A (0:01)

When the holidays start to feel a bit repetitive, reach for a Sprite Winter Spiced Cranberry and put your twist on tradition. A bold cranberry and winter spice Flavor Fusion Sprite Winter Spice Cranberry is a refreshing way to shake things up this sipping season, and only for a limited time. Sprite obey your thirst. Ford Bluecruise Hands Free highway driving takes the work out of being behind the wheel, allowing you to relax and reconnect while also staying in control. Enjoy the drive in blue cruise enabled vehicles like the F150 Explorer and Mustang Mach E available feature on equipped vehicles. Terms apply. Does not replace safe driving. See Ford.com BlueCruise for more details. Running a business comes with a lot of what ifs, but luckily there's a simple answer to Shopify. It's the commerce platform behind millions of businesses including Thrive Cosmetics and Momofuku, and it'll help you with everything you need, from website design and marketing to boosting sales and expanding operations. Shopify can get the job done and make your dream a reality. Turn those what ifs into Sign up for your $1 per month trial@shopify.com specialoffer.

B (1:23)

All right, so social media consensus has converged on the fact that or the belief that the AI current finances and economics are a bubble. And I have a little bit of a contrarian take. Now, I'm not saying that there aren't problems, but I'm also going to say that it's not really fully a bubble. So if you want to hear a fully contrarian take, let's take a look at the structure of AI investment right now and then you can decide for yourself at the end. Now, first and foremost, social media has consensus that AI is a bubble. The conversation on platforms like X and LinkedIn are dominated by comparisons to dot com bust and even the tulip craze. The narrative points to massive spending, inflated stock prices, and a perceived lack of immediate widespread utility. So hype, bubble.com, fOMO, market correction, the greater fool theory, and all that kind of fun stuff. And you might say, like Dave, you just published a video about why OpenAI has no future. Clearly you're on the side of a bubble. I'm not saying that there aren't structural problems, I'm just saying it's not that simple. So here's one of the first structural differences is today's giants are built on profit, not promises. The.compeak, the PE ratio, the price to earning ratio, was greater than 100. So that was in the year 2000 and the valuation was sky high based on eyeballs and the future promise of eyeballs. Now today though, the pe ratio is 30x, which is still high. You know, typical pe ratios are 15 to 20, so this is still higher than average, but it's not insane. So valuations are based on massive cash flow. Microsoft, Google, Meta, Amazon generated over $300 billion in operating cash flow last year. That's not nothing. So this is just first and foremost, this is a very, very big structural difference between then and now. Now the biggest case for a bubble is in the CapEx spend. So the annual gap, there's a, there's a, they're spending $600 or sorry, $600 billion against a lower generative AI revenue. So this was, this is as of 2025. So you might say, okay, they're overspending. Why are they overspending? Why are they leveraging with debt and that sort of stuff? What are they betting on? But if they're, if their operating revenue is already $300 billion and they're basically saying, hey, we can, we can spend $600 billion to build out and capture more revenue, you know, that kind of makes sense. Like that's, that's what business investment is for. That's what leveraging with debt is for. Now you might say that's the point of a bubble. It's too much debt, they're over leveraged. But that's technically not the definition of a debt or of a bubble. So this pattern is very familiar. So when we're talking about industrial revolutions and technological changes, what we're really seeing, the structure that we're seeing, according to Carlota Perez, is we're seeing the installation phase and then the deployment phase comes next. A massive infrastructure buildout, overspending, financial capital and control and bubble perception. So we're here where basically it's like, hey, it's a new gold rush, there's a new, there's a new major technology and we really need to invest. And so then what we're going to see in starting 2027, starting 2027 or 2028 and then running through 2030 to 2032 is the deployment phase. So that's the ramp up of widespread adoption, production capital and then golden age of utility. So you might say it maps actually pretty well to the Gartner Garter Gartner Garter Gartman hype cycle. My brain isn't fully online yet. So anyways, the hype cycle, where you have the trough of disillusionment, which arguably we pass through this year and, and now everyone's looking at it realistically, like saying, okay, AI didn't become Skynet. It also didn't save the world. So now what? So now it's the, the long slope of enlightenment where we all kind of have a more realistic worldview as to what this thing is. But what I'm here to tell you is that the trough of disillusionment doesn't mean it's a gigantic bubble that's about to pop. It just means that this is a predictable infrastructure build out as many businesses start to pivot towards this new technology. So this is called the Solo paradox and the J curve of productivity. So general purpose technologies like electricity, PCs or AI today Force companies to shift spending from immediate production to intangible capital. This initially looks like a drop in productivity. So the investment cycle of the immediate costs are retraining staff, redesigning workflows, restructuring databases, buying that technology. And then the output is a lagging indicator. The new technology isn't fully integrated. So output remains flat or even decreases while the costs increase. This has happened plenty of times. So what didn't make it onto the slide deck is that, is that Solo? I think it was Robert. Anyways, this economist in 1987, I think he said, he said there's evidence of the computer revolution everywhere, except in the economic productivity data. So computers were already everywhere, companies were deploying them, but GDP hadn't gone up yet. And the reason is because going from pre computer era to the computer era, it wasn't just a matter of buying a couple of PCs and a couple of mainframes. You had to put in the networking, you had to change workflows, you had to train people. There are movies from the 80s and 90s of like, you know, oh, our office just got our first computer. And then like, you know, the people were scared of it and they like poke it and it beeped at them and they're like, ah, can you imagine actually taking that seriously today? Like computers are just everywhere. So retraining. So that's the J curve problem, like retraining, installation. We're right there now with AI. And the companies that know that AI is the next big thing, they're taking the risk of that big investment, which means sometimes spending a lot of cash, sometimes taking out a lot of debt, but they know that it's the next big thing. It's the same as the Internet, it's the same as the PC, it's the same as electricity. It's just a matter. It's A foregone conclusion that it's the next big thing. So it looks bubble like right now, but it's really just investment and build up. So here's how long that lag is. History shows that there's a pretty long lag. Steam, steam power was like 100 years. Electricity was about 30 year lag period for the J curve. Computers were about 20 years, software as a service was about 10 years. So artificial intelligence is expected to be about two to five years. So we've had this shortening of every, every cycle, every adoption cycle has gotten shorter and shorter. And artificial intelligence is actually good because we've actually spent the last 15 or 20 years building up the readiness via software as a service, which means Internet connected everything in cloud services. So AI is actually going to be a little bit quicker to adopt. And we're already seeing this because it piggybacks on software as a service and Internet services. But the point is, is that we've seen this J curve before and it has played out plenty of times with all with electricity based rollouts because electricity is what enabled computers in the Internet and computers in the Internet are what enabled AI. AI is just a fourth or fifth, you know, downstream consequence, fourth or fifth order consequence of electricity. So you know, and here's the quotation by the way, you can see the computer age everywhere but in the productivity statistics. There we go. 1987, Robert Solow. So it did make it into the slide deck. So we're not seeing the productivity yet of AI. What we're seeing is the cost of investment and the cost of scaling up just in the same way that we had to do the same thing for electricity, computers, Internet and AI and software as a service. The critical difference is building for unmet demand versus theoretical demand. And I see that it duplicated it. So here's the thing. In 2000 with the dot com revolution and crash it was, the idea was if you build it, they will come. It was a supply, supply side theory that if we, if we spent $100 billion laying cables hoping traffic would eventually arrive. It took five years to arrive. So bankruptcies due to dark unused infrastructure and websites that didn't generate traffic and that sort of thing. However, today we have a demand pull. Please stop asking for GPUs, we're out of stock. Google and Microsoft are capacity constrained, turning away high end compute customers. We see this with Claude and Anthropic where you still run out of messages. That is a sign of unmet demand. Now you might say, well that unmet demand is Mostly for free users. And they're still working out the business model, you know, whether it's going to be ad based, subscription based, hybrid based or whatever. The point is, is that the world wants more AI than we have. And the fact that we want that we all want more AI from individual consumers as well as other businesses, and it hasn't even gotten good yet tells you that this is a demand side problem, not a supply side problem. And this, this I believe is the biggest structural difference with the dot com revolution and the.com buildup. They knew that the Internet was the next big thing, but it took a long time for people to even figure out how do we build Facebook, how do we build Reddit, how do we build, you know, YouTube and Netflix and all those other things that ultimately meant that the Internet was, you know, and those are all consumer side stuff for the Internet, by the way. But anyways, it took a while to figure out what to use the Internet for with artificial intelligence. We already know what we want the GPUs to do and we're continuously exploring more things to use them for. So that's why we have Nvidia, ARM, intel, all of these people getting into deep into GPUs, TPUs, NPUs, whatever you want to call them, Qualcomm, you know, on the mobile side as well. Like the differences between fire sale and sold out. And we're in the sold out phase. The market signal is sold out, not for sale. And that's supposed to be fire sale. Sorry. The evolving shortage. The nature of the supply constraint has matured indicating a deep structural demand. 2023-2024 shortage. Individual high end GPUs were unavailable. Late 2025 shortage. Individual GPUs are available for rent, but large network clusters of GPUs are sold out until late 2026 and 2027. Also memory is sold out. There's memes going around the Internet that like, you know, sticks of RAM that would have cost you $200 last year, cost you $1,000 this year. Memory is sold out, GPUs are sold out. The demand is very, very real. That's why you see those prices going up. So the real bottlenecks are not just silicon, it's also the interconnects. So the networking fabric links power and memory. So there are multiple supply side bottlenecks that are being driven by that insane insatiable demand. And that demand is only, it's only increasing. Google noted internally that they needed to double capacity every six months just to keep up with current model training schedules. That is also to keep up with demand. So the demand is we want more AI, we want smarter AI, we want more tokens and that sort of thing. So we're experiencing the biggest industrial revolution buildup in human history. Just by sheer numbers alone, it rivals the Apollo program. Not quite. Because the Apollo program cost like, I don't know, a sign an a not insignificant portion of gdp. I don't know if this build out has rivaled the Apollo program yet, but it's bigger than the Manhattan Project in terms of equal dollars. So GPUs are money printers, not tulips. This is another big structural difference. Tulips were zero yield assets whose value was purely speculative. GPUs are capital assets with a rental yield used to produce a commodity with a market price. So the money printer, an Nvidia H100 GPU costs about $25,000 to $30,000. The annual revenue at 60% utilization as of today is about $13,000, meaning the payback period is two to two and a half years. So that is exactly what you want to see because not only. So here's the thing. Not only is, is that insurance against deprecation because you get to amateurize that whenever a company invests in hardware, it deprecates. So they get to, they get to, they get to claim that against taxes. So eventually this GPU is going to be free anyways. And the entire time that it was free and in terms of total cost, it was also printing money. So even if it's obsolete in a year or two, it doesn't matter. So this is, this is a standard industrial equipment payback cycle, similar to a CNC machine or a commercial truck. This is an investment in a productive capacity, not a greater fool speculation. And you've seen this mentioned a couple of times. The greater fool was basically you buy a tulips hoping that you can sell it to a greater fool, or you buy, you know, a meme coin hoping you can sell it to a greater fool. GPUs are not being sold like you're not reselling them. So you know, and this, this is the rapid obsoles, you know, so the real risk isn't a bubble, it's rapid obsolescence. And I, I personally don't even believe that this is a major risk for the reasons that I just outlined. But some people do. So it made it into the data. In 2000. Cisco was a massively profitable company selling virtual vital infrastructure. Its stock was so overvalued at 200xpe. It took 25 years to recover its peak price, which just happened despite the company's success. This is a valuation risk, not a technology risk. By the way, I used to work at Cisco. Even when its stock price was below its peak, it was an enormous, powerful company. I wouldn't go back because Now I'm a YouTuber, which is the dream job. But the point is, it was a good company to work for. Moore's Law on steroids. Jensen Huang said that we are on. We're seeing Moore's Law squared. The pace of AI hardware improvement is ferocious. The new Blackwell B200 chips are vastly more efficient than the H1 hundreds purchased just two years prior to the risk isn't that Nvidia goes to zero. It's that a company that spent billions on H1 hundreds in 2024 finds itself owning a fleet of economically uncompetitive hardware in 2026. This is creative destruction, not a speculative collapse. So again, this is a sign that the demand is insatiable rather than the demand that the demand is ephemeral. So, you know, yes, rapid obsolescence is a problem because by the time you even get your order of B2 hundreds, they're going to be obsolete, but at the same time they're still going to be money money printers while they're working. So we didn't just wait for the computer boom, we changed how we measure it. So here's the thing back then, this is, this is a little bit more complex. In the 80s and 90s, we didn't know how to adjust for cost to say like, what is the investment? Because they considered that like the investment in software and hardware was just opex and it was just money out the door. So they had to do what was called the hedonic adjustment. And so prices were adjusted for quality, basically saying, hey, this, this investment actually made me more productive and it made. Made my quality of life better. So a computer that was twice as fast for the same price was treated as a 50% price drop, boosting real GDP. So basically as. And to. To ignore. Ignore these for a second. The reason is because as computers and software get better and better, they basically double performance every year or two. They stay mostly the same price or even get lower in price. And they didn't know how to price that into gdp. And so this is why they had such a hard time measuring the solo paradox was they're like, why aren't we seeing the productivity boost in the numbers? And it's because it's actually deflationary because you're actually spending less money to get more value. And so they didn't know how to price that into gdp. So then the signal to watch for AI is the same will happen for AI. The objective signal that the J curve is turning up won't just be in company earnings, but in BEA methodology. So that's the, that's one of the labor trackers. So future reclassification, looking at the R D of training runs that if it's not already classified as tax deductible, it should be because it's the same as software development. So when training, when data training runs are reclassified from R and D expense to capital investment, the official productivities number will spike. This is the same pattern that we saw with ordinary software and computer development back in the 80s and 90s. So basically what I'm here to say is that people are like, oh well, AI hasn't hit GDP yet. It's partly a methodological error and it's partly a classification error. So now finally a new framework for the AI economy. The bubble narrative, which is a flawed analogy. Driver was speculation and mania. Core asset zero yield, such as tulips or websites that don't have traffic yet, or companies that don't have any cash flow yet. Demand was theoretical, so a supply push, the key risk was the greater fool disappears. And then what we actually have today is the industrial revolution framework, which is a more historical pattern, a general purpose technology and productive capital assets such as GPUs. The demand is real and unmet, so that's the demand side. And the key risk is valuation and obsolescence. So yes, we might still see some overvaluation and we might see some obsolescence. But a GPU going obsolete is not the same as a tulip losing its value. It is very fundamentally structurally different. So that's my presentation on the AI bubble. I know it's a little bit of a contrarian take, but this is why whenever I see more and more videos saying it's a bubble and there's, you know, people are paying each other, it's like, who cares? Like, look at the structural differences. Look at how this compares to other historical industrial revolutions rather than just a tech cycle. This is not just the same as a tech cycle. This is much more fundamentally structural, like the Internet itself, like the PC itself, like the electricity itself. All right, anyways, that's my rant for the morning. Cheers, have a good one.