Economy & Markets
Eye On The Market podcast
Subscribe today and never miss an episode
Join Michael Cembalest, Chairman of Market and Investment Strategy, as he explores a wide variety of investment topics, including the economy, policy and markets. Your subscription, using your preferred podcast provider, will feature new episodes with the release of each Eye on the Market publication.
Listen on Apple Podcasts
Listen on SpotifyEye on the Market Outlook 2026: Smothering Heights
Happy New Year everybody. Welcome to the 2026 Eye on the Market Outlook podcast. Recording this in the last week of December from Santa Fe there—in New Mexico—there are no fish at all here, but there's lots of quilts and incense everywhere, so that's different. So, I wanted to think about different topics for this year, and I decided to take a deep dive into this question of whether this giant moat that's supporting the global equity markets is really as indefensible as it's sometimes described to be.
And by that, I'm referring to the giant moat of the big hyperscalers and the semiconductor companies that support them. And that is 80% of the Outlook, and then we have a few pages on everything else. So in this podcast, I'm going to walk quickly through the outline of what we look at in this year, and then some discussion about U.S. equity outperformance versus the rest of the world.
So let's get into it. Is this moat really indestructible? The cover art this year is Game of Thrones-inspired, you know, vaguely, loosely. And it's a dragon, and he's sitting on top of a castle, and the and the flags of the four hyperscalers are there, and the bricks of this moat are made up of Nvidia, TSMC, AMD and ASML, which is the Dutch lithography company.
And the dragon’s spitting a fiery stream of TSMC's two nanometer chips down on the people at the bottom of the moat. So that's kind of the sense you get when you read about this moat that it's this giant, indestructible thing. And I understand why people see it that way.
And I partially see it that way. We're talking about just eight stocks here that were worth $3 trillion in 2018 and are now worth $18 trillion. So they've gone up by a factor of six in just a few years. And so here are some facts and figures about this moat. The four hyperscalers have alone spent $1.3 trillion in capex and R&D since the fourth quarter of 2022, when ChatGPT was launched.
If we broaden the moat to the eight companies, which are the four hyperscalers plus Nvidia, TSMC, AMD and ASML, they represent 20% of global equity market cap, which is kind of incredible for eight stocks. Then if we broaden it more to 42 AI-related companies, in the U.S. equity markets, they account for 65 to 75% of S&P earnings, revenues and capital spending since Q4 2022. And then in the GDP accounts, and this is based on the numbers that just came out at the end of December, tech capital spending accounted for 40 to 45% of U.S. GDP growth in the first three quarters of 2025, not GDP, GDP growth, but still, those are kind of some amazing numbers. And just to put the tech capital spending numbers in context, there's a chart that we included.
So tech capital spending was in the neighborhood of 2% of GDP in 2025. That's the sum of the peak-year spending on the Interstate highway system that was started by Eisenhower, the Apollo project, the moon landing, electrification of farming, the Manhattan Project and several of the FDR public works projects from the 1930s, including the Triborough Bridge, the Midtown Tunnel, and LaGuardia airport.
So 2025 tech capital spending was equal to the sum of all those things, you know, in real, relative to GDP at the time, dollars. So that's kind of amazing. And the stakes are high that there's got to be some really substantial productivity and profits outcome from all this spending. So here's the good news, and you've heard me talk about this before: Most of the companies that are in our direct AI basket make a lot of money. When we look at the free cash flow to revenue ratios of these stocks, they're substantially above the average for the market. And really, Oracle and Intel are the only two laggards at the end that don't.
But the hyperscalers are betting the ranch and these numbers are sometimes even hard to believe in until you triple-check them. But the average S&P company spends about 10% of its revenues on capex and R&D. But okay, that includes a lot of companies that aren't necessarily very capex intensive. Let's just look at the information technology space. That number's about 17%. So the average overall tech company is about 17% capex in R&D, well, as a share of revenues. The four hyperscalers are bunched up between 35 and 40%. And then the Meta number is just nuts, at almost 70% at this point. So these guys are definitely betting the ranch on this.
Now so far, this capex boom we've been seeing has been financed mostly from internal cash flow. And that's what makes it different from, you know, casinos and airlines and gas pipelines and things from prior decades.
And the broadband boom at the end of the 1990s, most prior capex cycles in the U.S were financed by debt, at least by the end of it: this one, not so much. We have a chart in here that looks at the share of capex in dividends financed with debt, rather than cash flow.
And even though the capex to sales numbers are going up sharply right now, the amount that's being financed with debt is not. And that's in very stark contrast to the end of the 1990s.
Now, that said, in the fourth quarter of 2025, all of a sudden, there was an explosion of hyperscaler debt financing. And by that I mean bonds, loans and leases to finance data centers and other AI-related expenditures. And Oracle, Google, Meta and Amazon were the big borrowers. Even so, you know, we started writing about Oracle in September, but if you strip out Oracle, most of the rest of these direct AI stocks that we're looking at here have net debt to EBITDA ratios that are either negative because they've got more cash and equivalents than debt, or they have, like Amazon and Microsoft, very, very low positive debt ratios that are way below the S&P median. So Meta and Oracle are kind of pushing the needle right now on debt financing of all this. But in some ways at least so far, they're outliers. And by the way, Oracle is paying the price and Oracle is showing the market what happens when you push too hard to finance this stuff. Their credit default swap spreads have been widening pretty sharply, about 100 basis points just over the last couple of months in 2025.
Now, when most people, or at least when a lot of people talk about this current situation we're in, people talk about valuation and they believe that we're in a valuation bubble. I'm not so sure.
The valuations are less extreme than you might think. And the way that we've been looking at that is to say, well, let's just not look at PE, right. Let's adjust PE for margins and profitability and earnings growth. And one way to do that is—we have a chart in here that looks at price-to-book ratios relative to return on earnings.
And if you look at the sectors that way, all of a sudden there's an inherent sensibility to the way the market's being priced. So you may believe that investors are paying too much for growth right now. And maybe they are. But there's a lot more internal coherence to the way the markets are being priced than a lot of things I read.
And even when we add the individual stocks of those hyperscalers and the four big semiconductor companies, they lie along the same curve. In other words, there's a linear relationship between return on earnings and margins, and what kind of valuation the markets are putting on. Another way to think about it is when you can look at a PEG ratio, which is a PE ratio adjusted for expected earnings growth.
And when you look at the PEG ratio for the tech sector globally, the PEG ratio is pretty much the same as the overall market. So investors are paying up for growth. But not in any kind of 2000-ish type way, at least so far, as far as we can tell. Now that said, if you look at the big economy-wide GDP accounts, private investment, private fixed investment and information processing equipment and software, as a share of GDP has now reached a new all-time high just past the 2000 peak.
So there's just another way of looking at this, a ton of money pouring into this space right now. So, you know, the lesson I've learned as an investor is when the markets are bombed out or after a sell-off, and when sentiment is negative, you're supposed to ask, what could go right, rather than obsessing over the things that led to the sell-off in the first place.
But when you're at all-time highs and there's a lot of enthusiasm priced into markets, you're supposed to ask what could go wrong. And so that's what we do in most of the 2026 Outlook, we ask what could go wrong. And we focus specifically on four medium-term risks.
Number one, a Metaverse moment for the hyperscalers. I'll explain what that means in a couple of minutes. But basically, you know, we had a Metaverse moment a couple of years ago, and the MAG7 stocks fell by 50% in 2022 simply because people lost confidence in the earnings projections.
The second thing is a power generation constraint.
The third risk is, China somehow scales the moat on its own with its own lithography and semiconductor technology. I think that's a when, not an if.
And then lastly, as Chinese dependance on Taiwan starts to go down, you have to start thinking more about risks to Taiwan. And TSMC and Western access to those chips. And then we have a section at the end that looks at all of the other stuff going on in terms of the fed and tariffs, labor supply, U.S. equities, China, Japan and things like that. So, we also have a section at the end that looks at the history of populism for investors.
So I'm not going to go through too much of this. But let me just talk about a few things. The first risk, and I think the most important one to think about, is this Metaverse moment. And as I mentioned, you know, the MAG7 stocks fell by 50% in 2022 simply because of a lack of confidence in their ability to sustain the level of earnings growth that they had been posting.
So what we do is we bucket the whole tsunami of information about AI into six things. We looked at the improved technological capabilities of these models. We looked at adoption rates of these models. We looked at the impact of these models on corporate profits and employment. We looked at hyperscaler revenues, free cash flow margins and cash balances.
And then we looked at the question around GPU and networking depreciation assumptions. Towards the end of the year, there were some well-known short sellers that started to focus on that, I think for good reasons. And so we take a deep dive into that. But I stacked up these Metaverse moment topics in order of importance. In other words, the stuff we end with in this section is more important than the stuff in the beginning.
So, yes, we all know about the improved technological capabilities of these models, but their impact on actual profits, revenues and labor productivity is really the more important thing for us. So, I'm not going to go through too much of the details. It's all in there. I will say that to me, the more robust the survey that we've seen—and surveys are just surveys, right? —but the more robust and detailed the survey, the less optimistic it is about the actual cost and revenue impact on companies adopting generative AI. And so I think it's important to keep that in mind.
Now there are some real labor productivity benefits happening. There's a few individual companies that have reported, you know, Anthropic themselves and Autodesk and Adobe and Deloitte and Tencent. I mean, of course, there are a lot of examples. But on a broad, marketwide basis, so far, this whole AI trade has been about the infrastructure of it rather than the benefits of it.
And what do we mean by that? A lot of the sell-side Wall Street firms have created three buckets of stocks, and we show ones here from one of them, but they're all pretty similar. And there's three buckets: AI infrastructure, which are the semiconductors, the electrical equipment, tech hardware, power suppliers. And then that's one bucket.
Another bucket is, companies that are supposed to benefit from selling products and services related to all this. And the third bucket are the companies that are supposed to benefit from productivity gains because they have a high labor share of sales. So far, only the first bucket is really doing well. The other two buckets are kind of flat to the broad market since GPT was launched.
And so far, other than the in the infrastructure trade, the rest of it so far has been a stockpicker’s game rather than a secular one. So what we're going to take a look at is, how sustainable is this? Well, the hyperscaler free cash flow margins are starting to trend down, particularly for Meta, which is outspending everybody else.
And their cash balances are going down, and that's important too. And the chart in here is kind of remarkable. Hyperscaler cash, and cash equivalents, were 40 to 50% of total assets a few years ago. And now they're converging to 10, let's call it 10 to 20%. So the glide path that we're on is these companies are getting closer and closer to where they're either going to have to start borrowing a lot of money, or they're going to have to start making a lot more money on their generative AI investments than they have so far.
And then we also spend some time looking specifically at OpenAI and the, to me, the best way to think about it is the good, the bad and the ugly, because there's a little bit of everything. The good is the projections the company's making about its future are, the stuff is flying off the shelves. And what I mean by that is, in the middle of 2025, they made a bunch of forecasts for 2027. Revenues, users, paid subscribers, tokens, you know, all of the widgets that are involved in this whole thing.
And then just three months later, in October of last year, they redid the forecast for 2027. And they had gone up dramatically in just three months. So the good news is, from OpenAI's perspective, its business prospects are booming. The bad is, for them to reach their long-term targets, they need 30GW of power by 2030, and I'll get into that more in a minute.
And the ugly is, you know, OpenAI is arguably the biggest individual risk to the moat, even more than Nvidia, and despite the fact that OpenAI is still a private company. OpenAI is on track to make about $10 to $20 billion in revenue, and they have commitments of $1.4 trillion to its corporate partners. And they currently survive almost exclusively on subscription fees and developer AI fees, with little or no search advertising revenue, cloud computing revenue or hardware sales. And, you know, Altman's response to Brad Gerstner on his podcast, when Gerstner kept asking him questions about this, and Altman replies, look, if you want to sell your shares, I'll find you a buyer. You know, I think this is going to be a very interesting place to watch over the next 12 to 18 months.
All right. Risk number two is data centers and energy. And you've all seen some of these charts from us from before.
Office building investment is collapsing for obvious reasons. Data center spending, electric power going up a lot. The question is, are we getting closer to a wall of power constraints? And as many of our clients know, I do probably six months of work a year on energy topics. And every March I publish our annual energy paper. And so they'll be a big data center section in there.
But, you know, you have to start asking yourself the question of whether or not some of these targets that these hyperscalers are throwing out there can be met. And I have found in my own personal experience, historically, some of the people at hyperscalers in charge of energy policy have not always been very realistic about the future.
A few years ago, Google had this kind of net-zero forecast thing that they shelved after a few years once they realized that it was infeasible.
There's a chart in here that I think is the most important one, which is, how much capacity the United States is adding each year. Right. Because that'll tell us whether or not 30GW for open AI is a reasonable number or not. But I have two lines on this chart. One is the nameplate capacity that gets added. And the other one is the nameplate capacity weighted by the effective load-carrying capability of all this new capacity.
In other words, adjusting nameplate capacity for its reliability and intermittency. Right? I mean, historically, from 1950 to around 2010, you didn't have to worry about that. But when the renewables revolution started, a lot of the capacity you're getting, whether it's hydro or wind or solar or even batteries, have certain intermittency and reliability issues. So once we adjust for that, the U.S. didn't add 65GW last year. The U.S. added 25GW last year. It's a very different picture. So a 30GW-build for OpenAI looks like a lot of money.
The other thing, too, is there's a data center backlash, and there's another important chart in here to help you understand why. We have a chart that shows that, for 10 or 12 utilities, how much are they charging their regular customers for electricity. And even with the extra amount that they charge the data centers as an extra fee for all of the additional generation, transmission, distribution and complexity that they bring to the table, almost none of these utilities are, even with those extra tariffs, charging enough to cover the cost of new generation. And so that's the reason why there's a backlash against the data centers within the power generation community. Because the power generation community and the ISOs, which are the independent system operators, want these guys to pay full freight for the cost related to what they're doing to the grid.
I go into more detail in the piece. There are some strenuous objections to what I just said. In papers, either academic papers or financed by Google, that argue that there's plenty of spare and slack natural gas capacity, etc., etc., the paper gets into more detail on that. But the bottom line is that there's just not enough production capacity of combined cycle turbines to meet demand.
And it is what it is. And over the last 24 months, the cost of a new combined cycle turbine has risen from about 1200 kilowatt dollars a kilowatt to 2500 delivery times for 3 to 7 years. And the same thing is now impacting the simpler single cycle turbines, whose costs have also doubled and whose delivery times have also gone up a lot.
So, I feel like we're getting closer to some kind of power constraint wall. I spoke at a conference last year, people from Brookfield disagree with that. You know, time will tell. One thing that's important to remember is on the tariff question, while the moat companies, which are the semiconductors, computer parts and peripherals, are mostly exempt from tariffs, right. They're—all of those products are 70 to 80% exempt from tariffs. That's not the case with the power generation ecosystem. So the moat companies did a better job lobbying the White House and the Commerce Department to exempt a lot of their products from tariffs. The power generation industry had no such luck. When we're talking about turbines, generators, transformers, electrical switches, solar panels, batteries, electric boards and things like that, motherboards, those things are subject to much fewer exemptions and therefore much higher terms.
And then just to wrap up on the OpenAI question, 30GW of power over five years is roughly equal to the peak capacity added during the nuclear boom in the United States in the 1970s. And then the one that took place in the 1980s. So just for context, that's a lot of power. Now, one country that is not having trouble adding more generation is China.
And so China is gearing up to compete with the U.S. on both brain and brawn. And what I mean by that is there's a lot of innovation going on in China. But the part, the gap they can't make up versus Nvidia based on pure innovation, they're going to out-broaden them by simply stringing more chips together, making bigger clusters.
Normally you wouldn't do that in the West because of the cost of power. But in China they're less sensitive about the cost of power, and they're more focused on national security and building out domestic supply chains and reducing exposure on Taiwan and the U.S. in the process. So the third risk we talk about in the piece is, what if China scales this moat eventually on its own?
Now, to be clear, China currently relies really heavily on the West. And when you look at, the origin of the chips that are used to train Chinese models, the vast majority of those are Western chips. Only a handful of them rely on Chinese chips. And when you look at installed semiconductor capacity, around the world, China has plenty—but the older chips that are 15 nanometers and above, that are going to be used in automobiles, refrigerators and simple mechanical devices.
When you start talking about advanced chips and specifically those less than 40 nanometers, China doesn't really have much production capacity at all. That's what they're now aiming to change. And we go into a fair bit of technical detail in the piece because I thought it was interesting, and I think it's important to understand if you're really trying to get at this China question. The bottom line is that the current generation of Huawei's chips are about a little more than two times the power draw per unit of computation when compared to Nvidia's current offerings. And when you look in the future, those ratios go even higher.
And so what's on the table is that Nvidia's power demands are expected to drop pretty substantially per unit of computation relative to Huawei. And so what Huawei is proposing instead is, let's string together 8000 or 15,000 individual processing units compared to the same cluster from Nvidia that would be either 140 or around 600. And so, that kind of daisy chain approach from China is very inefficient, but they're aiming to make up for it with massive subsidies, and a lot more investment in power generation and distribution.
So the bottom line is, China has a long way to scale the moat if you're using a Western lens because they're 1 to 3 generations behind on almost everything. But they have a different approach than the West and its Asian allies. As I mentioned, massive industry subsidies are willing to absorb higher costs in the interest of national security, a no-holds-barred build-out of all forms of power generation. And also no small amount of industrial espionage, whether it's against TSMC or ASML or Nvidia or everybody else. And what it feels like to me is that China is gradually reducing its technological reliance on Taiwan, and TSMC specifically, just as it continues to ramp up its military hardware focused on Taiwan. And so the last of the four risks that we all need to think about is, Taiwan may be the most blockade-sensitive, advanced economy in the entire world.
And that is by no means an exaggeration and might actually be an understatement, because look at this chart we have in here. Taiwan imports 90% of its primary energy consumption. I mean, there's only a couple of countries in the world like Singapore and Cyprus and Morocco that are at similar levels. So 90% of fossil fuels, imported.
It represents 90% of their entire primary energy consumption. And by the way, it's not just energy, it's also food. So Taiwan is in the top ten of food imports as a percentage of the domestic food supply. And the thing that's interesting about this list of countries, it's the only country not in the Middle East that has this kind of food import sensitivity.
Right? And you can understand why Qatar and Kuwait and Saudi Arabia and Oman and the UAE import a lot of food based on the topography of where they live. But Taiwan has plenty of arable land and yet still is in the top ten list of countries that import a lot of the domestic food supply.
And so, I'm just going to close this out by talking about and showing this last chart. At the end of last year, China had somewhere between 60 and 100% of all of its military assets deployed in the Taiwan Strait. Whether we're talking about surface vessels, submarines, air force, special mission aircraft, specifically destroyers, frigates, you know, amphibious assault ships—100% of the amphibious assault ships that China has are in the Taiwan Strait. So, you know, I don't think anybody has a crystal ball here, but when we start thinking about risks, we have to start thinking about this.
And then what is the U.S. trying to do? Well, Secretary of Commerce Lutnick has talked about the U.S. becoming 40% self-sufficient in semiconductor production by the end of the decade. It's ambitious. If you look at the roadmaps, they're a little imprecise. I wish we had a little bit more hard numbers from some of the companies involved, but we drew out the roadmap of what TSMC is pointing to building in Arizona and currently building. We look at existing Intel production and then future production, potentially in both Arizona and Oregon. And then we look at estimated future production from Samsung for Qualcomm and Tesla in Texas. And if you add all of that up and you make projections of U.S. semiconductor advanced node demand, by the end of the decade, the U.S. could get to 30 to 35%. But it's a stretch ambition. And that would mean the U.S. is still highly reliant on TSMC and Taiwan. And note that even today, all the chips made in Arizona are still sent back to Taiwan for packaging, dicing and testing. At least until Amcor builds this new facility they're working on, which is targeted for 2028. So we're in an interesting vortex here, where it looks like it'll be the end of the decade at the earliest, when the U.S. starts to achieve something that looks like partial self-sufficiency in advanced semiconductors.
So there's a ton more information in the Outlook. Take a look, at those four sections if you're interested in getting into the details. I have a few pages at the end of the Outlook on everything else. You know, the Fed, tariffs, immigration, the U.S. capital markets recovery, healthcare stocks, China, Japan, as I mentioned, a page on the history of populism for investors…And then for those of you that remember Robert McNamara and John Mitchell, we have some comments on them.
Let me just mention a couple things, and, let me close this out and thank you all for listening. Let me close this out with a discussion of the U.S. versus the rest of the world in equities.
So the U.S lagged the rest of the world, or almost all of it other than India in 2025. So, and, if you look at U.S. large cap, U.S. small cap, it trailed almost every country. So I can understand somebody saying, wow, you know, that was an opportunity missed. This year was a great year to have been diversified.
The challenge has been, the average investment strategist, the average sell-side Wall Street strategist has been saying that for the last 11 years. And over the last 11 years, the PE discount for non-U.S. stocks, which started at parity in 2009, kept going down and down and down and down. And it turns out that the beginning of last year was rock bottom, when non-U.S. stocks traded at a 40% PE discount versus the U.S.
And so, last year some of that got caught up. But I thought the right framework for thinking about this. Look, if you switched from the U.S. to non-U.S. stocks at the beginning of last year, you're a genius, right? Because that was perfect timing. But look at this last chart I want to show you. Suppose you had switched at any point since 2010? You would have vastly underperformed, never switching at all.
So the first bar on this chart shows that you would have made almost 900% on your money cumulatively, by investing in the S&P 500 since January 2009. If you had swapped out of the S&P into non-U.S. stocks, which is this MSCI World ex USA Index in all the years in between, you would have substantially underperformed. So you would’ve needed to have avoided the siren song calls—and by the way, I'm really looking forward to the new Christopher Nolan Odyssey movie, I saw the trailer—you would have needed to avoid all the siren song calls to go into international stocks because they were cheaper every year for the last 11 years and just hit the bid at the beginning of last year in order to have navigated this properly.
So, for those of you that missed it last year, but were invested most in the S&P, you're generally way ahead of everybody else. And I'm not sure how much room there is for the non-U.S. stocks to continue to appreciate. They're currently trading at about a 30% discount to the U.S. I think 20% is about as good as it gets because of how much more profitable and higher earnings growth that U.S. sectors tend to generate compared to Europe, Japan and China.
So that is the Smothering Heights podcast. Thank you very much for listening. And take a look at the piece, which is in your inbox this morning, and I look forward to talking to you again soon. And so long from incense-ridden Santa Fe.
(DESCRIPTION)
Text: JP Morgan, Eye on the Market.
(SPEECH)
[WHOOSHING]
(DESCRIPTION)
Painting, a red dragon stands on top of a stone castle wall and breathes fire. The stones of the castle read Nvidia, AMD, ASML, and TSMC. At the top of the wall are flags that have logos for Microsoft, Google, Meta, and Amazon on them. Inside the fire that the dragon breathes are TSMC chips. Text: Outlook 2026, Smothering Heights. The speaker, Michael Cembalest, sits in front of a background that shows an office with views of a city.
(SPEECH)
Happy New Year, everybody. Welcome to the 2026 Eye on the Market outlook podcast. Recording this in the last week of December from Santa Fe there, New Mexico. There are no fish at all here, but there's lots of quilts and incense everywhere, so that's different. So I wanted to think about different topics for this year, and I decided to take a deep dive into this question of whether this giant moat that's supporting the global equity markets is really as indefensible as it's sometimes described to be.
And by that, I'm referring to the giant moat of the big hyperscalers and the semiconductor companies that support them. And that is the 80% of the outlook, and then we have a few pages on everything else. So in this podcast, I'm going to walk quickly through the outline of what we look at in this year, and then some discussion about US equity outperformance versus the rest of the world. So let's get into it.
So is this moat really indestructible? The cover art this year is Game of Thrones-inspired vaguely, loosely. And it's a dragon and he's sitting on top of a castle, and the flags of the four hyperscalers are there, and the bricks of this moat are made up of NVIDIA, TSMC, AMD, and ASML, which is the Dutch lithography company, and the dragon's spitting a fiery stream of TSMC's 2 nanometer chips down on the people at the bottom of the moat.
(DESCRIPTION)
A chart is entitled 8 largest Hyperscaler and Semi scaler market caps, US dollars, trillions. The chart shows o to 20 trillion dollars on the y axis and years from 2018 to 2025 on the x axis. Gold and blue lines show that these companies' values have risen dramatically, as Michael Cembalest explains.
(SPEECH)
So that's the sense you get when you read about this moat, that it's this giant, indestructible thing. And I understand why people see it that way, and I partially see it that way. We're talking about just eight stocks here that were worth $3 trillion in 2018 and are now worth $18 trillion. So they've gone up by a factor of 6 in just a few years. And so here are some facts and figures about this moat.
The four hyperscalers have alone spent 1.3 trillion in CapEx and R&D since the fourth quarter of 2022, when ChatGPT was launched. If we broaden the moat to the eight companies, which are the four hyperscalers plus NVIDIA, TSMC, AMD, and ASML, they represent 20% of global equity market cap, which is kind of incredible from eight stocks. Then if we broaden it more to 42 AI related companies in the US equity markets, they account for 65% to 75% of S&P earnings, revenues and capital spending since Q4 2022.
And then in the GDP accounts-- and this is based on the numbers that just came out at the end of December-- tech capital spending accounted for 40% to 45% of US GDP growth in the first three quarters of 2025, not GDP. GDP growth. But still, those are some amazing numbers.
(DESCRIPTION)
A bar chart is entitled Capex in context, Tech capital spending in 2025 versus spending on major U S infrastructure projects, Peak annual project percent of GDP. The y axis shows percentages from 0 to 2.4%. The x axis shows different infrastructure projects from the 1930s public works, to the Manhattan Project, Electricity, Apollo Project, Interstate Highway, Broadband, and Tech capex in 2025. The first four categories combined are only a little more than the tech capex in 2025.
(SPEECH)
And just to put the tech capital spending numbers in context, there's a chart that we included. So tech capital spending was in the neighborhood of 2% of GDP in 2025.
That's the sum of the peak year spending on the interstate highway system that was started by Eisenhower, the Apollo project, the moon landing, electrification of farming, the Manhattan Project, and several of the FDR public works projects from the 1930s, including the Triborough Bridge, the Midtown Tunnel, and LaGuardia Airport. So 2025 tech capital spending was equal to the sum of all those things in real-- relative to GDP at the time-- dollars. So that's kind of amazing.
And the stakes are high that there's got to be some really substantial productivity and profits outcome from all this spending.
(DESCRIPTION)
A bar chart is entitled Most direct AI companies are highly profitable. Free cash flow to revenue ratios of Direct Al stocks. The chart shows that most tech companies have plenty of free cash flow.
(SPEECH)
So here's the good news, and you've heard me talk about this before. Most of the companies that are in our direct AI basket make a lot of money. When we look at the free cash flow to revenue ratios of these stocks, they're substantially above the average for the market. And really, Oracle and Intel are the only two laggards at the end that don't.
(DESCRIPTION)
A line chart is entitled, But the hyper scalers are betting the ranch. Hyperscaler capex and R&D as a share of revenues. The y axis shows percentages from 0 to 70. The x axis shows years from 2017 to 2025. Colorful lines show Meta, Alphabet, Amazon and Microsoft's expenditures, which are all way over the S&P 500.
(SPEECH)
Now, but the hyperscalers are betting the ranch. And these numbers are sometimes even hard to believe until you triple check them. But the average S&P company spends about 10% of its revenues on CapEx and R&D, but OK, that includes a lot of companies that aren't necessarily very CapEx intensive. Let's just look at the information technology space. That number is about 17%. So the average overall tech company is about 17% CapEx and R&D as a share of revenues.
The four hyperscalers are bunched up between 35% and 40%, and then the Meta number is just nuts. It's almost 70% at this point. So these guys are definitely betting the ranch on this.
(DESCRIPTION)
A line chart with the title: financed mostly from internal cash flow Capex financing versus capex cycle. The y axis has percentages from 0 to 45. The x axis has years from 1996 to 2026. He explains the results.
(SPEECH)
Now, so far, this CapEx boom we've been seeing has been financed mostly from internal cash flow. And that's what makes it different from casinos, and airlines, and gas pipelines, and things from prior decades, and the broadband boom at the end of the 1990s.
Most prior CapEx cycles in the US were financed by debt, at least by the end of it. This one, not so much. We have a chart in here that looks at the share of CapEx and dividends financed with debt, rather than cash flow. And even though the CapEx to sales numbers are going up sharply there right now, the amount that's being financed with debt is not. And that's in very stark contrast to the end of the 1990s.
(DESCRIPTION)
A bar chart with the title: ...until Q4 2025, Annual change in hyper scaler long term debt, (bonds, loans and leases). US dollars, billions. The y axis is dollar amounts from 0 to 200. The x axis is years from 2015 to 2025. The money spent has gone way up, including in just the fourth quarter of 2025.
(SPEECH)
Now, that said, in the fourth quarter of 2025, all of a sudden, there was an explosion of hyperscaler debt financing. And by that, I mean bonds, loans and leases to finance data centers and other AI related expenditures. And Oracle, Google, Meta, and Amazon were the big borrowers.
(DESCRIPTION)
A bar chart is entitled: which moved the needle for Oracle and Meta. Net debt to EBITDA ratios of Direct AI stocks. Multiple: includes bonds, loans and SPV triple net leases. The y axis goes from 0 times to 7 times. The chart shows that the numbers have gone up, especially in the third and fourth quarter of 2025.
(SPEECH)
Even so, we started writing about Oracle in September. But if you strip out Oracle, most of the rest of these direct AI stocks that we're looking at here have net debt to EBITDA ratios that are either negative because they've got more cash and equivalents than debt, or they have, like Amazon and Microsoft, very, very low positive debt ratios that are way below the S&P median.
So Meta and Oracle are kind of pushing the needle right now on debt financing of all this. But in some ways, at least so far, they're outliers.
(DESCRIPTION)
A line chart. Text: causing Oracle to pay the price for higher debt and lower FCF than the rest of the direct AI stocks (as we anticipated in September 2025). Credit default swap comparison, select AI stocks. The y axis has numbers from 20 to 160. The x axis shows years from 2022 to 2026. The lines represent companies. Oracle has a very high rate.
(SPEECH)
And by the way, Oracle is paying the price. And Oracle is showing the market what happens when you push too hard to finance this stuff. Their credit default swap spreads have been widening pretty sharply, about 100 basis points just over the last couple of months in 2025.
(DESCRIPTION)
Two scatterplot charts. Text: Valuations are less extreme than you might think. The first chart has the title S&P 500 price to book and R O E. The second is S&P 500 versus hyper scalers and semi scalers. He explains.
(SPEECH)
Now, when most people, or at least when a lot of people talk about this current situation we're in, people talk about valuation, and they believe that we're in a valuation bubble. I'm not so sure. The valuations are less extreme than you might think. And the way that we've been looking at that is to say, well, let's just not look at PEs, right? Let's adjust PE for margins and profitability and earnings growth.
And one way to do that is to look at-- we have a chart in here that looks at price to book ratios relative to return on earnings. And if you look at the sectors that way, all of a sudden, there's an inherent sensibility to the way the market's being priced. So you may believe that investors are paying too much for growth right now, and maybe they are. But there's a lot more internal coherence to the way the markets are being priced than a lot of things I read.
And even when we add the individual stocks of those hyperscalers and the four big semiconductor companies, they lie along the same curve. In other words, there's a linear relationship between return on earnings and margins and what kind of valuation the markets are putting on another way.
(DESCRIPTION)
A line chart with the title: Adjusted for earnings growth, tech stocks are priced in-line with the rest of the market, PEG ratios for global tech stocks versus the market. The y axis has numbers from 0.5 to 2.5. The x axis shows years from 1995 to 2025. A gold line shows MSCI all world equity index for info tech, and a blue line shows all world equity index. The lines pretty much track with each other as they go up and down.
(SPEECH)
Another way to think about it is when you can look at a peg ratio, which is a PE ratio adjusted for expected earnings growth.
And when you look at the peg ratio for the tech sector globally, the peg ratio is pretty much the same as the overall market. So investors are paying up for growth, but not in any kind of 2000-ish type way, at least so far, as far as we can tell.
(DESCRIPTION)
A line chart with the text: But we have reached a new peak, so. Private fixed investment in information processing equipment & software as a share of potential GDP, percent. The y axis has percentages from 3.2 to 4.8. The x axis has years from 1995 to 2025. The line goes up, sharply down in 2003, up a little, down in 2008, then steadily up.
(SPEECH)
Now, that said, if you look at the big economy-wide GDP accounts, private investment, private fixed investment and information processing equipment and software as a share of GDP has now reached a new all time high just past the 2000 peak.
So there's just another way of looking at this, a ton of money pouring into this space right now. So the lessons I've learned as an investor is, when the markets are bombed out or after a sell off, and when sentiment is negative, you're supposed to ask, what could go right, rather than obsessing over the things that led to the sell off in the first place. But when you're at all time highs and there's a lot of enthusiasm priced into markets, you're supposed to ask what could go wrong. And so that's what we do in most of the 2026 outlook. We ask what could go wrong.
And we focused specifically on four medium term risks. Number one, a metaverse moment for the hyperscalers. I'll explain what that means in a couple of minutes. But basically, we had a metaverse moment a couple of years ago, and the Mag Seven stocks fell by 50% in 2022, simply because people lost confidence in the earnings projections. The second thing is a power generation constraint. The third thing is that China-- the third risk is China somehow scales the moat on its own with its own lithography and semiconductor technology. I think that's "when," not an "if."
And then lastly, as Chinese dependence on Taiwan starts to go down, you have to start thinking more about risks to Taiwan, and TSMC and Western access to those chips. And then we have a section at the end that looks at all of the other stuff going on in terms of the Fed, and tariffs, labor supply, US equities, China, Japan, and things like that. So we also have a section at the end that looks at the history of populism for investors.
So I'm not going to go through too much of this,
(DESCRIPTION)
A line chart with the title 1. A metaverse moment, Ancient history: Mag 7 stocks fell by 50% in 2022. Total return index. The y axis shows numbers from 5000 to 40,000. The x axis shows January in each year from 2021 to 2025. The line goes mostly up, with some dips.
(SPEECH)
but let me just talk about a few things. The first risk, and I think the most important one to think about, is this metaverse moment. And as I mentioned, the Mag Seven stocks fell by 50% in 2022, simply because of a lack of confidence in their ability to sustain the level of earnings growth that they had been boasting.
So what we do is we bucket the whole tsunami of information about AI into six things. We look at the improved technological capabilities of these models. We look at adoption rates of these models. We look at the impact of these models on corporate profits and employment. We look at hyperscaler revenues, free cash flow margins and cash balances. And then we look at the question around GPU and networking depreciation assumptions.
Towards the end of the year, there were some well-known short sellers that started to focus on that, I think for good reasons. And so we take a deep dive into that. But I stacked up these metaverse moment topics in order of importance. In other words, the stuff we end with in this section is more important than the stuff in the beginning. So yes, we all know about the improved technological capabilities of these models, but their impact on actual profits, revenues and labor productivity is really the more important thing for us.
(DESCRIPTION)
A chart with the title: The survey says. Cost decreases and revenue includes from generative Al use. The chart has two halves, cost decreases on the left and revenue increases on the right. He explains the results.
(SPEECH)
So I'm not going to go through too much of the details. It's all in there. I will say, though, to me, the more robust the survey that we've seen-- and surveys are just surveys. But the more robust and detailed the survey, the less optimistic it is about the actual cost and revenue impact on companies adopting generative AI. And so I think it's important to keep that in mind.
(DESCRIPTION)
A bar chart with the title Some real productivity benefits happening, Company estimates of the effect of AI on labor productivity. The y axis has percentages from 0 to 90. Each bar shows a different company's percent.
(SPEECH)
Now, there are some real labor productivity benefits happening. There's a few individual companies that have reported-- Anthropic themselves, and Autodesk, and Adobe, and Deloitte, and Tencent. There are, of course, there are a lot of examples. But on a broad market-wide basis, so far, this whole AI trade has been about the infrastructure of it rather than the benefits of it. And what do I mean by that?
(DESCRIPTION)
A line chart called, but not at market-wide levels So far, AI infrastructure company benefits only. The y axis shows numbers from 80 to 170. The x axis shows June and December for each year from 2023 to 2025. Lines show different types of AI usage, from AI infrastructure companies to AI enabled revenues to regular companies.
(SPEECH)
A lot of the sell side Wall Street firms have created three buckets of stocks, and we show ones here from one of them, but they're all pretty similar. And there's three buckets, AI infrastructure, which are the semiconductors, the electrical equipment, tech hardware, power suppliers. And then that's one bucket. Another bucket is companies that are supposed to benefit from selling products and services related to all this. And the third bucket are the companies that are supposed to benefit from productivity gains because they have a high labor share of sales.
So far, only the first bucket is really doing well. The other two buckets are kind of flat to the broad market since GPT was launched. And so, so far, this has been a stock picker's-- other than in the infrastructure trade, the rest of it so far has been a stock picker's game, rather than a secular one.
(DESCRIPTION)
A line chart. Text: Tracking the hyper scalers via changes in free cash flow, Hyper scaler free cash flow margins, (net of capex), Percent trailing 12 months. The y axis shows percents from negative 10 to positive 40. The x axis shows years from 2021 to 2026. The lines show the four hyperscaler companies and the S&P 500. Only Amazon is below the S&P 500.
(SPEECH)
So what we're going to take a look at is, how sustainable is this? Well, the hyperscaler free cash flow margins are starting to trend down, particularly for Meta, which is outspending everybody else.
(DESCRIPTION)
Another line chart with the same companies. The title is, and by declining cash balances. Hyperscaler cash and cash equivalents, Share of total assets. The y axis shows percentages from 10 to 55. The x axis shows years from 2019 to 2026. All the lines go down.
(SPEECH)
And their cash balances are going down, and that's important, too. And the chart in here is kind of remarkable. Hyperscale or cash and cash equivalents were 40% to 50% of total assets a few years ago, and now they're converging to, let's call it, 10% to 20%. So the glide path that we're on is these companies are getting closer and closer to where they're either going to have to start borrowing a lot of money, or they're going to have to start making a lot more money on their generative AI investments than they have so far.
(DESCRIPTION)
A table is entitled, Open AI, the good. It shows Open AI forecasts for 2027, as made in mid 2025 and October 31, 2025.
(SPEECH)
And then we also spend some time looking specifically at OpenAI. And to me, the best way to think about it is the good, the bad, and the ugly, because there's a little bit of everything. The good is the projections the company is making about its future are-- the stuff is flying off the shelves. And what I mean by that is, in the middle of 2025, they made a bunch of forecasts for 2027, revenues, users, paid subscribers, tokens, all of the widgets that are involved in this whole thing.
And then, just three months later, in October of last year, they redid the forecast for 2027, and they had gone up dramatically in just three months. So the good news is, from OpenAI's perspective, its business prospects are booming. The bad is, for them to reach their long term targets, they need 30 gigawatts of power by 2030, and I'll get into that more in a minute. And the ugly is, OpenAI is arguably the biggest individual risk to the moat, even more than NVIDIA, and despite the fact that OpenAI is still a private company,
OpenAI is on track to make about $10 to $20 billion in revenue, and they have commitments of $1.4 trillion to its corporate partners. And they currently survive almost exclusively on subscription fees and developer AI fees, with little or no search advertising revenue, cloud computing revenue, or hardware sales. And Altman's response to Brad Gerstner on his podcast when Gerstner kept asking him questions about this, and Altman replies, look, if you want to sell your shares, I'll find you a buyer. I think this is going to be a very interesting place to watch over the next 12 to 18 months.
All right.
(DESCRIPTION)
Text: Data Centers and Energy. Two line charts. The first one is U S live data center capacity. The line goes up. The second is U S construction spending. The lines for data centers and electricity go up, while office buildings comes down.
(SPEECH)
Risk number two is data centers and energy. And you've all seen some of these charts from us from before. Office building investment is collapsing for obvious reasons. Data center spending, electric power, going up a lot. The question is, are we getting closer to a wall of power constraints? And as many of our clients know, I do probably six months of work a year on energy topics, and every March, I publish our annual energy paper. And so there'll be a big data center section in there.
But you have to start asking yourself the question of whether or not some of these targets that these hyperscalers are throwing out there can be met. And I have found in my own personal experience, historically, some of the people at hyperscalers in charge of energy policy have not always been very realistic about the future. A few years ago, Google had this kind of net zero forecast thing that they shelved after a few years once they realized that it was infeasible.
(DESCRIPTION)
A line chart with the title: How much power is the U S actually adding each year? US electricity generation and storage capacity additions: nameplate and E L C C-weighted, Gigawatts, annual. The y axis shows gigawatts from 0 to 70. The x axis shows years from 1950 to 2020. A blue lie shows nameplate capacity. A yellow line shows nameplate capacity weighted to effective load carrying capability factors. They go up and down together, with a steep hike between 2000 and 2010, and then the blue line goes up by itself.
(SPEECH)
So there's a chart in here that I think is the most important one, which is, how much capacity is the United States adding each year? Because that'll tell us whether or not 30 gigawatts for OpenAI is a reasonable number or not. But I have two lines on this chart. One is the nameplate capacity that gets added, and the other one is the nameplate capacity, weighted by the effective load carrying capability of all this new capacity. In other words, adjusting nameplate capacity for its reliability and intermittency.
And historically, from 1950 to around 2010, you didn't have to worry about that. But when the renewables revolution started, a lot of the capacity you're getting, whether it's hydro, or wind, or solar, or even batteries, have certain intermittency and reliability issues. So once we adjust for that, the US didn't add 65 gigawatts last year. The US added 25 gigawatts last year. It's a very different picture. So a 30 gigawatt build for OpenAI looks like a lot of money.
The other thing, too, is there's a data center backlash. And there's another important chart in here to help you understand why.
(DESCRIPTION)
A bar chart with the title Why is there a data center backlash? Specialized data center power rates still trail cost of new generation. The y axis shows different power companies. The x axis shows dollar amounts from 0 to 100. He explains the bars.
(SPEECH)
We have a chart that shows that for 10 or 12 utilities, how much are they charging their regular customers for electricity? And even with the extra amount that they charge the data centers, as an extra fee for all of the additional generation, transmission, distribution, and complexity that they bring to the table, almost none of these utilities are, even with those extra tariffs, charging enough to cover the cost of new generation.
And so that's the reason why there's a backlash against the data centers within the power generation community, because the power generation community and the ISOs, which are the independent system operators, want these guys to pay full freight for the cost related to what they're doing to the grid. I go into more detail in the piece. There are some strenuous objections to what I just said in papers, either academic papers or financed by Google, that argue that there's plenty of spare and slack natural gas capacity, et cetera, et cetera. The paper gets into more detail on that.
(DESCRIPTION)
A bar chart with the title Not enough production capacity, Global gas turbine orders. The y axis shows numbers from 0 to 110. the x axis shows years from 2001 to 2028, with the future years being projections. A dotted line shows our current production limit. Some past years and all the future years are above the production limit.
(SPEECH)
But the bottom line is that there's just not enough production capacity of combined cycle turbines to meet demand. And it is what it is. And over the last 24 months, the cost of a new combined cycle turbine has risen from about $1,200 a kilowatt to 2,500. Delivery times are three to seven years. And the same thing is now impacting the simpler single cycle turbines, whose costs have also doubled, and whose delivery times have also gone up a lot.
So I feel like we're getting closer to some kind of power constraint wall. I spoke at a conference last year. People from Brookfield disagree with that. Time will tell.
(DESCRIPTION)
A bar chart entitled The impact of tariffs, Power generation versus semiconductor tariff exclusions, Percent of 2024 U S product imports excluded from Trump tariffs. The y axis shows different items like computers, transformers, and other equipment. The x axis shows percentages from 0 to 100. The semiconductor ecosystem is in red, and it has the most.
(SPEECH)
One thing that's important to remember is on the tariff question. While the moat companies, which are the semiconductors, computer parts and peripherals, are mostly exempt from tariffs-- all of those products are 70% to 80% exempt from tariffs-- that's not the case with the power generation ecosystem.
So the moat companies did a better job lobbying the White House and the Commerce Department to exempt a lot of their products from tariffs. The power generation industry had no such luck. So when we're talking about turbines, generators, transformers, electrical switches, solar panels, batteries, electric boards, and things like that, motherboards, those things are subject to much fewer exemptions, and therefore much higher tariffs.
(DESCRIPTION)
A line chart with the title, Open AI's 30 GW in context, U S nuclear plants built by year of completion, GW, 5 year rotating sum.. The y axis shows gigwatts from 0 to 40. The x axis shows years from 1960 to 2030. The line is mostly very low, except from 1970 to 1990.
(SPEECH)
And then just to wrap up on the OpenAI question, 30 gigawatts of power over five years is roughly equal to the peak capacity added during the nuclear boom in the United States in the 1970s, and then the one that took place in the 1980s. So just for context, that's a lot of power.
(DESCRIPTION)
Text: 3. China is gearing up to compete on both brain and brawn. A line chart shows electricity generation in terawatts. The y axis shows terawatts, and the x axis shows years from 1990 to 2025. A blue line shows the U S, and itis mostly very flat. A red line shows China, and its line goes upward dramatically.
(SPEECH)
Now, one country that is not having trouble adding more generation is China. And so China is gearing up to compete with the US on both brain and brawn.
And what I mean by that is there's a lot of innovation going on in China. But the gap they can't make up versus NVIDIA based on pure innovation, they're going to out brawn them by simply stringing more chips together, making bigger clusters. Normally, you wouldn't do that in the West because of the cost of power. But in China, they're less sensitive about the cost of power, and they're more focused on national security, and domestic supply, and building out domestic supply chains, and reducing exposure on Taiwan and the US in the process.
(DESCRIPTION)
A bar chart with the title, Status quo: China relies heavily on the West, Origin of chips used to train Chinese LLMs, Number of models. The y axis shows numbers from 0 to 35. The x axis shows years from 2017 to 2025. The bars show what percentage is Western, Chinese, or combination. The bars show that the vast majority of chips are Western.
(SPEECH)
So the third risk we talk about in the piece is, what if China scales this moat eventually on its own? Now, to be clear, China currently relies really heavily on the West. And when you look at the origin of the chips that are used to train Chinese models, the vast majority of those are Western chips. Only a handful of them rely on Chinese chips. And when you look at installed semiconductor capacity around the world,
(DESCRIPTION)
A bar chart with the title, particularly for advanced chips, Installed semiconductor capacity by market and node, Millions of 200 n m equivalent wafer starts per month. The y axis shows numbers from 0 to 12. The x axis shows size of chips, from 7 n m to greater than 130. Different colors show China, the U S and its allies, and other.
(SPEECH)
China has plenty, but the older chips that are 15 nanometers and above that are going to be used in automobiles, refrigerators, and simple mechanical devices.
When you start talking about advanced chips, and specifically those less than 14 nanometers, China doesn't really have much production capacity at all. That's what they're now aiming to change. And we go into to a fair bit of technical detail in the piece, because I thought it was interesting. And I think it's important to understand if you're really trying to get at this China question.
(DESCRIPTION)
A table entitled Chip level. He explains it.
(SPEECH)
The bottom line is that the current generation of Huawei's chips are about a little more than 2 times-- have 2 times the power draw per unit of computation when compared to NVIDIA's current offerings.
And when you look in the future, those ratios go even higher. And so what's on the table is that NVIDIA's power demands are expected to drop pretty substantially per unit of computation relative to Huawei. And so what Huawei is proposing instead is, let's string together 8,000 or 15,000 individual processing units compared to the same cluster from NVIDIA that would be either 140 or around 600. And so that kind of daisy chain approach from China is very inefficient, but they're aiming to make up for it with massive subsidies and a lot more investment in power generation and distribution.
So the bottom line is, China has a long way to scale the moat if you're using a Western lens, because they're one to three generations behind on almost everything. But they have a different approach than the West and its Asian allies. As I mentioned, massive industry subsidies, are willing to absorb higher costs in the interest of national security, a no-holds-barred build out of all forms of power generation, and also no small amount of industrial espionage, whether it's against TSMC, or ASML, or NVIDIA, or everybody else.
And what it feels like to me is that China is gradually reducing its technological reliance on Taiwan and TSMC, specifically, just as it continues to ramp up its military hardware focused on Taiwan.
(DESCRIPTION)
Text: 4. Taiwan may be the most blockade-sensitive advanced economy in the world. A bar chart shows Net imports of fossil fuels as a share of primary energy consumption. The y axis shows percents from 0 to 100. Each bar represents a different country, and Taiwan is third.
(SPEECH)
And so the last of the four risks that we all need to think about is, Taiwan may be the most blockade sensitive, advanced economy in the entire world. And that is by no means an exaggeration and might actually be an understatement.
Because look at this chart we have in here. Taiwan imports 90% of its primary energy consumption. There's only a couple of countries in the world, like Singapore, and Cyprus and Morocco that are at similar levels. So 90% of fossil fuels imported represents 90% of their entire primary energy consumption. And by the way, it's not just energy. It's also food.
(DESCRIPTION)
A bar chart shows value of food imports as a % of domestic food supply. The y axis shows percents from 0 to 100. Each bar represents a country. Taiwan is ninth.
(SPEECH)
So Taiwan is in the top 10 of food imports as a percentage of the domestic food supply. And the thing that's interesting about this list of countries, it's the only country not in the Middle East that has this kind of food import sensitivity. And you can understand why Qatar, and Kuwait, and Saudi Arabia, and Oman, and the UAE import a lot of food based on the topography of where they live. But Taiwan has plenty of arable land, and yet still is in the top 10 list of countries that import a lot of the domestic food supply.
And so I'm just going to close this out by talking about and showing this last chart.
(DESCRIPTION)
A bar chart shows Share of Chinese military assets deployed in the Taiwan Strait, Percent. The y axis shows percents from 0 to 100. The bars represent different types of military assets such as ships, subs, and planes.
(SPEECH)
At the end of last year, China had somewhere between 60% and 100% of all of its military assets deployed in the Taiwan Strait, whether we're talking about surface vessels, submarines, air force, special mission aircraft, specifically, destroyers, frigates, amphibious assault ships. 100% of the amphibious assault ships that China has are in the Taiwan Strait.
So I don't think anybody has a crystal ball here. But when we start thinking about risks, we have to start thinking about this.
(DESCRIPTION)
A bar chart with the title US ambition: 30%-35% advanced node self-sufficiency by the end of the decade. U S might reach 30 to 35% of advanced node production by 2028 to 2031. He describes the chart.
(SPEECH)
And then, what is the US trying to do? Well, the Secretary of Commerce Lutnick has talked about the US becoming 40% self-sufficient in semiconductor production by the end of the decade. It's ambitious. If you look at the roadmaps, they're a little imprecise. I wish we had a little bit more hard numbers from some of the companies involved.
But we drew out the roadmap of what TSMC is pointing to building in Arizona and currently building. We look at existing Intel production, and then future production, potentially, in both Arizona and Oregon. And then we look at estimated future production from Samsung for Qualcomm and Tesla in Texas. And if you add all of that up and you make projections of US semiconductor advanced node demand by the end of the decade, the US could get to 30% to 35%.
But it's a stretch ambition. And that would mean the US is still highly reliant on TSMC and Taiwan. And note that even today, all the chips made in Arizona are still sent back to Taiwan for packaging, dicing and testing, at least until Amcor builds this new facility they're working on, which is targeted for 2028. So we're in an interesting vortex here, where it looks like it'll be the end of the decade at the earliest when the US starts to achieve something that looks like partial self-sufficiency in advanced semiconductors.
So there's a ton more information in the outlook. Take a look at those four sections if you're interested in getting into the details. I have a few pages at the end of the outlook on everything else. The Fed, tariffs, immigration, the US capital markets recovery, health care stocks, China, Japan, as I mentioned, a page on the history of populism for investors. And then for those of you that remember Robert McNamara and John Mitchell, we have some comments on them.
(DESCRIPTION)
A bar chart with the title U S equities lagged the R o W in 2025, YTD total returns for major equity markets/regions, Percent, U S dollars. The y axis shows percents from 0 to 100. Each bar represents a country or region.
(SPEECH)
Let me just mention a couple of things and let me close this out. And thank you all for listening. Let me close this out with a discussion of the US versus the rest of the world in equities. So the US lagged the rest of the world, or almost all of it other than India, in 2025. And if you look at US large cap, US small cap, it trailed almost every country. So I can understand somebody saying, wow, that was an opportunity missed. This year was a great year to have been diversified.
(DESCRIPTION)
A line chart entitled Turns out rock bottom was a 40% P/E discount. Non-US P/E discount versus U S M S C I World ex-U S fwd P/E divided by US had P/E. The y axis shows percentages from 60 to 130. The x axis shows years from 2005 to 2025. The line goes up and down, but the trend is steadily down, and very steeply down.
(SPEECH)
The challenge has been, the average investment strategist, the average sell side Wall Street strategist, has been saying that for the last 11 years. And over the last 11 years, the PE discount for non-US stocks, which started at parity in 2009, kept going down, and down, and down, and down. And it turns out that the beginning of last year was rock bottom, when non-US stocks traded at a 40% PE discount versus the US.
And so last year, some of that got caught up. But I thought the right framework for thinking about this. Look. If you switch to the-- if you switched from the US to non-US stocks at the beginning of last year, you're a genius because that was perfect timing. But look at this last chart I want to show you.
(DESCRIPTION)
A bar chart entitled It made sense to wait. Switching from the S&P to the M S C I World, Total return index January 2009 to December 2025. The y axis shows the percentages and the x axis shows the years. He explains the chart.
(SPEECH)
Suppose you had switched at any point since 2010. You would have vastly underperformed never switching at all.
So the first bar on this chart shows that you would have made almost 900% on your money cumulatively by investing in the S&P 500 since January, 2009. If you had swapped out of the S&P into non-US stocks, which is this MSCI World XUS Index, in all the years in between would have substantially underperformed. So you would need it to have avoided the siren song calls. And by the way, I'm really looking forward to the new Christopher Nolan Odyssey movie. I saw the trailer.
And you would have needed to avoid all the siren song calls to go into international stocks because they were cheaper every year for the last 11 years, and just hit the bid at the beginning of last year in order to have navigated this properly. So for those of you that missed it last year, but were invested most in the S&P, you're generally way ahead of everybody else.
(DESCRIPTION)
Three tables with the title: I think a 20% M S C I World ex-US P/E discount would be as good as it gets
(SPEECH)
And I'm not sure how much room there is for the non-US stocks to continue to appreciate.
They're currently trading at about a 30% discount to the US. I think 20% is about as good as it gets because of how much more profitable and the higher earnings growth that US sectors tend to generate compared to Europe, Japan and China. So that is the Smothering Heights podcast. Thank you very much for listening. And take a look at the piece, which is in your inbox this morning, and I look forward to talking to you again soon. And so long from incense ridden Santa Fe.
[WHOOSHING]
(DESCRIPTION)
Text: JP Morgan, Eye on the Market.
