Is AI Trade DEAD? 🤖 $835B Capex vs $2.4T Backlog: The Real Math💥

Source: YouTube

Welcome and framing

Speaker 1: 00:05 Hello, everybody. Yesterday was a special day. Shout out to Silicon Phoenix: first day waking up in a world of flying cars. That demo was spectacular, but it's also causing more polarity across the two different sides of the aisle.

Today we're going to talk about whether the AI trade is dead or alive. There are a lot of signs; it's been a very rough week. We're going to break out all the numbers, all the charts, and go deep into why it could be dead and why it could be alive. Then a very detailed quant verdict at the very end, because that's what I do here: if it's dead or alive.

Thank you to the moderators: Cheshaw, D Cypher, Sliver, DD, Tesla, K8. Let's go. This one is a little tricky, but boom or bust—that's the button we press to determine yes or no. It'd be fun to do a little poll: is it dead or alive? If somebody can do that, let me know.

This is not financial advice. Shout out to Hodlnader: “Glad to be back on Patreon. Missed you guys.” Some people leave for the bear. Never leave for the bear. The bear is where all the goodness is; that's where the billions are made. Anyhow, let's get into it.

Sentiment and market performance

01:22 Sentiment and signals: Fear and Greed is at 19, extreme fear in the stock market. I'm not talking about the crypto market today because that's not relevant.

Look at the week we've had. This snapshot I grabbed about half an hour ago, and we have gone into the power hour, so things bounced off these levels. But you can see it was a brutal week across the board. There were no survivors. Nobody got away, unless you were in oil or interest rates; you probably were okay. There's some rotation into some of these stodgier sectors.

Micron down 14%. Of course, the Kospi, South Korea, got hit hard. Marvel nearly 17%. ALAB nearly 16%. ASML 7%. Tesla 7%. SpaceX 11.6%. Palantir, at last, fell down. But remember, it's been going up nonstop since 106, hit 210, and did a little mean reversion. Cerebrus, very bad, down 22%. SK Hynix down nearly 19%. Solana down 7%. Michael Shaggy down 12%, on rumors they're going to be booted from the MSCI or MCSI World Index Fund.

02:42 But this is just a week. If we look at what's happened year to date, the numbers are wild. Micron over 200%. AMD 150%. Marvel 170%. Alabs 72%—started strong, fell down hard, grew again. ASML 54%. That's after many of these names lost 20% this week. This is still very positive for the year, so always have that zoom-out perspective.

The bust case: cycles and construction constraints

03:08 Let's start with the tricky part. I always like to start with the bad news first, get it out of the way. This is the case for the bust. I'm going to go through this pretty fast and pull up all the research I did and all the stuff I monitor and track every single day, because I'm heavily invested in AI. I need to know if it's going to survive or not.

First of all, Bank of America: they come out and think, “We're seeing signs of boom and bust that we've seen before,” historic cycles, et cetera. All the people who lived through the dot-com bust still have the scar tissue from that. They're hoping this will be the same thing. That's why they believe it's the same all over again. They believe people are not investing; they are cheerleading. That's why the banks are saying, “No, things are still slowing down. Things are looking bad. Things look ugly. This is classic capital-cycle risk.”

Thank you, Boo McDuck. Happy Friday, everybody, as well, and thank you to the mods in the chat.

04:15 In addition, data centers—this one took me by surprise. It's from TechRadar. Data centers are more unpopular in the United States than the Iran war. Think about that for a second, because there are groups spreading FUD about data centers: “They use all the water. They make noise. They do this.” Remember, these are like—it's like trying to listen to here or factory. It's all FUD. Anyway, I'm not going to go there.

What is crazy is nearly 70% of the US population are against AI data centers because they were told to be against them. Can they stop progress? We shall see.

The other problem is these things take years to build. Hooking up the electric that's required, whether it's before or after the meter, is extremely difficult. Building them is extremely difficult. Getting permits takes years. All the power and electric, the facilities, the wiring, et cetera—not just sourcing the semiconductors or the memory, which is also a headache—is really difficult.

Building takes years. There are a lot of big talkers saying, “We're going to have 10 gigawatts, 20 gigawatts, 50 gigawatts of compute next year.” No, they're not. There's only one player that can pull that off. We'll get to that later.

05:36 The other problem is the bottleneck is extremely physical. Again, this is not the chips, but all the other stuff: the skilled trades, the connectivity, the substrates, the data centers, the land, the power. All of this takes a lot of time.

There is huge demand. They want to try to deploy another 43 gigawatts of new AI compute next year, but that's going to be very difficult, many believe, with all the restrictions. OpenAI and Anthropic both want to add more than five gigawatts each. That's 10 gigawatts in total. Ten gigawatts is a lot. We'll share some details around exactly what that means in a minute.

The bust case: credit, IPO timing and the capex bill

06:23 Also, the trade-credit market for AI is getting completely smashed this week. If you are a neoscaler, one of the neoclouds, you're getting absolutely brutalized.

I have been talking for a long time about things like Oracle potentially being in hot water. CoreWeave, for the last 18 months—the financials don't make sense to me. Now the market is painting with a broad brush, and they think everybody is going to get hit by this. The credit markets are flinching, which means the credit dries up, which means it's very difficult to find money to fund these things.

Credit is getting much more expensive. We have the Fed raised 25 basis points, 10-year Treasury is rising over 100 basis points, close to 6%. Because AI is soaking up so much capital, that actually also raised the discount rate. It makes it harder to raise money, even for other companies and also for the hyperscalers.

The credit-default-swap rates are at record-wide spreads as well. You can see some of the top players here. Oracle, really high; CoreWeave as well. All the other players are on there too. It's getting expensive to fund and finance these things.

07:39 In addition, we just found out Anthropic is weighing delaying their IPO beyond November. They were supposed to IPO in October. They said, “Let's wait till November.” Now they're going, “Maybe not November.” They're also toying with a valuation rumored down to $1.5 trillion from $2 trillion. That may reflect some of the recent numbers too. It's like somebody's put a whole bunch of sand in the gears of this whole AI boom, big time.

The bill is huge. This is the bill from just the big four or something: Amazon, $220 billion of capex, up from $200 billion. Google up to $205 billion, up from $195 billion. Microsoft, $175 billion. Meta, another $145 billion. Big four, tons of money: $735 billion to $750 billion, and next year even higher. Add in Oracle, another $100 billion, takes it to $835 billion. It's still going up.

How are we going to pay for this? How are they going to pay for it? Can they even pay for it? That is all the negative stuff around the case against the AI trade. Could it be cooked?

The boom case: intelligence and diversified demand

08:50 Now let's talk about the case for the AI boom. This is the same approach, looking at the same different type of data. I'm trying to put a picture together in my little mind.

First of all, Joe in the channel yesterday—we were talking about how I've always said, for years now: have you ever heard anybody saying they have enough intelligence? The answer is no.

UNKNOWN: 09:13.040 So.

Speaker 1: 09:13.520 As you get more intelligence, you want more intelligence. That's the whole Jevons paradox. As tokens get cheaper, as intelligence gets cheaper, it'll drive more demand. The penetration is still so low right now, and training has gotten so much better.

It used to cost $20 million to train the first GPT-1 model. It can now be done for $100 today. It used to cost $20 million; now it's $100. That's a million X. Usage is supposed to go up a million X as well over the same decade.

We have that math problem that had 10,000 agents solve it, $4 million of compute, that humans couldn't solve. Math is completely cooked as well. Everything is changing very fast. Tons of demand, and demand comes in thirds.

09:59 A lot of people think, “All the money, all the demand, is just coming from OpenAI, Anthropic, et cetera?” No, it's not. This is from Jensen. He talks about his market segment for NVIDIA GPUs. One third is classic computing. One third is across five to six frontier labs, which includes names like SpaceX and Google, Amazon, OpenAI, Anthropic, Meta.

Then you have another one third. It's beautifully spread across startups and neoclouds. There are thousands of them, and they all use open-source models. We have a new group coming in, the sovereign AI group, driven by players like Palantir that want people to build their own AI using open-source models as well.

The other key point: a lot of people say, “If the market cap of OpenAI and Anthropic is so high, they need to do a trillion dollars in revenue.” The answer is no, because there are rumors that some of this business they do is very high margin. We'll get to that in a second.

Inference, cloud growth and investment capital

11:02 This one is inference. I always talk about how, in the future, inference will be everything. But inference is also where the money is.

This little excerpt I copied here from Semi Analytics, I think it was, talks about how the gains per GPU are incredible in terms of efficiency. Pricing has gone down, and blended prices per million tokens have dropped in the past few months. Despite that, inference margins for the frontier models have gone from 40% to over 70% just in a few months, despite prices falling.

It's extremely lucrative. In the future, when everything is learned and all the models are stable and mature, 99% of all compute will be inference. Ninety-nine percent. If that's lucrative and high margin, it's great for everybody.

11:54 Revenue is beginning to show up. People thought, “The Mag 7 are spending far too much on AI and compute.” But look at the revenue growth for these companies: Azure up 43%, Google Cloud up 82% year to date, Amazon up 37%. That is extremely big cloud growth. The investment is paying off for them as well.

By the way, there's no compute available. If you're looking for compute out to 2027 or 2028, it's all sold out. There's none available. More has to be built.

The other thing—I touched on this briefly yesterday—is: are we early or late? This ties into the Jevons paradox as well, but this is the double exponential. First of all, AI needs more compute per task. The reasoning models, and as AI agents explode, they will need more options and, of course, more compute, and more demand for more tasks. Everything will go up to the tune of a billion X from here. There is no slowing down by any stretch. These AI agents, these bots, are going to drive things parabolic.

13:00 Also, the money is still coming. It's $3 trillion to $4 trillion—the AI infrastructure spend. You want to follow the money because that will all trickle down to your investments, as long as you're positioned correctly. AI infrastructure is expected to reach $3 trillion to $4 trillion per year by 2030.

You've got all the FUDsters, all the Bank of Americas, all the newspapers loving this. They love to grave-dance and say, “The AI boom is over. The money's dried up. It's gone.” No, it's not.

What you also have to understand is this is the only growth in the economy. Look at the people close to the Cantillon machine, close to the money printer. What are they going to do? Think of the Apollos, the BlackRocks, et cetera. They get all the money, and they have to deploy it. Where are they going to deploy it? In cellular carriers, food makers, Nike, retailers? No, it's going to go straight into AI, because that's where the return is. That's where the growth is, and that's what people miss. There is tons and tons of money to be there, and it's going to continue to grow.

GPU supply, TPUs and spending commitments

14:06 This is a good-news, bad-news piece, so it's hard to figure out where to put it. I put it in the good-news area because this could be construed as a headwind for NVIDIA.

NVIDIA has given us notice that they're expecting to double growth next year. They have infinite demand for their GPUs. In fact, it's rumored that 40% of the Vera Rubin supply has been locked up by Elon Musk alone—one player. Remember, five players can buy everything. What about the thousands of other companies trying to get their hands on a Vera Rubin? They can't.

Apparently, on the black market, they cost $5 million each, when normally they used to retail at $2 million. These things are like hen's teeth, very hard to track.

The good news is this impacts other parts of our IA13. Demand for Google TPUs is off the hook. The expected growth rates, as you can see here, go from TPU sales of $2 billion in 2026 to $18 billion in 2027, $43 billion in 2028, $52 billion in 2029, and $60 billion in 2030. It's tapering off a little in 2029 and 2030, but the point is huge demand.

Who makes those TPUs? The AMDs, the Broadcoms, the Marvels, et cetera. Again, there are not enough chips. There's this huge demand for chips and this huge amount of capital chasing all those chips. Don't fret that stuff. It's very good for everybody.

There's no slowdown in capex spending. We haven't heard any players saying, “We said we're going to spend $200 billion, and we're just not going to do it at all.” That has not happened. In fact, what they have said is they're going to spend even more next year—a lot more—and they have to.

Frontier-lab revenues and compute plans

15:51 In addition, although Anthropic, you could argue, has tapered off a little, when you look at the growth quarter to quarter, they did $4.7 billion in Q1, $11.6 billion in Q2, and $16.8 billion in Q3. I don't know if they're going to get higher in Q4. We shall see. There's a lot more competition now than there was before.

The point is, annualized, they hope to get to $110 billion, which means a little over $26 billion, $27 billion a quarter in Q4. We'll see. OpenAI is looking at $50 billion annualized, but it could get a lot higher. It's probably going to be more like $70 billion annualized by year-end, given the growth rate I'm seeing. These big two are not slowing down.

16:35 In addition, SpaceX has told us they want to get to 2.7 gigawatts of compute by year-end. That would be $100 billion ARR. Remember, they're hoping to use 60% of their compute for themselves, for Optimus and other things, and lease 40% out to other players. The other players have infinite demand for what they want, and they can get the return on investment for every gigawatt they turn on in less than a year, per the CFO.

Will they get to 2.7 gigawatts by year-end? Elon thinks so. This is his map, his plan. Shout out to Michael Dell, Elon Musk, Jensen Wang, and everybody else.

How much will they get to by next year? Elon Musk: more than 10 gigawatts by year-end. So, 2.7 by this year-end, 10 gigawatts by year-end 2027, which means adding nearly two gigawatts a quarter, almost—a little less—to ramp that. To pull that off, they probably need $100 billion in GPUs as well. The plans are there. The stories are huge, but there's more.

17:51 As I mentioned before, OpenAI and Anthropic plan to add 10 to 12 gigawatts of compute next year. That's between them, five to six gigawatts each. For that to happen, they need a player that can do it. Who is that player? I don't know, but I think Elon Web Services, SpaceX, will be providing a lot of that data, because the others don't have it.

Google doesn't have extra compute. Amazon doesn't have extra compute. They're all going to have to go somewhere. Microsoft is running out as well, and that's why they're building their Maia chip.

OpenAI and Anthropic's combined revenue is expected—this is a crazy number—to go from about $100 billion this year to over $500 billion in the next 12 months. I don't see that happening. SemiAnalysis came up with that number. I do not see them making half a trillion dollars in revenue in the next 12 months.

But if they do, can you imagine how much money would be floating around? Even if they get to $200 billion at a 10X multiple, 20X multiple, that's $2 trillion to $4 trillion in value there. Then they can be very easily justified.

Again, they say they will get there using only about 30% of that compute for inference—that's the high-margin stuff. Seventy percent goes to training. In fact, it's rumored that OpenAI is taking a huge amount of their training compute and putting it towards versions six and seven that they're keeping for themselves.

I think they're going to try to develop drugs and become drug makers, cure things like cancer. We are the beneficiaries of all this investment. Our lives get better. This is what people don't understand. They're anti-AI, but it's actually very good for us. Tell your friends. Convince them, especially when it comes to health and medicine.

Token demand and the memory crunch

19:38 Another simple little chart: the amount of token demand is real and rising. Can anybody see any blips in this chart? Some slight little ones, maybe from some weird anomalies or how it was measured, but the trend is up. That's up to, I think, 800 billion tokens, or maybe 800—I don't know what that measurement is. But it's monstrous, and it keeps on demanding.

20:08 Then let's get to memory. This is another big piece. I know this week the memory stocks got clobbered. Whether you're in Samsung, SK Hynix or Micron, they got completely smashed.

Then Michael Dell says he expects the memory crunch to get worse. There's already no memory. You can't get it. You cannot get it. No memory anywhere. They're completely sold out for next year.

Elon said, “For sure in 2027, but if China ramps LP5 and LP6 as fast as they've ramped other manufactured products, 2028 will be a different story.” But this is the difference: high-bandwidth memory is a tougher hill to climb, and that's not what the Chinese have, although a lot of people believe they do. Grand Duke, thank you for coming. It's not true. HBM will still be extremely hard to get.

This was rumored to hit the wires yesterday: Google is trying to secure high-bandwidth memory not for 2028 or 2027, but out to 2035, because there's none. Everybody is trying to lock up supply nearly a decade out. Think about that. That's Google going to Micron and saying, “Please, sir, can I have some more?”

If anybody here is old enough to remember that movie, let me know. I thought it was a good image that I hobbled together yesterday, thanks to AI.

Can the companies pay for the buildout?

21:29 Let's get to the conclusion. A little bit of numbers here. I'm going to measure up, in a fair way, the pros and cons and see if it's dead or alive.

First of all, paying for cash: is it over? Can they afford it? Are we running out of cash? Is there a trade-credit risk for AI? The answer is no for over half of them. They still have enough free cash flow to pay for all this capex.

I think one player was slightly negative for one of the quarters. Amazon was slightly negative, but the rest are all free-cash-flow positive: Microsoft, Alphabet, Meta, making tons of money. They're using that to fund their capex.

Then the verdict is yes for the big four, with conditions, because of the growth in their revenue, the cash flow that's spinning off, the amount they're investing, and the depreciation timelines for GPUs.

A lot of people think GPUs would burn out in one or two years. No, they last five to six years. There are companies that have H100s that are running now for seven years, no problems, and they're still renting them out at top dollar.

The credit will stay open for these big players. I'm not sure about Oracle and CoreWeave, et cetera. But when it comes to Microsoft, Google, Meta, NVIDIA and all these players, not a problem. They'll have no problem securing funding, having free cash flow to pay for it all.

Bust versus boom: the quantitative verdict

22:52 I'm going to turn off my camera for this one. This is my quick-and-dirty little summary.

If you add up the math on the left, on the bust side, you've got about $835 billion of capex in 2026. Bond issuance has gone nutty, up 79%. Alphabet's Q2 cash flow did turn negative just for a minute, but it's going to be back positive next year. You've got the CDS rates for CoreWeave and Oracle, totally in the toilet. Power is short by 10 to 15 gigawatts. There's some circular financing in there as well. That's all the nasty side.

On the boom-case side, you've got $2.4 trillion of backlog. You've got money coming in from all the big Wall Street players. Cloud is growing at 82%. Anthropic revenue is growing by 3.6X. Labs are targeting 5X growth next year. There's $3 trillion to $4 trillion of planned AI infrastructure investment per year in 2030. It's not slowing down.

The backlog covers the 2026 capex by nearly 3X. When I look at the simple numbers in an objective manner, the good far exceeds the bad. Are we going to go bust? No. Are we going to go boom? Yes. Will there be casualties along the way? Yes, that's always the case. When you're running at 100 miles an hour, breaking eggs, of course there will be casualties.

Let's finish, everybody with me: “Do you're so basic.”

Closing perspective

24:27 Net-net, the AI trade is not over, my friends. Not at all. But there are some signs of pain out there in the marketplace.

This is the fun little “Come on, do something” meme guy poking a stick. Micron is the box. Micron blew out earnings. Forward guidance is insane, and the stock didn't move. I think a lot of people felt a bit hurt by that one.

Remember the year-to-date heatmap of the returns: 220% for Micron, even after the money it gave back this week and last week. We just have to get through the storm.

September—my fear was September was good, which is normally not. Therefore October will be bad, and that's exactly what's panned out.

With that, my friends, don't lose faith. Keep on stacking. If you have not enough AI exposure, things are on sale right now. We are here. We've got a little power, a nice little recovery out there too, which is good. Things are even up a little bit in aftermarkets.

I hope you all learned something new today. Once again, thank you to the mods in the chat. Thank you to Boo McDuck, DBF and Grand Duke, and of course Cypher, Sniper, Jacob, Mavasu there too, TND, Tesla, Sean D and K8. Thanks, all, for coming. Looks like it's sticking up a big Friday. See you all soon. Bye.