology, so let's not confuse it with short-term trends—now they can actually make more money over time. So chips become more valuable over time, not cheaper. I think he has a very convincing argument that the data centers that are currently being built are not operating in the catastrophic scenario that people expected. Because this is an unfortunate reality: when you look at historical examples of creating very expensive infrastructure, usually the first wave of investors goes bankrupt because it takes too long for the technology to start making money.

Let's take the railways. You have to build a railroad before you can run trains on it, before it becomes valuable and people can use it. But the big difference with AI is that yes, you're still building infrastructure, but it's not like with railroads, where at first it's zero value, zero value, zero value, and then finally it shows up. And throughout the period of zero value, you were forced to invest huge amounts of money in building stations , guarding lines, and laying tracks. Yes, this will all last an extremely long time, but there is a huge delay in generating income.

And Huang and other people who are optimistic about AI say, "Guys, AI does n't work like that." You are earning income right now. And not only are you making a profit right now, and yes, maybe the profit is n't coming in as quickly as we expected, partly because of China's actions with open source models, but Anthropic's revenue is growing faster than any other company's revenue in history. Maybe not in percentage terms, but if you look at real money, it's just crazy. Crazy amount of billions of dollars they have added to their balance sheet in just the last 5 months.

So, let's realize that this is not the railways, and this technology is already being used now. Yes, this technology will depreciate over time, but for now we see it living for at least 6 years, and the cost of computing is increasing. Okay, this is all very important. Now I will express my, perhaps more balanced opinion: yes, this is true, since we are still in the process of building, and right now the demand exceeds our real ability to satisfy it. However, you don't see many other big companies emerging, and that raises questions.

This, of course, does not put an end to it , but it makes you think. Where will other companies come from to meet this growing demand? Because it will probably take some time for AI to penetrate deeper into the economy. I always say: make your refrigerator smart, make your blender smart. All of this just takes time. And if we reach that upper limit, where there is no longer the income growth that we need, overproduction of computing power will begin. Then very quickly the old chips will start to lose value, or if China or someone else creates more efficient algorithms and you need less computation for the same result—there are several scenarios where the value of these chips could drop , not just because of reduced demand.

So, demand may continue to grow, but the need for computing power may decline to meet that demand, causing their cost to fall. That's why Jensen obviously doesn't guarantee 100% of the value of these chips. He only takes on 25%. So he's, uh , confident, but not completely. In short, it's a much more balanced view of what's probably actually happening. This comes at a time when companies are regretting cutting staff for the sake of cost savings, which is forcing them to try to rehire all laid-off workers.

Okay, I hate this argument because it's complete nonsense. So, here's what's really happening. So, there are people, and I'm sure a certain number of them just gave in to the excessive hype. It sounds like, "Oh, AI can do all this." "AI has limitations. AI fails every now and then. There are things it does well and things it does terribly. The number of things it does well is increasing every day. The number of things it does terribly is decreasing every day. And so, um, what's probably going to happen to jobs is that some of them are just going to disappear completely.

They're never going to exist again. No human will ever do that again . Um, data entry, um, data analysis, things like that—that's just going to be a thing of the past. AI is doing it much better. Or, if it doesn't disappear, the number of those jobs will be drastically reduced. Same with driving: eventually, one data analyst will be able to oversee a thousand AIs working, collecting data and processing it, and they'll just be checking that everything is working properly. So instead of a whole department—one person and a bunch of AIs.

But that doesn't mean that new jobs won't be created. So this the story we start telling ourselves, like, "Oh, we thought AI was cool." Oh, that's not really cool. We have to hire all these people back.” …that’s not true. First, just because you fired a bunch of people and now you’re hiring new people doesn’t mean you’re looking for the same skill set. You may have fired everyone who was resistant to AI. Now you’re hiring people with similar job descriptions or even the same job title, but you’re already hiring a different type of person who will scale AI in their jobs, as opposed to someone who might have been very resistant to it or just didn’t know how to work with it .

Um, they ’re not “digital natives” or whatever. So this is where you have to be very careful before you buy into this narrative. If you’re not using AI, it’s very easy to fall into the hype: either believing that it can do more than they say (because you haven’t tested it), or falling into a doomsday pessimism like, “This is all [ __ ], it’s not real, and it’s going to be gone soon.” ". But if you're in this every day, like me, you quickly realize: yes, there are a lot of things it does poorly, but damn, new features and capabilities are coming out all the time.

They're coming out so fast. It's truly impressive. So, in short , don't be fooled by these narratives. Market research from Robert Half shows that about 29-32% of companies that laid off staff due to early AI predictions have already started bringing employees back into the same roles. Which begs the question: What the hell were these companies hoping to do to live up to those predictions? Even though companies are hiring again, finding a job in this market feels like an impossible task. Whether it's because of "fake" job postings or the extinction of entry-level positions, finding an employer who will actually look at your resume and give you a chance is like finding a needle in a haystack.

If you're trying to make the career progress you want, career A coach from today's sponsor, strawberry.me, can help you. We let these people rate their content. strawberry.me connects you with a professional coach who will help you stay focused on achieving your career goals. Once you've been matched with a coach, you'll have one-on-one video sessions to assess your current situation. This is different from a typical coach or therapist: instead of just discussing your feelings or why you're stuck, they get straight to the point of achieving results.

Your coach will help you get a clear picture of where you want to go in your career and what steps you need to take to get there . They'll help you identify and overcome blind spots, close skill gaps, and create a structured plan. So, if you know you're qualified for a better job but feel lost in the sea of online job openings, head over to strawberry.me/colehaastings (it's the first link in the description) and you'll get 50% off your first coaching session. Meanwhile, while these companies are hiring again after layoffs, one country is beating us in AI by a huge margin and at a much lower cost— China.

If there's one thing China does well, aside from turning its cities into social media cyberpunk cities, it's building scalable, low -cost products, accepting lower profits in order to capture the market and beat international competitors. Okay, I agree with that. China is extremely good at allocating budget funds to industries so that they thrive, and they're very good at optimizing. To say they're ahead of us in AI is a lie. They're catching up fast, though, and that's certainly an area where we need to keep our finger on the pulse from a national defense perspective.

But it's pretty clear now that the United States is the real vanguard of advanced AI models. China is following suit, building open-source models that work, let's say , quite well and are much cheaper. So I would said they would win the price war, but they haven’t won the real battle for technological supremacy yet. So the question is: how important will that supremacy be? On the road to superintelligence, that supremacy becomes extremely important . And their open-scale AI model could soon capture a significant share of the market.

True. Released on July 27, 2026, Kimi K3 is an open-scale model from the Chinese company Moonshot AI. Unlike the closed-scale models OpenAI and Anthropic, Kimi K3 can be downloaded, modified, and run on your own hardware without relying on Moonshot’s servers. Which, by the way, is just great, and we should strive to see more of it. Unfortunately, this could negatively affect the leading research models in the US. It might make it harder for them to make the kind of profits they were hoping for, but it’s ultimately better for people .

Being able to get AI for essentially free—for the cost of running your own local server—is It's really incredible. And they're already saying that it reduces the cost of tokens by, I think, 60%. So that's pretty impressive. Because of its efficiency, the Kimi K3 can run for a fraction of the inference cost of many leading models, while still delivering comparable performance on many benchmarks. By the way, can we just stop and laugh at how ridiculous the graphics are ? Well, what the hell does "maximizing intelligence per dollar" mean?

Well, I'm glad you asked. That's, oddly enough, one of the most important things about the idea of ​​AI. You see, in the end, whoever can get the most intelligence for the least money wins. So that's the whole point. Now I get it. It looks a little funny on the spreadsheet, but it's really like saying, I can get the smartest employee in the world for the cheapest price. And whoever can do that, who can hire the smartest people for the least money, will have a huge advantage for their companies.

It’s just a fact. So now the question about models is: How much intelligence am I getting for my money? That’s value for money, but in AI terms . In any case, Chinese models are already starting to take up a huge share of AI traffic. According to OpenRouter, Chinese models increased their share of token traffic from 1.2% in 2024 to more than half in the summer of 2026, overtaking American models in terms of total volume. Think about the potential implications of this. Some of the largest American companies have made huge investments in OpenAI and Anthropic.

In the first quarter of 2026, other revenues accounted for 60% of Google’s revenue , 51% of Amazon’s, and 27% of Nvidia’s. Some of these investment returns come from stakes in leading AI companies like Anthropic and OpenAI. If a significant portion of the market switches to Chinese models, OpenAI and Anthropic could see revenue shrink, reducing the billions they're spending right now on Nvidia chips, cloud infrastructure, and huge new data centers. But if you want the AI ​​bubble to burst faster, that would be great. Yes, that's exactly what we've been talking about before.

China is a real threat to the AI ​​industry. If they start washing away some of the revenue, it's going to slow things down , and we're back to the railroad argument. If debt is growing faster than revenue, the whole industry becomes a series of questions about whether they can get to positive cash flow before the debt explodes. And if the debt becomes unsustainable, and they need to raise new money to stay solvent, but they can't raise it because the revenue is coming in so slowly—the industry suddenly grinds to a halt.

Now, if it grinds to a halt, it's not necessarily the end of the world. The problem is if the company never reaches profitability and can't become financially independent. Here What's the question? If you look at Microsoft or Google, they're all cash flow positive. So if they stop building , they'll be fine. But Anthropic and OpenAI are not. They're not even close to being cash flow positive. So these guys could really implode from within. So it all comes down to this: What do you think will happen?

Will the U.S. government step in or not? Your answer to that question tells you how big of a risk you think we're at. Let's talk about the next threat to AI. The rise of negative public attitudes toward anything related to artificial intelligence. From 2023 to 2025, the percentage of Americans who said the impact of AI would be very or somewhat positive actually went from 15% to 27%. However, over the same period, the number who said it would be very or somewhat negative went from 40% to 47%.

Yeah. Hmm. I wonder why broad segments of the population hate AI so much. Could it be that the people who are building these models have weak moral compasses and only care about growth and profits, not about helping humanity? Not really. I think people tolerate an incredibly obscene amount of immoral nonsense and offer little resistance. People care about how their own lives are. If my life is good , then I'm fine . If my life is bad, I start looking for reasons. I think AI creates such a level of anxiety that people can't look at it from a fundamental perspective, and that's what creates a lot of our problems.

If you think AI is going to make your life worse, make it meaningless, take away your purpose and economic prospects in the future. If it creates so much uncertainty about the future that you feel terrible, you just dismiss it. People say it's because OpenAI is run by a moral monster. That's not the real reason. You know, I think AI will probably end the world, but in the meantime, there will be great companies built on serious machine learning. Could it be? No, that's a terrible thing to say.

There's no doubt about it. It sounds completely absurd, but I don't think it's a real driving force. Because of the fact that working-class neighborhoods are being torn down to make way for data centers that make incredible noise and suck up all the water? Or maybe it's because of the flood of AI content that's fueling entire dropshipping empires? How are you guys doing? Smith's on the phone. Today, I'm going to show you how the video will be. Now, again, I'm going to do a deeper study of data centers.

I think the reports that come out of cities that have already opened data centers are calling them a real miracle, because they generate so much tax revenue that they create more jobs. Now, I'm convinced that they bring in much more more good than bad, but we'll get into that later. But AI garbage is a whole other story, and people are going to have to find a way to clean it out of their lives. For example, this video where I had to turn off the music because of copyright issues is making me about $30,000 a week.

Maybe it's the way it manipulates or acts like a little kid. Hey, can you see me? I can't see you yet . If you want me to see what's going on, you need to turn on the camera. It'll get better over time. I just turned on the camera. Okay, okay, now I can see you. What's going on? What do I look like? I see a person in front of... If you're not looking at the camera, he didn't turn on his camera. that many uses of AI are aimed at destroying human connections, critical thinking, and creativity.

But perhaps the most egregious example that really annoys me, especially as an author who just published his debut novel ( link in the description ), is that AI companies are destroying rare books to train their models. The scandal dates back to last year, with the first major copyright settlement for AI. Court records, first reported by the Washington Post, revealed that Anthropic was buying millions of paper books to build a database to train its models. The internal project, called " Panama," relied on a process called destructive scanning.

Machines would cut the spine off the book, run individual pages through industrial scanners, and then throw the original away. Their justification? Well, physical, old books are certainly not AI trash, and since AI is learning from its own trash, books written by humans have become extremely valuable to them. I understand why some people do n't like it, but I have a completely different view of it. We have a problem in America right now : kids can't read. Kids are n't running out to buy these books. We don't want to risk losing some huge treasure trove of knowledge because that AI is destroying these books.

We risk losing this information forever because it's stored in a format that humans don't even work with anymore. So, in any case , I find it much more interesting to record these books. And if we could record these books in a way that they could be read later, that's what I would choose. But at the very least, feeding them to an AI will allow it to learn from the patterns that are there, and you can interact with that knowledge. In my opinion, being able to interact with the patterns that are discovered is more important and more useful than just reading the book itself.

Of course, I would never want to have to choose between those two options. I don't want to have to choose between books disappearing and me never hearing anyone's words, and just having the generalized patterns that emerge from them. That would be terrible. But if you ask me, does it bother me that they take one copy of a book and destroy it, knowing that there are a limited number of books , but they don't destroy all of them—no. Also, I did some additional research on this issue.

It seems like they’ve overstated the situation a bit. It seems like it was Anthropic, they reached out to a lot of bookstores. The vast majority of them didn’t even respond. There’s no evidence that they’ve received millions of books and are destroying them. So I guess it’s a lot of ado about nothing, although I still wouldn’t want those books to be lost to history. Anthropic was ordered to pay $3,000 per book, and most of the authors they “borrowed” content from were paid. However, the court documents didn’t reveal how the companies found the millions of physical copies in the first place .

A possible answer comes from a report by 404 Media. ISBN DB, known for its metadata support for International Standard Book Numbers, appears to be gearing up for the AI ​​boom . The company advertised a service that would help AI developers get anywhere from 1,000 to 1 million printed books for training large language models, describing it as being designed for “ large-scale AI queries.” The web pages were later taken down, and ISBN DB said the service never launched. So we still don’t know exactly how Anthropic and other AI labs got their hands on these books in large numbers , but the existence of the offer suggests that a market for supplying physical books for AI training was already forming.

The sneaky part is happening now. This month, Dutch secondhand book dealer Peter de Vries received a strange email from someone named Natalia, who represented a company called 2077 AI. She—or the company—wanted to place a fairly large order of books and attached a spreadsheet of over 3,000 ISBNs, asking for the corresponding titles, prices, and estimated shipping costs to China. At first, the secondhand book dealer dismissed the request as spam and barely read the first few lines. It wasn’t until a few weeks later, after a journalist investigating the request contacted him, that he realized it had something to do with the industry.

AI. He later shared an email and a spreadsheet with Fortune, revealing a list of 3,001 mostly academic titles published between 2020 and 2021 by Elsevier, Wiley, Routledge, Oxford University Press, and Emerald Publishing. Now imagine that these are purely academic books that no one will ever read again. They're gathering dust, and someone says, "We're going to turn this into useful knowledge that the whole world can use." And if they're headed to China, they're likely going to end up in an open model that everyone can access, and we're somehow complaining.

That's the part I don't understand. It's like people trying to protect something that they never cared about until they heard that someone else was using it. If they were just burning books and not wanting anyone to have access to them, I could understand why people would riot. I would riot myself. But the reality is that we're translating this in a format that we can all use that information in. And I actually completely reject the idea that when you upload something to an AI, you're stealing it.

Is there a way that it can be stolen? Is there a way that you can take, like, oh my God, what was that? Figma? Figma uploads all of their data to OpenAI or Anthropic, they're really excited, you know, they work with the code, they create something new, and then one of them ends up turning it into their own version of the software and selling it. Okay, cool. It literally means you take it directly and then you republish, essentially, a free version of that software. Okay, that's [ __ ].

If someone took, say, Stephen King and then republished Stephen King, that would be bad. But if you learn from all the authors, listen, I know that happens to me. I know that with the thousands of hours of my content that's out there, other people created " Tombots" that show you what it would look like if you asked me a question. And the answers are often pretty accurate. And that's just the way it is. To have access to all the other information in the world, it's a small price to pay to say, "All my old content is now there, just as any of you could look at it if you wanted to, and then learn how I answer questions and do the same." But to me, it's...

you've heard it before. You might hear, "Oh, this person is so-and-so's kid, " meaning intellectually. They go and learn from them, and essentially relay their ideas. That's how it's always worked: we all learn from the people who came before us. We say we stand on the shoulders of giants. The fact that we've now made it much more efficient, by taking all of human knowledge and making it accessible to anyone with an internet connection, is a trade-off that's well worth it. Now, moving go ahead, I'm the only one who can create new versions of what I would say to make something amazing, and so no one can take that away from you.

But in any case, I reject the idea that this is theft. According to the seller, several other Dutch second-hand booksellers have received the same request and also assumed that it is a scam. Yes, these companies are no longer just after mass-produced books. They are hunting for rare antiquarian editions. It's almost like a second attempt to burn the Alexandrian her library. AI companies, so again , knowledge is preserved. They are not being destroyed. AI, desperate for valuable human data, resorts to removing the spines of antique books, running pages through scanners, and throwing away text that will never be read by other eyes again.

Because people are already so busy reading this. Despite the talk of synthetic data replacing human knowledge, the industry's most valuable resource remains authentic human writing, and it is becoming increasingly scarce. This is where things get weird. On the one hand, AI companies are scouring the world for increasingly scarce human knowledge because they still need fresh, high-quality data. On the other hand, investors are betting that AI is about to become smart enough to self-improve , completely eliminating the need for humans. These are two completely different stories.

And whether the second one comes true may determine whether today's AI investment boom will look far-sighted or become one of the biggest bubbles in tech history. Of course, as with all videos, I'll try to leave it OK. So, the incredible thing that everyone needs to realize is that it's really important to have an understanding when you're thinking about your own portfolio, what should I do with my money, so that you have a better understanding of what AI is likely to come up with . What metrics matter?

So I think my view of where we are lies somewhere between the pessimism that you've seen here and someone like Raul Pal, who is constantly set up only for the market to fall. He says AI can't get too big. There's no amount of money that could be invested that would make him think, " Okay, this is complete madness." Um, this is the most important technology of all time. It's better to think of it as an arms race. This means that the US government will support it.

The Chinese government will support it. These are not industries that will be allowed to fade away. However, debt matters . The rate at which, uh, your income grows compared to the rate at which you take on this debt matters . For the people who invested, not necessarily for the long-term viability of the technology itself. And the thing that I find really interesting, what I'm , uh, focused on right now , is what , uh, Jensen Huang is talking about, about the fact that right now, at least today, the depreciation schedule for this equipment is not what people think.

It is longer now. Although it's probably a little more shaky than he wants you to believe in terms of things that could knock him off that path, the fact that we haven't yet been able to meet the demand for intelligence, computing, uh, tells you something. That, despite all the money that's pouring into it, it's not like the railroads, where we just have to wait and wait. Income is growing at an incredible pace . Usage levels are growing at an incredible rate. While we are using it, innovations are emerging in both the US and China.

Expenses are decreasing, but revenues are still increasing. Whether this will remain true in the long term is a more complicated question. And this raises the last thing that everyone needs to analyze: not whether we are in a bubble, but where exactly are we in the bubble of this bubble? I think this is a perfectly reasonable answer. If you enjoyed this conversation, check out this episode to learn more. I want you to explain socialism, but don't use the words " wealth" and " taxes." So, this is a video in which we look at the real causes, effects, and mechanisms of how economic systems emerge.

Go deeper

• All Impact Theory episode guides