Roundup
Investor Spotlight
Mohnish Pabrai | Managing Partner at Pabrai Investment Funds | New Money
Called the S&P 500 likely “ridiculously overvalued” and pitched Berkshire (BRK.B) as a safer index alternative, with ~40% cash and Greg Abel ready to deploy in a dislocation.
Put the AI “gold rush” — memory makers, Google, Adobe, even SpaceX — in Buffett’s “too-hard pile,” favoring boring no-brainers like Kazakh super-app Kaspi.
Stressed singular focus, non-negotiable management integrity (”why be slightly in bed with a crook?”), and holding moated compounders like Costco and Visa unless egregiously overpriced.
Aug 2 | 50 min | Digest ⬇️
The Economy and the FED
Michael Kantrowitz | Chief Investment Strategist at Piper Sandler | The Exchange by CNBC
Laid out five reasons the industrial revival has room to run, tying a rising ISM (at 55, seven straight months of expansion) to better credit conditions, easier lending, and broader earnings.
Explained why the Magnificent 7 have gone flat year-to-date — their multiples have compressed as much as their earnings have grown — while the rally broadens to names like Caterpillar, Generac, Target, and Ford.
Argued record valuations are supported by all-time-high margins, and that the real threat to the rally isn’t the Buffett Indicator but high interest rates and the Fed.
Aug 4 | 2:00 - 9:40 (8 min) | Digest ⬇️
Michael Darda | Chief Economist and Market Strategist at Roth | The Exchange by CNBC
Attributed the recent rise in yields largely to a jump in the term premium, signaling markets are jittery about Fed Chair Kevin Warsh’s reaction function.
Called the manufacturing recovery genuinely encouraging — seven consecutive months of the ISM above 50 after a roughly three-year slump.
Dismissed the “C-shaped economy” framing as more entertainment than data, noting real wages rose just 0.26% year over year, and drew a sharp line between one or two modest Fed hikes and 75-to-100-plus basis points.
Aug 4 | 31:35 - 40:20 (9 min) | Digest ⬇️
Cloud and AI Infrastructure
Paddy Srinivasan | CEO at DigitalOcean | The Exchange by CNBC
Detailed the company’s pivot to AI infrastructure: 29% revenue growth, an 800% jump in inference revenue, and AI revenue past $200 million and tripling year over year.
Positioned DigitalOcean as an AWS for the AI-native ecosystem, serving names like Cursor, Character AI, and Hippocratic AI rather than legacy enterprises.
Stressed capital discipline — guiding to 50% growth next year while staying cash-flow positive with a 24% adjusted operating margin and little to no debt.
Aug 4 | 20:30 - 24:00 (4 min) | Digest ⬇️
Gavin Baker | Managing Partner & CIO at Atreides Management | Invest Like the Best
Pressure-tested the AI selloff (July was “2022 in a month”) and found no negative quantitative metric — GPU pricing, DRAM, and token growth all accelerating.
Argues hyperscalers under-earn: contracted compute trades below spot, so as it reprices, operating cash flow funds the buildout — cutting ~$700B of credit.
Sees the real risk as regulatory/PR (NY’s data-center moratorium), with memory LTAs, Nvidia’s “credit wrapper,” and SpaceX’s compute ramp as key battlegrounds.
Aug 4 | 1 hr 18 min | Digest ⬇️
Patrick Boyle | Finance Lecturer & Former Hedge Fund Manager | Patrick Boyle On Finance
Argued the ~$1.65 trillion of “hidden” big-tech debt isn’t Enron-style fraud but mostly lease and chip-purchase commitments disclosed in the footnotes — legal, and hiding in plain sight.
Flagged the real concerns: aggressive adjusted-earnings and stock-comp accounting, circular AI vendor financing (Nvidia, OpenAI, SoftBank), and a “big market delusion” needing $2.5T/year in revenue the industry is nowhere near.
Explained why disclosure still fools people — footnotes are tedious to read and arbitrage has limits — so burying inconvenient numbers where few will look is a rational bet.
Aug 4 | 34 min | Digest ⬇️
Aswath Damodaran | Professor of Finance at NYU Stern | CNBC
Believes AI peaked months ago with more consolidation ahead — the shakeout hits smaller players (à la Situational Awareness), not the cash-rich Mag 7.
Reads the earnings-driven rebound as FOMO, not a fundamentals rethink — managers just don’t want to be left out, and capex keeps surging.
Flags falling marginal returns on capital at Meta, Alphabet, and Microsoft; still calls Micron richly priced on normalized earnings despite the bull case.
Aug 3 | 4 min | Digest ⬇️
Kim Forrest | Chief Investment Officer at Boca Capital Partners | The Exchange by CNBC
Expected a strong AMD report and said sold-out-product commentary tends to lift the entire semiconductor group.
Framed the AI cycle as “the end of the beginning,” where changing compute needs favor nimble players like Intel and, especially, AMD.
Named NVIDIA as the incumbent most at risk if it stays anchored to its GPU-and-gaming roots rather than adapting to customers.
Aug 4 | 10:28 - 14:10 (4 min) | Digest ⬇️
Retail
James Daunt | CEO at Barnes & Noble | Fortune Magazine
Turnaround from near-death (100 closures, 6 CEOs, $1B Nook loss) to 717 stores — 90 above 2019 — after Elliott bought it in 2019 and installed Daunt as CEO.
Playbook: killed pay-for-display and lets each store curate locally, run its own merchandising and social — flexible sizes, bookseller-led, neighborhood feel.
Growth thesis: books aren’t zero-sum — better stores inspire buyers (”come for one, leave with three”), expanding the market; he downplays IPO chatter.
Aug 3 | 8 min | Digest ⬇️
Situational Awareness & Leveraged ETFs
Michael Green | Chief Strategist and Portfolio Manager at Simplify Asset Management | Prof G Markets
explain how excessive leverage triggered the rapid collapse of Leopold Aschenbrenner’s Situational Awareness fund
how leveraged ETFs and their forced daily rebalancing create volatility drag and destabilizing “endogenous flows” in the semiconductor market
why the record $218 billion pile of US leveraged ETFs risks a South Korea-style implosion absent regulatory intervention
Aug 4 | 14 min | Digest ⬇️
Digest
Mohnish Pabrai on Berkshire as an index, the AI “gold rush,” and hunting no-brainers like Kaspi
On what to tell a brand-new investor today: With the market at all-time highs, the S&P 500 concentrated in the Mag 7, and the Shiller PE near tech-bubble levels, what should someone opening a brokerage account actually do?
The kind of investing I do is we make very few bets. They tend to be large and infrequent, and my opinion on the things we don’t buy is irrelevant — what matters is what we act on. So I’m not an investor in the S&P 500. I generally feel it’s overheated, but whether it’s overheated, fair, or overpriced doesn’t matter, because we aren’t placing a bet there.
We try to place bets where the odds are heavily in our favor, approaching no-brainer territory. Anytime things become murky or debatable, it’s an automatic pass.
Normally I’d tell an individual investor with a long runway to dollar-cost average into an index. But with the S&P where it is, that’s probably not the best direction — at worst it’s not a no-brainer, and more likely it’s ridiculously overvalued.
An alternative that may work for a lot of people is to think of Berkshire Hathaway as an index. Don’t buy the S&P — buy BRK.B. Something like 40% of the market cap is cash, another 25-to-30% is very good publicly traded businesses, and then there are a lot of great wholly owned businesses. Berkshire is probably either fairly priced or underpriced, but not overpriced. And the big advantage is that if we get a dislocation of any kind, Greg Abel is going to step up to the bat.
On whether Berkshire stays the “Buffett ETF” now that Abel has the reins: With Buffett stepping down, does Berkshire remain the Buffett ETF over the next 10 to 20 years, or become something different?
Even if Buffett had a life expectancy of 200 years and decided to run Berkshire for another 70 or 80, he’d have the same problem Greg Abel has: size is a big anchor. Warren Buffett with a trillion or more to manage has his hands tied behind his back versus a Warren Buffett with $50 or $100 billion.
It’s not a Greg Abel versus Warren Buffett situation — size is the 800-pound gorilla in the room. If you gave Greg Abel a billion, or even $50 billion, to manage, I think he’d be killing it in a way he can’t now.
The reason I like the Berkshire index is that if we don’t get a dislocation, we’re not taking a lot of risk. If we do get a dislocation, we may be looking at a double in a few years, because Greg definitely has the temperament to get very aggressive. So you’ve got optionality on some great pre-optionality — and pre-optionality is a very good thing.
On the roughly $400 billion cash pile: Would you do anything differently — be more aggressive returning it to shareholders — or is holding and waiting the right call?
We’ll just hold it, wait, and be patient. They did recently put about $30 billion to work in Google, so they’re generating $50, $60 billion or more in cash flow and they did put a decent chunk to work.
But people shouldn’t read too much into it, because Google is a $4 trillion market-cap company. Warren has to make those big bets because his universe is so small — he can’t play the Mickey Mouse games I get to play, which are a lot of fun. He’s way past all that.
I don’t think they should be distributing cash. We’ve had a long run of the S&P going up, and it’s not unlikely that in the next five or ten years we see a big dislocation. If we do, that might wipe out all their cash.
On who really made Berkshire’s Google call: It looked like it might have been Abel taking Berkshire in a new direction, but Buffett said on CNBC that he initiated it — what do you make of that?
It makes more sense that Warren did it than Greg. Warren has studied Google for a long time. Geico has been a customer, and they kicked themselves for not buying Google back when they knew Geico was being charged $20 a click that cost Google less than a cent — they could see how incredibly profitable it was. And before Google went public, Sergey and Larry came to see Warren for advice because they wanted to run it like Berkshire. He’s had a front-row seat for a long time.
But we have to temper that with how small his universe is. If he were managing $100 billion, there’d be no Google bet. In that CNBC snippet he almost downplayed it — he said he doesn’t like it as well as at least four or five other businesses they own.
And Google is now a company with debt, which it never was, and with high capex, which it never had. All these guys have high capex now. Zuckerberg came out and said that whether the bet works or not, they have to play. Think about that statement — he’s saying, I’m putting a lot of money to work and I don’t know whether it’ll pay off, but I don’t have a choice.
So I think what happens in AI when it shakes out is two or three players make out big time, and there are a lot of carcasses on the roadside. We don’t know who the carcasses are and we don’t know who the winners are. It’s a very difficult game right now.
On the AI “gold rush” and buying the pickaxe makers: Is the move to own the pick-and-shovel suppliers — the memory makers and data-center essentials — the smarter way to play it?
When Google spends $100 billion in 2027, that’s the equivalent of spending $20 billion five or six years ago — that’s how much the prices of what they’re buying have gone up. So it’s the gold rush, and the people selling the pickaxes are making it coming and going. The numbers look big — $100 billion, $200 billion, $400 billion — but you have to calibrate that it would have been a fourth or a fifth of that a few years back.
But even the pickaxe makers, for me, go in the too-hard pile. Take the memory guys. There are three of them and enormous barriers to entry — patents, scale. Years ago I was talking to the CFO of Micron, and he told me that if one of their fabs burned down and they tried to replicate it — with all the patents, engineers, process people, everyone who built it the first time — they’re not sure they could get the same throughput out of it. He said there’s a part of this business that’s black magic.
So the memory guys look insulated. The three of them can’t keep up, they’re on allocation telling people to take a number, and they’re jacking up prices unbelievably. That’s why there’s a four-or-five-to-one delta between what Google was paying and what it will pay. But even there, the question is where this is three or five years out — does one of the three jump in front of the others? We don’t know.
So these areas that everyone talks about on YouTube and podcasts — the simple Mohnish rule is that’s not where we want to invest. We want to invest where nobody’s interested, nobody’s talking about it, the boring stuff, where it’s a no-brainer and you’re being hit over the head by a 2x4.
On what “boring” actually looks like: If not even the fallen software-as-a-service names, then what?
We want things that are orgasmic. Let me give you an example: a company based in Kazakhstan called Kaspi. It’s listed on the NASDAQ, it’s a US-listed company, and it cash-flows $2 billion a year. Kaspi is the WeChat of Kazakhstan — a super app that does everything.
About ten years ago it was a broken, failing bank. A rock-star manager came in as CEO — he now owns about 40-to-43% of the company — and converted it into a super app. Whether you need your driver’s license or you want to do some shopping, anything under the sun, you do it there. Kazakhstan has 10 million people, and there’s no other company in the country producing $2 billion of cash flow. That’s an outlier.
He was pumping out a billion a year in dividends because they’re so profitable. Then he decided to replicate what he did in Kazakhstan in Turkey — he bought a small bank and a failing fintech there. To fund it, he told shareholders he was shutting off the dividend for a year or two, and they took the stock out back and shot it. He made his bets, then turned the dividend back on after a 12-to-18-month pause. The dividend yield now approaches 10%, and it trades at like five to seven times cash flow.
The way I look at it: the core Kazakhstan business is still growing, a monopoly adding bells and whistles, and then there’s this moonshot in Turkey. If the moonshot doesn’t work at all, you make two or three times your money. And if it does work, we don’t know. It’s not “heads I win, tails I don’t lose much” — it’s heads I win, tails I win.
On how you find these ideas: Is it just turning over stones, or the curiosity to keep chasing them?
We need to go back about 2,500 years to the Upanishads in India. One verse was written for our benefit: as is your wish, so is your will; as is your will, so is your deed; as is your deed, so is your destiny. And the punchline — your deepest desire is your destiny.
You have to buy into that hook, line, and sinker, not as a skeptic. And you cannot have three deepest desires; you have to have one. If you’re burning-passionate about one thing, it’s going to happen. If your deepest desire is to buy stocks at a PE of one, then out of 50,000 stocks in the world you’ll find them. If it’s a PE of 50, you’ll find those too. So it’s very important to have the right desires — don’t blow it with some stupid desires. It has to be something that resonates with your soul.
Look at Bezos driving to Seattle — he just wanted to build the world’s largest bookstore, and he went all in. Elon wants to die on Mars, just not on impact. That’s the power of a singular focus.
My focus is that I want to make investments I can explain to a 10-year-old in about four sentences. And finding them is easy — there’s a free website called Value Investors Club. You give them your email and see all the ideas with about a 60-day delay, which is irrelevant. Just read every write-up until an idea hits you with a 2x4. The company and the person are both there on a platter; you just have to read it. Like someone asked Warren where to start with 5,000 US stocks, and he said, start with the A’s. If it’s your deepest desire, it won’t feel like work — it’ll be like watching the highlights of the World Cup final.
On drawing the line of your circle of competence: You visited Turkey personally with investor Matt Peterson — where do you draw the line before pushing too far into unfamiliar markets?
Buffett says that in investing there are no called strikes. In baseball, if you let three good pitches go by, you’re out. In investing, you’ve got unlimited dot balls — you decide which one you want to hit, and it’s not going to hurt you to let the others go.
So you don’t need to, and shouldn’t, invest in Turkey or anywhere else until you’re 5,000% all-in. If you’re harboring doubts, the answer is simple: we move on. Buffett says he can let 10,000 balls go by, and when there’s a big fat pitch coming down the center — when the ball looks like a watermelon — that’s when he swings. You might look at Kaspi and not see a watermelon; you might see a tiny marble in the corner of your strike zone, and you let it go.
This is a very forgiving business. You don’t need to invest in the same things I do, and I don’t need to invest in the same things Warren does. Because we have auction-driven markets and humans vacillating between fear and greed, there’s always one part of the market where everyone’s exuberant and hyperactive — let them have their fun, leave it alone. We hang out in the other unpopular nooks and crevices. With less capital, your opportunity set actually gets larger — you don’t need to make Nvidia-type bets, you can make a lot of Mickey Mouse bets. But you have to have the discipline to know your circle of competence, know when something’s a total no-brainer, and have the patience to let it play out.
On handling uncertainty: Using Adobe — down 36% over the past year and 63% from its 2024 highs amid AI image-generation fears — what steps do you take to protect the downside when investors see uncertainty?
Adobe is a very simple case. If you have strong conviction on what the minimal cash flows would be over the next five, ten, or fifteen years, and you discount them back, the decision becomes obvious. If those cash flows are robust and point to a cheap stock and you have confidence in them, you proceed. If you can’t have high conviction, we move on.
Adobe may be a no-brainer for one person and a too-hard pile for another. For me, it goes in the too-hard pile. Compare it to Kaspi: Kaspi also has risk, but I can’t see anything hitting those cash flows from left field — in fact there’s a case they go higher, because they’re still innovating in a rich country. Turkey may or may not work, but if it does, it has eight times the population of Kazakhstan, and they might do three more countries after that.
So if you put a gun to my head and said I could make only one of two investments, Adobe or Kaspi, I’d choose Kaspi. In your case, you could just say neither — if you’re not comfortable, keep looking, because there are 50,000 companies. You keep turning the pages.
On the SpaceX IPO: After the biggest IPO in the world and what’s happened to the stock since, how do you react?
Elon is not human — he’s an alien. I don’t know if there’s any manager on the planet as good as him; his capabilities are superhuman. The guy doesn’t know anything about rockets and he kills all the rocket companies, landing two rockets simultaneously backwards, and they all laughed when he first said he’d do it. Then Starlink, and on and on. As Munger said, never underestimate someone who overestimates themselves. So Elon is the kind of person one should never short.
Now, going long Elon — for Mohnish, too-hard pile. You can read SpaceX’s offering documents about mining asteroids and intergalactic adventures, and it may come about because he’s superhuman, but I don’t need to make that bet. Kaspi versus SpaceX, I’ll go to Kaspi — easier to understand, and if I lose the money I’ll know how.
But the SpaceX business is phenomenal. He’s brought together people executing things they never thought they were capable of — he personally interviewed the first 3,000 hires — while running five companies and going to fix the US government on the side. I’m so proud to be an American that this country attracted a finished product like Elon; he couldn’t have done this in South Africa. He lives eight miles from me in Westlake Hills and is single-handedly driving the Texas economy. I hope we have 100 Elons. So SpaceX doing well as a business is almost a no-brainer — but SpaceX as an investment is a very different question, and that goes in the too-hard pile.
On the discipline of the “too-hard pile”: Why do so many investors resist using it early in their careers?
Humans have a high ego. They’re not willing to admit, this is something I can’t figure out. Our natural tendency is: SpaceX, of course I can figure it out; Adobe, of course I can figure it out. So it’s an exercise in humility. The too-hard pile should be used very aggressively — because that’s when you’re being true to yourself.
On the “passive investing bubble”: Is the flood of passive money distorting price discovery at the top of the market and setting up trouble?
My first feeling is to put that question in the too-hard pile. We have a saying in Hindi: why ask the address of a home we’re never going to visit? I’m not making an investment in the S&P, so I don’t need to convince myself I can answer that. More than likely we’re very far from that point, but whether we’re far or close isn’t relevant, because it’s not the game we’re playing. Much more important is what the hell is going to happen to Kaspi.
One of the fallacies smart people fall into is wanting to have an opinion on everything. It’s a big exercise in humility to say I don’t know much and don’t have much of an opinion on most things. We don’t need to Monday-morning-quarterback everything.
On the non-negotiables of a checklist: What will really run you into trouble if you’re not careful?
One thing that’s absolutely non-negotiable is understanding the nature and competence of management and owners. These have to be very high-integrity people with very high capability, and you need enough information to know that.
I’ll give you a recent example. There’s a US homebuilder, NVR, famous for buying back 80-to-90% of its stock over two decades with no dividends, delivering terrific returns. But the original founder is gone, and now something like 40-to-50% of the shares they buy back end up in the pockets of the managers. The feast the knights managing the castle are having is excessive — and when I see that, it immediately says we’re done.
A fund-manager friend who likes the business said that because of this issue it’s a very small position, and without it, it’d be a very large one. I told her that makes no sense — why would you want to be slightly in bed with a crook? I don’t even want to be in the same room. There are 50,000 stocks; this is a very basic issue. I don’t want to be in bed with greedy managers. With Elon, you might call him a greedy manager, but he sets such crazy targets that hitting them is so hard, so I don’t have much issue with him on comp — just with other things that are hard to figure out.
On judging capability of management: What exactly should people look for?
Warren and Charlie answer this really well: they just look at the track record. They don’t care what the manager says he’s going to do in the next five or ten years — they look at what’s happened over the last 20, 10, or five years and calibrate the nature of the manager. With Kaspi, I’ve never had any interaction with the manager, but I spent a lot of time studying the track record, and it gave me tremendous confidence I was dealing with someone of very high integrity and capability. Many times an investment isn’t obvious, and when it’s not obvious it goes in the too-hard pile. It’s only the anomalies that don’t.
On your single most influential mental model: Which one matters most in your own life?
Focus. It’s another way of talking about the deepest desire — we have to be all in on one little thing, and it has to be everything to us. If we do that, life is going to be amazing.
On what you’d do differently starting over: With today’s brain, how would you redo your investing journey?
I was very overdosed on Graham and very underdosed on Fisher and Munger. That realization only came to me seven or eight years ago — and I’ve been an investor for about 32 years, so for almost a quarter century I was wandering in the wilderness, aimlessly misdirected. If I went back, I’d focus a lot more on the Fisher-Munger model — maybe 20% Ben Graham, 80% Fisher-Munger.
On selling versus holding great compounders: You changed Brandon’s mind on this — how do you now decide whether to hold a high-quality business or take profits when it runs up?
Capitalism is very brutal. Any time you have a successful enterprise, there are lots of people waiting to take you down. So it’s the exception to the rule that a business survives a long time and does well. What happens in a very small sliver of businesses is that a moat gets built almost accidentally — Visa, Mastercard, FICO, Moody’s, American Express, Ferrari. The founders themselves didn’t expect these to become as moaty as they became.
If we’re in the fortunate situation of owning part of one of these, we don’t want to sell a Costco, a Coke, a Visa, a Mastercard, or an Amex unless they’re egregiously overpriced — not just overpriced, egregiously. The valuation has to be so extreme you cannot justify it, and in all those names, that’s currently not the case. Costco trades at about 50 times trailing earnings and has never been egregiously overpriced; 250 times normalized earnings would be. They may be overpriced, but they haven’t gotten to crazy numbers.
It is so rare to have a great business that survives 50, 100, or 150 years on terms agreeable to its owners — so if you’re a partial owner of one, don’t step away quickly.
On private-credit redemptions and bank exposure: A subscriber asks whether you’re worried about the recent problems in private credit and how exposed the banks are.
Mohnish has no intelligent thoughts on that. It’s in the too-hard pile. I’ve never spent any time thinking about it, so I have no idea. I see it pop up in the media occasionally, and I think there are problems, but I don’t know if I’m the right guy to talk about them. Like I said — stay focused.
Michael Kantrowitz: the industrial revival, the flat Mag 7, and whether record valuations can hold on CNBC
On the rally making new highs despite the Fed: With bond yields falling and Kevin Warsh not standing in the market’s way, how does he read the Fed’s role in the move?
The market has struggled in the last three years when we’ve been at these levels of the 10-year yield. But this year, the impact of higher rates has not been as negative.
There are two reasons for that. One, the earnings numbers are just ridiculously strong, for lack of a better term, and they’re also quite broad — and the macro data is quite strong too.
And number two, a lot of the recent move in rates is tied to oil prices. If you look at how the market’s been trading with oil, it’s just not as afraid that we’re going to have a long-term negative situation. Today’s another good example of that.
So if we do open the strait again, we’ll likely also see lower rates and, importantly, a lower chance of the Fed raising rates in September and likely the rest of the year.
On whether semiconductors are still the market’s leadership: After a sharp NASDAQ selloff, with the memory names bottoming and Intel up 10%, is the group still directive for the rest of the market?
Yeah, I think it is. But in the wake of the really sharp decline in the NASDAQ and semiconductors, we did see the equal-weighted S&P 500 continue to do well.
That’s because there are other areas of the economy doing much better than in prior years. There are a lot of other stocks seeing upward earnings revisions, aside from just semis.
About a week ago we wrote that it was time to look for names within the tech space as earnings season had begun. We thought earnings was a much-needed positive catalyst to help support those stocks once again.
On the five reasons the ISM revival matters: He’s told clients a strong ISM report like yesterday’s matters to them — can he run through why?
Sure. One of the top questions we’re getting from clients is whether this cyclical improvement in the ISM this year is over. We put together a list of five reasons it’s a positive message for that data continuing to improve.
One, when the ISM improves, you typically also see small business credit conditions improve — because the ISM is such an important measure of cyclical macro breadth. When it’s going up, usually most of the economy is doing better, and that’s something we didn’t have the prior years.
The second is easier lending standards. Yesterday we got the quarterly loan officer survey, which indicated commercial banks are easing lending conditions.
The third is improving fundamental earnings breadth, which we’ve certainly seen all year. The fourth is healthier market breadth — how many stocks are in a rising trend above their 200-day moving average. Better PMIs typically correspond with better market breadth.
And lastly, one of our high-conviction views this year is to tell investors to keep owning stocks across all 11 sectors that are highly sensitive to the ISM. Generac is the most ISM-sensitive name within Industrials. Even within Staples, a sector that hasn’t done well for years, Target is the most cyclical name. And within Discretionary, Ford is the most cyclical name there. The list is longer, but those are three examples we continue to like.
On what to do with the Magnificent 7: When you can make the kind of money Caterpillar has made this year, where does that leave Amazon, Microsoft, Google, and the rest?
The difference this year — well, two differences. One, the earnings outside the Mag 7 are fantastic this year. That wasn’t true in 2024 or 2023.
Also, the P/E of the S&P 500 is down 10% this year. Just imagine if the P/E were flat: earnings are up 25%, so the S&P would be up 25% this year. A lot of that compression in multiples is coming from the Mag 7.
As strong as their earnings are — still double-digit numbers — their P/Es are down just as much, giving you a pretty flat year-to-date performance for the group. Meanwhile there are stocks seeing better earnings where multiples are expanding and getting rewarded. There are plenty of opportunities in this macro backdrop.
On the Buffett Indicator and record valuations: With market cap relative to GDP at the highest levels on record, is this sustainable, or is a right-sizing coming?
There are two questions: why are multiples up, and what could change that?
I’d argue multiples are up because, fundamentally, margins are at all-time record highs. Margins have the highest correlation in the fundamental data to the level of the market’s P/E.
Number two, look back to 2022, when multiples came down sharply. They didn’t come down because they were high — they came down because we had an interest rate and inflation shock. Today the multiple is down 10%, and everything, whether it’s market cap to GDP, the Shiller index, the forward earnings number, or the Fed model, has come down a little this year. That’s due to higher oil and the higher odds the market’s pricing for a Fed rate hike.
So it’s really not about whether P/Es are too high. It’s what risk could show up that could really spook the markets. Right now, that’s probably still high interest rates and the potential for the Fed.
Roth Chief Economist, Michael Darda, talks the industrial revival, Warsh’s yield jitters, and the “C-shaped economy”
On whether Kevin Warsh is causing chaos or resetting Fed policy: Asked for his read on the new Fed chair’s moves, where does he land?
All we can do is look at markets and try to interpret the message. If we look at the move in bond yields between the last two Fed meetings that Chair Warsh presided over, the move up was dominated and dictated by a rise in the term premium.
So it seems markets have gotten a little jittery about what the Fed’s reaction function is, or is likely to be going forward, and that uncertainty has pushed rates up. That’s not really what you want to see if you’re a Fed official.
But there’s certainly time for things to calm down. We’re going to get a lot of data before the next FOMC meeting — two CPI reports and two employment reports before September.
On the strength of the economy and the industrial revival: If an uncertainty premium is lifting yields, how should the Fed respond to an economy showing this kind of industrial revival?
I actually think the surge in yields maybe over-rotated a little. Most of it has been driven by an economy that’s been resilient. If you look at macroeconomic surprises, they’ve risen quite substantially this year, and that characterizes most of the rise in bond yields. What’s happened over the last month is something different, related to an uncertainty premium.
But it’s a good story, at least on the manufacturing side. This is seven consecutive months now with the ISM manufacturing index above 50, and that’s after essentially a three-year slump. So this is something new, and it’s quite encouraging to see as a sustained move.
On whether the economy is a K, a C, or something else: With the Treasury secretary declaring the K-shaped economy dead in favor of a C-shape, what letter would he assign?
A lot of these letters are better served for entertainment purposes, really just describing what’s going on. The reality is that median incomes tend to go up when wealth rises, so this whole notion of a K is a bit of a misstatement.
That said, when you get things like energy shocks, real wages went negative with the spike in headline inflation — which hopefully is reversing now, and will continue to. So as long as the labor market stays intact and the Fed’s not forced into wrenching adjustments, it’s a pretty good story: real wage growth will recover as headline inflation eases. And that’s not a K, that’s something else.
On whether the Treasury secretary is overstating the recovery: Pressed on whether that means Scott Bessent is overstating things —
No, he’s not wrong, but maybe not as rosy as portrayed, in the sense that real wages are just barely keeping pace with inflation. Year over year — I’m just looking at these now — it’s a 0.26% gain in real wages. It’s slight.
On calls for the Fed to be forced into hikes: Some, including JP Morgan, now think the Fed will be forced to hike later this year — might Warsh’s uncertainty force its hand?
“Forced to hike” is one thing if we’re talking about modest adjustments upward — that’s being priced in already. For September, the markets are now looking for more than a 50% probability that the Fed lifts the policy rate. But one or two adjustments is quite different than 75 or 100 plus. That distinction is really the difference.
It’s going to depend on where the economy is. If you just look at the GDP data, it looks like we’re overheating — the underlying nominal growth rate in Q2 was almost pushing into double-digit territory. That looks like the Fed is way behind the curve.
But the labor market data doesn’t look anything like that. There’s a nominal spending-income proxy you can get by summing payrolls, hours, and wages, and that looks much steadier — in the mid-4s. The bond market is telling you that data is more accurate.
On what the Fed should actually do: Should it be targeting that kind of nominal growth, and where does that leave policy?
If they were targeting that kind of thing, they’d be slamming the brakes right now. But I don’t know that anyone’s trying to bring in a different framework.
What they should do is essentially what Powell and company have been doing over the last few years. The reaction function was much better. Obviously they made an inflationary mistake in ‘21 and ‘22.
But absent these supply-side disturbances, we were getting to a pretty good place in terms of the business cycle and inflation. And the bond market thinks we’re still there. That’s pretty important, in my opinion.
DigitalOcean CEO, Paddy Srinivasan, on DigitalOcean’s pivot to AI infrastructure, 800% inference growth, and disciplined capex
On whether DigitalOcean is an AI story: After a beat on the top and bottom lines, is this an AI story?
Absolutely, it is. As most people know, we’re one of the most beloved names in the world of developers — they just love being on DigitalOcean’s cloud. Over the last 18 months, we’ve pivoted squarely into the AI infrastructure space.
We just reported a blockbuster quarter, growing at record levels — 29% year-over-year growth, and historically our best new ARR add quarter, predominantly driven by AI.
Our AI inference cloud just reported an 800% increase in inference revenue year over year. We’re already at $200 million-plus in AI revenue, and it’s tripling year over year for the last several quarters. So we are absolutely an AI story.
On how the business actually works: In plain terms, customers rent computing power instead of running their own servers, and DigitalOcean handles the hardware — is that basically right, with so many players now in the space?
Yeah, we are exactly like AWS, but focused on the AI-native cloud ecosystem — some of the names you mentioned, including Cursor and others. We focus squarely not on the large, classic brick-and-mortar enterprises, but on the AI-native ecosystem.
Think about this as the Netflix, Airbnb, and Ubers when they were in the 2009, 2010 timeframe. Our target market is these AI-native companies that are re-scripting software.
On how long the AI capex wave can last: Can the whole ecosystem keep benefiting from rising spend — and how does he think about how long it runs?
It’s a great question. I’m not a betting man, but we’re taking a very different approach. We’re a very disciplined company when it comes to investments.
We’ve all seen the headline capex numbers from the large hyperscalers. We’ve increased our capex significantly as well, but we are one of the very few companies projecting 50% growth for next year, and we’re disciplined in that we’re positive cash flow this year and into the foreseeable future.
We just reported a 24% adjusted operating income margin, and we’re very cash-flow positive — the last 12 months we’re at 17% cash flow. So we’re pursuing the AI infrastructure trade, but doing it in a profitable manner with little to no debt on our books.
Gavin Baker talks pressure-testing the AI selloff, GPU prices going vertical, and hyperscaler under-earning on Invest Like the Best
On the month that just happened: AI names are down 40 to 60% in a straight line — what’s on your mind after a stretch like that?
I would describe July as 2022 in a month. There are some fundamental negatives we should talk about, but on the whole, the balance of fundamentals is improving significantly. Loads of AI names are down 40 to 60% from their highs in a month.
And you asked me before we started — I’ve been out here for the summer, and I haven’t heard a single negative quantitative metric about AI. Not one instance of deceleration. In fact, every metric is accelerating. However you cut it — GPU availability, GPU rental pricing, the spot price of DRAM this month, token growth — everything has actually accelerated.
On what the market can’t see: If the fundamentals are accelerating, why is the tape falling apart?
A big part of the problem is that the market does not have visibility into Anthropic and OpenAI, and into these open-source inference clouds that monetize inference here in America — Fireworks, Baseten, Modal, Together. The picture looks very different when you see that, because open source has accelerated massively because of GLM 5.2 and Kimi K3, and Nemotron continues to chug along. OpenAI has accelerated, and Anthropic continues to grow really strongly and is almost certainly pumping out significant amounts of free cash flow.
There’s this chart everybody looks at of semiconductor cash flow going like this and hyperscaler free cash flow going like that — and you’re missing these private companies. But that chart also misses something important: everybody in 2024 and 2025, even if you were really bullish, thought the price to rent a GPU would decline slowly, and if you were bearish, precipitously. I don’t think anyone thought the prices of old GPUs would be going vertical in 2026.
On the contract-versus-spot gap: Why does that vertical move in GPU prices matter so much?
Everybody thought they were going to be smart and sign these long-term contracts, and a lot of the neoclouds had to, because they needed an offtake agreement to finance the GPUs. So you have the contracted base of installed compute trading at a massive discount to the current spot market. As those contracts roll off and compute gets repriced higher, spot can decline and compute will still get repriced higher.
You’ve started to see that this quarter. Operating cash flow — not free cash flow — from Microsoft, Meta, and Amazon accelerated from 28 to 32. There were an unusual amount of one-times this quarter, mostly EU fines; strip those out and you went from 28 to 35. That’s a material acceleration at this scale, and that’s before they light up the Rubins, which will come at a meaningful premium, and before these contracts reprice.
On walking through how we got here: Take me through the month — what actually drove the selloff?
First, Meta was going to rent out compute, and this was seen as very bearish — excess capacity, cutting capex, a disaster. That’s not at all what it was. They saw SpaceX sell some big, trading-optimized clusters into the market at a truly massive premium to contracted rates, and they saw an opportunity. There’s speculation they’ll show strong IRRs on a small chunk of capacity, then raise equity and probably raise capex. They didn’t cut capex — none of the telemetry into Meta’s plans had shifted; if anything it got more aggressive. And shortly after, they released their best model in a long time, Muse 1.1, overshadowed by Grok 4.5.
Then Kimi came out and there was a huge freakout about open source, and the Silicon Data token index dipped and flattened. What that index captures is mix — because of GLM 5.2 and Kimi, there was a shift from more expensive frontier tokens toward open-source tokens. The market thought this was negative, but a token is a token: it takes the same amount of flops, memory, and watts to make. All open source taking share does is take margin dollars out of the frontier model layer and, through elasticity, drive more token demand — pushing more margin dollars into the AI infrastructure layer.
Then China had a DUV machine, which caused a huge selloff in semicap equipment. And then we get to what is in a lot of ways the real concern: real yields have gone up, and spreads widened. Meta priced a bond last week and it did not price where you’d think a Meta bond would. CDS is blowing out for everybody. Smart private-capital people say it’s just banks hedging their commitments — but it doesn’t look good, and these are undeniable facts. That would be really scary if we needed debt to finance this buildout.
On whether this buildout actually needs credit: How much of it has to be debt-financed?
This is where the differential between spot and contract pricing for the installed base is so important. If you look at the gigawatts supposed to come on in consensus estimates for hyperscalers — gigawatts of Blackwell and Rubin — they’re essentially modeled to monetize at the rate of Ampere, the A100, two generations behind. That gets you $1.3 to $1.4 trillion in hyperscaler operating cash flow.
I think it’s very unlikely they monetize at the rate of Ampere. If they just monetize at a discount to current Blackwells, it’s more like $2 trillion of operating cash flow, and that takes about $700 billion of credit demand out. Ironically, as these installed bases reprice, the credit metrics look better and it gets easier to finance with credit — whether they choose to or not, we’ll see.
On the “Blackwell air pocket” risk you flagged before: Are you still worried the early Blackwells get used for training that doesn’t generate a return?
We really saw that in the first quarter. One reason I got comfortable with it two months ago was that you were seeing such incredible things out of Anthropic that the market was going to look past it — and it did, in April, May, and June. Then in July, because of this confluence of things, it stopped looking past it, just as operating cash flow started to really accelerate. That’s just a fact — it is accelerating at big scale. Microsoft brought on a huge slug of capacity in June that didn’t even show up in the second quarter.
On the discomfort of a selloff you can’t explain: In 2022 you knew exactly what the market feared. What’s different now?
In 2022, you knew what it was — recession, rates, inflation. DeepSeek, Liberation Day — you knew what it was worried about. In a weird way that’s comforting. Here, all these specific things, with the exception of credit, are just kind of ridiculous. And the fact that it’s still going down — a technician would say that’s a little scary. It’s definitionally the bullet you don’t see that gets you.
I think the three most important words in investing aren’t “margin of safety,” but “I don’t know.” So I’ve been out here pressure-testing every one of these.
On the anecdote that captures it: What are you actually hearing on the ground about GPU pricing?
I literally spoke to a company this morning — one of the sexiest startups people want to be in business with — that rented a cluster of several thousand Blackwells, B200s, somewhere in the mid-$2 per GPU-hour. They’re renting the exact same cluster seven months later and hoping to pay just under $4. You would expect a gentle decline in prices to be bullish; instead we’re up 50 to 60% in six or seven months.
One of the inference clouds went on a podcast and said they’re planning to pay 100% more for Blackwells when their contract expires. That just means all the hyperscalers are under-earning.
On the one negative you could find: Was there a single bearish data point out there?
The main thing people are saying is that third-party data suggests Anthropic’s curve started to go off its trajectory a little bit. That’s the only thing, and it may very well be true. But then OpenAI and open source are massively accelerating, so the sum of the labs is net accelerating. Open source is a little like dark matter to the public markets — hard to measure — but if you track what these inference clouds are saying, demand is clearly accelerating, which makes sense after the capability leap from GLM 5.2 and Kimi K3.
The underlying fundamentals are improving, and Nvidia, as we record this, is at its lowest forward PE of the last 10 years. The only times the semis have been cheaper were Liberation Day and DeepSeek, and those were V-bottoms. The market 100% thinks they’re significantly over-earning. Maybe they are — we need to be humble. But my mission this week was to look for negative data points as hard as I could, and I haven’t been able to find one quantitative metric other than that contested Anthropic third-party data.
On why the whole market moves as one: Is something structural amplifying these narratives?
There’s Mike Mauboussin’s theory that a breakdown in diversity is what leads to bubbles and crashes. Everyone I know in the public equity business, retail or institutional, immediately feeds every piece of news into Claude — sometimes a Claude agent. It’s probabilistic, but there’s probably not much variation in how it interprets that news. It’s almost like Claude is Walter Cronkite for the stock market, and everybody believes whatever it says. It’s really smart, but it’s not always right.
You see the effect. There’s this guy TBU, part of the anonymous semiconductor mafia, who posted a chart of Japanese capacitor stocks and said we’ve had an entire capacitor cycle in six weeks. And it’s true — vertical, then whoosh, before the actual fundamentals even hit.
On the efficiency innovation you’re seeing: Has anything on the long-lead-time research side made you especially curious?
A lot of people seem to feel they are very close to solving continual learning and sample-efficient learning. If those are solved, it’s possible you get a temporary discontinuity in training demand — I was effectively trained on 20 billion tokens, and these models are trained on 300 trillion. But training as a percentage of compute is going to asymptote to something very small anyway. SSI says they’re coming out with their model in August, and there’s a whole generation of new labs focused on this. This would be awesome for the world, and it’s hard for me to believe it would be negative for AI infrastructure demand — but I’m trying to be really open-minded.
On what would actually flip you bearish: What set of circumstances would get you really scared?
If operating cash flow doesn’t continue to accelerate, so we need to debt-finance this — that would be negative, and it’s a function of how Anthropic, OpenAI, Grok, Cursor, and open source do. A dramatic, sustained contraction in GPU prices would be worrisome; the market would react instantly. If the sum of the labs plateaus or declines, that’s really negative, unless it’s just open source growing the pie.
But have you heard anyone say they have too many GPUs? Not a single person — in fact, it’s the opposite. It sounds like a drug market.
On the multi-model future: Doesn’t cheaper open-source inference cut into AI demand?
A lot of people hear “half the cost” and think that’s bad for AI demand. It’s not, because the cost the user pays is a function of the margin on the tokens. You’re shifting from really expensive tokens with 90% gross margins to tokens with maybe 30% gross margins — that’s where the savings come from — but the tokens cost the same compute to produce.
A public company sets up a router, which cuts its AI spend but may actually increase the tokens it generates by shifting to cheaper open-source tokens. That’s just more compute. The GPU compute hours probably go up. And then you have this whole wave of AI natives leaning into it so hard, not hiring humans, putting it into tokens — they’re not slowing down. Meanwhile companies on the East Coast have barely adopted AI, and Europe is trying to figure out how to regulate it before using it. There are these differential waves of adoption all happening at once.
On where the customer cash flow comes from: The pushback online is: even if hyperscalers are under-earning, where does the accelerating operating cash flow ultimately come from?
Definitionally it has to come from faster economic growth through productivity — Satya’s comment that either we start growing 10% or we don’t — or from labor substitution. In a lot of AI natives you’re seeing substitution, not because they’re firing people, but because they’re not hiring nearly as many humans. Gross profit dollars per FTE are vertical relative to past generations of startups.
In the really AI-pilled companies, tokens as a percent of total comp spend get really high — 20 to 25%. Our friend Dylan Patel is at 30%; I’ve heard of 50%. There’s $25 trillion in knowledge work, so take 20% and that’s $5 trillion, which either comes from labor substitution or faster growth — and we really want it to come from growth. Interestingly, founder-controlled companies, adjusted for COVID-era overhiring, aren’t really doing mass layoffs, which tells you something about where they think there’s still opportunity for people. And you’ve seen charts from Cognition, Ramp, and Stripe that the companies spending the most on AI are growing meaningfully faster.
On the game theory of memory: Everyone’s citing long-term agreements — how should we think about them?
Everything’s in a shortage right now. We’re shifting, particularly for memory, from crushing short-term numbers to trading short-term upside for long-term supply-chain agreements — LTAs — where the customer prepays and there’s a floor and a ceiling.
Think about the game theory of breaking one. There are four companies that matter at scale: Amazon with Trainium, Google with TPUs, AMD, and Nvidia, who’s much bigger than everybody else combined. Memory is the dominant axis — the more memory you put with flops for a given unit of compute, the more tokens you get out, which lowers costs, which is why demand hasn’t responded negatively at all. If you break your LTA and leverage shifts back to the memory guys for any reason, you’re out of business — your allocations get cut, and this is a cyclical industry where oversupply is followed by undersupply. That was never the case before. Apple could do whatever it wanted because it had no competitor. Now you have at least four players and all the startups, and if you break an LTA, they break the volume agreement and give it to your competitor.
On Nvidia’s position and its new “credit wrapper”: Given how much this environment favors Nvidia, why is it trading at such a low multiple?
Nothing is more financeable than an Nvidia GPU, and they’re doing a very good job matchmaking land and power. They’ve rolled out a really clever new business model I’d describe as a credit wrapper with a revenue share if GPU prices are above a floor. This could give them a giant cloud business through royalties really quickly, and it alleviates the cash-flow mismatch. It’s not vendor financing — someone else loans the buyer the money. They still make equity investments, and they write in that the money can’t be used to buy Nvidia chips, though money is fungible.
If I were the CEO of SK Hynix, I’d do the exact same thing — go to the buyers and participate in a credit wrapper. It’s a logical extension of the LTAs: trade upside for durability and get a royalty on recurring revenues. Essentially every time Nvidia hasn’t taken an equity stake in something, it’s been a mistake — they’ve taken stakes in everything except, for a long while, the memory companies. Jensen sees every lab and every advance, and what he sees makes him bullish, so take some equity upside plus a revenue share while generating hundreds of billions in free cash flow to bridge the gap.
On the labs’ arms race: After watching OpenAI and xAI come roaring back, is anyone going to take their foot off the gas?
Anthropic, if they’d been as aggressive on compute as OpenAI, would have run away with it. Now OpenAI is back in the game, and I think Grok is in the game — those are the companies on the Pareto frontier, and they have the compute. Four months ago Dario was talking about how hard it is: if you buy too much compute you could go bankrupt, but if you don’t buy enough you could lose. We saw what happened — OpenAI got back in, and xAI is in it in a big way with Grok 4.5 and Cursor. From a game-theory perspective, nobody is backing off anytime soon, especially if it can be funded out of operating cash flow.
On the most bullish people you met: Did you meet anyone way more bullish than you, and what do they believe that you don’t?
Essentially everyone out here is more bullish than me. I read something arguing that renting an H100 for a year could cost $250,000 — about 15x the current spot. That wasn’t even in my considered-but-dismissed set of unlikely outcomes. The argument was that margins on compute are going up, the amount of compute is going up, and inference margins are going up — multiply those three and that’s how you get this crazy acceleration in the sum of the labs plus open source. I look at the stock market and feel like a foolish optimist, and then I talk to people at the labs and I’m bearish relative to essentially everyone, which is a strange state of affairs.
On the DUV news out of China: Is this the equivalent of ASML in 2001, or the first chapter of a new story for global cutting-edge compute?
Both could be true. If a DUV machine is a propeller plane and EUV is a jet turbine, they didn’t have it before and now allegedly do — that’s a phase transition, like going from liquid to solid. But that jet engine is 25 years behind. It’s learning by doing; you can’t teleport into the future, you have to go through the cycles. Is it significant? Yes. Did the market overreact? Probably — if it ever hits ASML’s orders, maybe in five years, and the market will have forgotten and re-feared it multiple times by then. It’s very hard as an American to have total clarity on China. For better or worse, we’re decoupling, it’s self-reinforcing on each side, and they’re not going to stop — neither are we.
On everything outside AI: What’s happening to the rest of the market?
Last month, everything but AI was vertical. Open source getting closer to the frontier, and companies like Fireworks making it easy to customize a model to beat frontier performance at meaningfully lower cost, is a godsend for the software industry and all these AI natives. Our friend Vishria said two years ago he’d never seen more companies go from founding to $50 million of revenue in nine months — but back then a lot of people dismissed them as ChatGPT wrappers. Now, with open source, you generate data unique to your use case. Fireworks came out with a product called Nexus — three lines of code and it ingests your data, RLs a model, and routes queries — and that’s the solution for every AI native. If you go from using two or three frontier models to using whatever’s optimal for 30 to 60% of your token consumption plus your own RL model, you’re not a wrapper anymore — you’re way more defensible.
It may be that lower-margin open-source tokens massively inflate the value of the most cutting-edge frontier tokens — if you have cheap 120-IQ open-source models, doesn’t that make a 160-IQ model that can orchestrate them more valuable? Frontier tokens may keep capturing the overwhelming majority of economic value but not all of it, while open-source tokens become the majority of tokens processed. And again, that’s great for infrastructure demand, because a token is a token and takes the same flops, watts, space, and cooling to make.
On the worst thing that could happen: Is the biggest risk regulatory?
Regulatory has to be the biggest risk — it’s the most obvious one. You can’t ignore New York doing a data-center moratorium, and we’re living in this weird post-factual political world. The AI industry has done a terrible job of PR. The narrative in Washington is that data centers raise your electricity prices, take your water, and take your job. The reality is that, given the deals being cut now, when a data center goes in, electricity prices generally go down for everyone around it because of behind-the-meter deals, and developers now build hospitals, schools, police and fire stations. The jobs are ongoing because you need plumbers, electricians, and HVAC contractors — data centers are one of the best things to happen for blue-collar wages in my lifetime, and yet the Democrats who ostensibly represent those workers are taking the jobs away.
A lie goes around the world faster than the truth gets out of bed. An author overestimated data-center water usage by 10,000x — not one or two orders of magnitude — and it got super debunked, like the Popeye spinach decimal error, but people still believe it. Somebody needs to tell the truth — maybe a foundation or a PAC running ads during the Final Four and NFL games showing what a data center actually does for your community. And AI is increasingly saving lives and curing rare diseases; at ASCO this year the vibe was the most scientific breakthroughs ever seen at a single conference. If you have a sick child or parent, AI meaningfully increases the odds of recovery. The industry has to tell that story, because New York feels like the first of many.
On what’s missing from the compute conversation: What are people overlooking on the hardware side?
What happens when you put these SRAM-based accelerators — not constrained by HBM DRAM, often on older nodes — onto the installed base. When you disaggregate inference, decode has two parts, attention and the feed-forward network. The holy grail is prefill on one chip, attention on a super-high-powered chip with HBM DRAM, and the feed-forward network on an SRAM chip. You just can’t beat SRAM for that feed-forward network, and no matter how you set the ratio of compute to HBM DRAM to SRAM on a chip, the workloads keep changing. Being able to disaggregate into these three parts is going to be really positive for the ROI on AI.
On dark horses: Can you imagine a player not currently on everyone’s mind reaching Game-of-Thrones scale?
Micron would be a sample answer to becoming as important as Anthropic, OpenAI, Microsoft, Amazon, or Nvidia. Lin at Fireworks is an absolute killer, and our friend Scott Wu at Cognition is one you’re here to. Those are the most obvious names.
On SpaceX in the public markets: What’s it been like watching it get digested — does the market understand it as a company?
It doesn’t really, because it’s such an everything company. The fundamentals have gotten better since the IPO — Grok 4.5, the Cursor acquisition, which has clearly accelerated meaningfully — and they’ve shown over three years they can bring on more compute faster than anyone at lower prices. One of the more bullish things for compute is that they put a vast amount into the market overnight and it wasn’t even a blip; the market utterly absorbed it.
A Substack writer thinks SpaceX is going to bring on eight gigawatts of compute in 18 months. I’ll never bet against Elon, but that would be a truly incredible feat, and rates have gone up since they signed those contracts, not down. They’re monetizing at something like $50 billion a gigawatt, and consensus for next year is $73 billion — that’s eight gigs at $50 billion a gig, and obviously not all lit up at the start of 2027. It seems very implausible; I almost don’t believe the report. But to this day the only companies that have brought on more than 500 megawatts of power in a year are the hyperscalers, CoreWeave, Crusoe, and SpaceX — and SpaceX has done it fastest and cheapest. Very little is built into that stock for the compute they might bring on, and I don’t think it’s anywhere near eight. As Elon says, they specialize in making the impossible late.
There’s a big New York hedge-fund short case that spot compute goes down 90% and it won’t generate the revenue people think. Maybe — but I’ve seen Elon’s companies do really impressive things. And I spent time at Starbase; orbital compute feels more real every day. Our friends at Benchmark funded StarCloud, an orbital compute company SpaceX is partnering with, letting them use the Starlink laser technology. Benchmark isn’t from the Elon ecosystem, and they chose to fund it without SpaceX’s internal launch cost — that’s a good sanity check. Maybe I’m crazy, and Elon’s crazy, and Benchmark and the SpaceX engineers are all crazy — that just doesn’t seem that probable.
Patrick Boyle: Is Big Tech’s Hidden AI Debt the Next Enron?
The accusation
A few weeks ago, Nikkei Asia reported that the five biggest US tech companies are carrying $1.65 trillion of debt that doesn’t appear anywhere on their balance sheets. Not the debt that you can see — a second, larger pile hidden behind it.
A few days later, the Financial Times found another $50 billion in leases that Nvidia had signed for a single data center in Texas, stuffed full of its own chips — a commitment nobody had known about.
And the number keeps moving. Nikkei did their count before most of these companies had even reported earnings. And when they did report days ago, three of them alone signed nearly $900 billion of new AI commitments in a single quarter. So $1.65 trillion is already an underestimate.
Which is why, if you spend any time on financial YouTube, you already know what people are calling this. The word being thrown around is Enron. Commentators — the ones with big social media followings and no obvious background in accounting — have looked at these numbers and concluded it’s Enron all over again.
So I’ve been practicing my shocked face in the mirror. It turns out to be surprisingly hard to hold that frozen, open-mouthed, pointing-at-a-red-line expression while also looking like you understand what a lease is.
Because here’s the question this video is actually about. Is any of that true? Is this fraud — the real numbers being hidden from investors the way Enron hid them, right up until the whole thing fell apart? Or is it something much more boring and much more interesting? Let’s find out.
What Enron actually was
First, though, because a lot of you weren’t following the financial news in 2001, a quick word on Enron, since the entire accusation rests on it.
Enron was an American energy giant that turned out to be a fraud. It had been hiding enormous debts and losses in a web of secret, off-the-books entities. The accounts investors could see were essentially a fiction. When it unraveled, the company collapsed in a matter of weeks. It took down Arthur Andersen, one of the five biggest accounting firms in the world, and wiped out the retirement savings of thousands of its own employees. It is still the definitive corporate accounting fraud.
So when someone points at big tech and says “Enron,” they aren’t complaining about confusing bookkeeping. They’re alleging deliberate fraud on a criminal scale. That’s the charge that we’re going to test.
Why most of this “hidden debt” isn’t hidden
When you see a headline claiming that tech giants are hiding over a trillion dollars in debt, it’s natural to assume a crime is being committed. But when you dig a bit deeper, a lot of this debt turns out to be long-term purchase agreements for graphics cards and leases on data centers that haven’t been built yet. Under standard accounting rules, if the goods haven’t been delivered, or if the building isn’t running, you don’t record it as a liability on the balance sheet. You disclose it in the footnotes.
When you sign a two-year phone contract, you’ve committed to paying the network something like $50 a month for the next 24 months. That’s a real obligation — you can’t just stop. And if you added it up, you’re on the hook for over $1,000. But you don’t sit down and record a $1,200 liability on your personal balance sheet the day you sign. You pay for it month by month as you use it.
The tech companies are doing the exact same thing, just with more zeros. Instead of a phone contract, it’s a 15-year lease on a data center in Ohio.
What it signals when a company borrows
For decades, financial commentators have been complaining about tech companies hoarding cash — that they use their cash flow to buy back shares instead of investing in anything new. Now these same companies are issuing securities to invest in new infrastructure, and the same commentators have found a way to be unhappy about that too. Which raises the question: what does it actually signal when a company chooses to borrow instead of raising equity? Because it tells you quite a lot.
Investors pay close attention to how a company pays for things, because the choices they make send clear signals to the market. Here’s the intuition. If you actually thought you’d build a machine that turns $1 into $5, you wouldn’t sell half of it to strangers to raise the money to build it. You’d instead try to borrow the money, build the machine, keep the entire $5 profit, and pay off the loan. You want to keep all of the ownership stake yourself in such a high-conviction investment. You only want to sell a percentage of the business when you’re less sure about the likelihood of it making a lot of money.
So debt isn’t always a bad sign. Borrowing to build tends to signal that management thinks the return on investment is worth keeping for the existing owners.
...but they’re also raising equity
Although — and this is where it gets interesting — these firms are also raising equity. In June, Alphabet completed the largest equity raise in corporate history, almost $85 billion, anchored by a $10 billion check from Berkshire Hathaway.
Berkshire might be the last name you’d expect on that list — a firm that generally regards buying back its own shares as more sensible than funding somebody else’s moonshot, and a firm which drives a notoriously hard bargain. It reportedly bought its Alphabet stock at around a 6% discount to the market price, because of course it did. This is not a firm that overpays for a story. But they decided that the AI buildout was worth a $10 billion investment.
When you see big tech raising money by every route available all at once — record debt, record equity, convertibles, the lot — the signal isn’t in which one they picked. It’s in the scale of the capital raise itself. You don’t raise capital like a company fighting for its life unless you think there’s something on the other side worth the fight. Whether they’re right about that is the rest of this video, but it’s hard to argue they’re hiding the spending when they announced part of it in the largest stock offering ever filed.
A timing difference — and the cleverer stuff
And a lot of this resolves itself over time. The leases that haven’t started yet will come onto the balance sheet as real liabilities once the data centers switch on. That part is just a timing difference. Meta alone signed $233 billion in new commitments last quarter — $96 billion of it leases that will move onto the balance sheet as the data centers come into use. The purchase commitments mostly turn into chips and buildings the companies actually own.
What doesn’t tidy itself up so neatly is the cleverer stuff — the joint ventures and off-balance-sheet vehicles engineered on purpose to stay off the books. But even that isn’t hidden in the Enron sense. The details are all there in the accounts; you just have to go looking for them. So it isn’t Enron-like fraud. It’s camouflage, which only really works on people who aren’t paying much attention.
Aggressive in plain sight: adjusted earnings
Now, none of this means that big tech plays it straight. They’re plenty aggressive. They just do it in plain sight, on the front page of the earnings report, under a heading that says “adjusted earnings.”
If you’re spending billions on data centers and chips, those assets wear out and need replacing. But many of these firms would rather talk to you about EBITDA — earnings before interest, tax, depreciation, and amortization. Charlie Munger suggested that every time you read the word EBITDA, you should replace it in your head with “BS earnings.” His point about depreciation was that it’s a kind of reverse float: you pay the cash upfront for the equipment, and the expense shows up later as the thing wears out. Leaving it out is just assuming that physical objects last forever, which is a lovely thought and very rarely true.
And you can see the strain in the real cash numbers. When the latest earnings landed, the four biggest hyperscalers posted their lowest combined free cash flow in a decade — $7 billion between them — and Alphabet went cash negative for the first time since it went public.
Stock-based compensation
Then there’s stock-based compensation. Tech companies love paying staff in stock and then taking that expense straight back out of the earnings they show investors, on the grounds that it’s non-cash. Aswath Damodaran at NYU has called adding back stock-based compensation one of the worst abuses in modern reporting. His point is that it isn’t a non-cash expense in the way depreciation is. It’s a barter. If a company sold shares on the market and used the cash to pay employees, everyone would call that a cash expense. Handing over the shares directly, instead of selling them and paying cash, doesn’t make the cost disappear.
Warren Buffett has been asking the same question for years: if options aren’t a form of compensation, what are they? If compensation isn’t an expense, what is it? And if expenses shouldn’t go into the calculation of earnings, where in the world should they go?
Buybacks on a treadmill
To stop all that stock they’re handing to staff from inflating the share count, the companies use real cash to buy their own shares back, which they present to investors as returning capital. Really, they’re running on an expensive treadmill just to stay in the same place.
And here’s the catch. A buyback only actually rewards the remaining shareholders if the shares are bought cheaply. But a company mopping up its own stock-based compensation doesn’t get to wait for a good price. It has to keep buying on a schedule, whatever the shares cost that quarter — which lately has not been cheap.
The real worry: circular financing
While tech executives might be aggressive with their accounting, the actual cash they’re spending on AI is very real. And it’s the way they’re spending it that has some investors worried.
Nvidia is currently working on a round of AI deals worth more than $750 billion. It’s in talks to backstop $250 billion to help OpenAI lease computing power, and to finance another $350 billion of OpenAI’s chip purchases. It threw $5 billion at a secretive new startup run by former OpenAI chief scientist Ilya Sutskever. Google has agreed to backstop lease payments for Anthropic, effectively handing it a $35 billion loan. SoftBank committed $65 billion to OpenAI and took out a $40 billion bridge loan just to finance the bet.
If you draw the diagram of who owns what, the companies at the center of the AI boom turn out to be mostly investing in each other.
Vendor financing — and what happens when it doesn’t work
Now, if you were a car salesman trying to hit your monthly quota, it might occur to you that lending a customer the money to buy a car from you, and then booking that as a sale, is a very effective way to move inventory — at least until the customer stops making the payments. And the fear in the market is that AI has turned into one enormous version of this: a web of companies funding their own revenue.
Nvidia’s CEO Jensen Huang has called the suggestion that any of this is circular “ridiculous,” which is a strong word to reach for while backstopping a quarter of a trillion dollars of purchases of your own product. But to be fair to him, he has a point. As the Financial Times pointed out, this is really just old-fashioned vendor financing. Telecom equipment makers and plane makers have been writing checks to help their customers buy their products for decades. And the argument for Nvidia doing it is just as reasonable: the AI boom is moving fast enough that a company like OpenAI couldn’t raise enough ordinary debt or equity to build the computing power it thinks it needs. So by stepping in, Nvidia locks in a customer, makes sure its chips actually get used, and if the bet pays off, ends up owning a slice of something that could be worth a fortune.
The trouble with vendor financing is what happens when it doesn’t work. Then it’s a double blow: you don’t just lose the customer, you lose the money you lent them to be your customer. And the credit guarantees make it worse. If an equity stake goes to zero, that’s just money wasted — annoying, but survivable. But a promise to cover a customer’s debts if things go wrong can turn a valuation problem into a solvency problem.
Right now, Nvidia throws off something like $200 billion a year in cash, so if one or two of these startups trip, it can take the hit. The question is what happens as the guarantees climb into the hundreds of billions, and a company that used to avoid debt is suddenly standing behind everyone else’s.
You don’t have to take my word that this matters. The clearest sign is in Nvidia’s own credit market. The cost of insuring its debt against default just jumped by the most on record in a single day, right as this round of deals landed. So the people whose actual job is to price the risk of Nvidia not paying its bills had a look at all of this and got noticeably less relaxed. Because the real risk was never just that the AI market turns out smaller than hoped. It’s that the people buying the chips and the people making the chips are increasingly the exact same people.
The big market delusion
All of this circular financing is happening because everyone involved is convinced that the market for AI is going to be so astronomically large that whatever they spend today will look like a rounding error tomorrow. Damodaran has a name for what happens next. He and his co-author Bradford Cornell call it the “big market delusion.”
The way it works is that a new technology shows up attached to a massive potential market. A crowd of companies crop up to serve it, and investors price each company as if it’s going to be the winner. This is not about the companies talking themselves up — it’s about the people buying the shares. Each cluster of investors looks at their chosen company and sees it as the obvious future giant.
The problem is that they can’t all be right. If you take these companies and add up what the market expects each of them to earn, you get a number bigger than the market itself. Everyone’s been priced to come in first in a race that can have only one winner. Which is how a whole market can be priced for a future that mathematically can’t happen. The story is doing all the work, and nobody’s minding the numbers.
How big is the story?
So how big is the story here? The Economist estimates that the AI buildout is on track to be the largest investment surge in history — around $900 billion this year alone being spent on chips, data centers, and power, with more than $400 billion of it borrowed.
And then they calculated what it would take to pay for all of that. Their estimate is that the industry would need to be earning something like $2.5 trillion a year in AI revenue — which is more than the entire global technology sector earns from everything it does today.
The adoption reality
The actual figure is not close. Adoption is real — around a fifth of American firms report using AI in some way — but a lot of them are using the free versions. According to a Bank of England study, the average American executive spends about 100 minutes a week using AI. That’s not a typo. The largest capital investment in the history of the species is being justified by an hour and a half per executive per week — so somewhere between lunch and the drive home.
And when users do pay, they don’t pay much. The fintech firm Ramp went through actual company spending and found that the median firm was spending, per employee per month, $10.66. So $2.5 trillion a year being spent to capture $10.66 per employee.
The most damning number that came out of the Bank of England’s research was that nine out of 10 executives said that AI had made no difference to their company’s productivity over the past three years. When a technology takes over the economy, people usually tend to notice.
Where AI actually works: small businesses
But that’s the view from the top. Look at the other end of the economy and the picture flips. The people getting real value out of AI aren’t the giants spending hundreds of billions on it — they’re the small ones. According to a survey by the payroll firm Gusto, the share of new business founders who used AI to get started doubled to 60% in two years. They are not using it to cure a disease or replace a department, but to build a website, handle the local paperwork, and do the things that used to mean hiring someone.
Now, some of this new business activity is people incorporating their hobbies, and a shrinking share of these firms will ever employ anyone but the founder — so let’s not oversell it. But the clearest real-world win for AI so far isn’t the company burning billions on it. It’s the person starting a one-man business paying about $20 a month.
When caution flips to hype: SpaceX
When a market gets priced as optimistically as AI has been, the people whose job is to sound a note of caution sometimes decide to do the opposite. Take SpaceX. When it went public in June, it wasn’t shy about the AI framing. Its own prospectus claimed a total addressable market of $28.5 trillion — and of that, $26.5 trillion, or 93% of it, was attributed to AI or Grok, which leaves about $2 trillion for everything else: the rockets, the launches, the satellites, the global broadband network, Twitter. The actual space company is the rounding error at the bottom.
That was the case for pricing the shares at $135. You’d think that would be ambitious enough for anyone, but within weeks a Wall Street analyst put a target of $800 a share on it, which would value the company north of $10 trillion — on a business that did under $19 billion of revenue last year.
Now, you might wonder why an analyst would look at a company losing money on $19 billion in revenue and decide it’s worth $10 trillion. As it happens, there is a reason, and it’s a good one. The reason is what SpaceX is about to do next. One analyst went through the prospectus and added up the spending the company is committed to — commitments that the filing discloses but never totals in one place — and they got to something like $235 billion by 2030. The IPO funds raised by SpaceX covered only a slice of that cash requirement. A gap of around $170 billion still needs to be filled by SpaceX issuing more stock and more debt, again and again, for years — which is a great deal of underwriting business for Wall Street.
The analyst chorus
And here’s the thing about those targets. According to Fortune, analysts at 18 of the banks that underwrote the IPO put out their price targets at almost the exact same time — around 25 days after the stock started trading, which is when the rules let them start talking. The notes were, in Fortune’s words, “almost uniformly bullish.” Morgan Stanley called SpaceX “AI’s final frontier.” Bank of America said it was “paving the superhighway to the stars.” Raymond James compared it to the invention of electricity, the railroads, and the internet. These are supposed to be equity research notes. Out of 30-odd analysts covering the stock, exactly one rated it a sell — and that one works at an independent research firm that doesn’t do any underwriting business.
The firewall that got scrapped
There was, for about 20 years, a rule that made this sort of arrangement awkward, but the SEC scrapped that rule last December. It was called the global research analyst settlement, and it dates to 2003, in the wreckage of the dotcom bubble. It built a firewall between the investment bankers and the research analysts at the same firms — the idea being to stop an analyst publicly rating a stock a buy to help his bank win a fee, while privately emailing colleagues to call the same company a pig. Which isn’t a hypothetical: in the cases that led to the settlement, one analyst did exactly that. Another described the stock he was recommending as a “POS,” which I’ll let you expand for yourself.
The firewall was taken seriously. Bankers and analysts at the same firm weren’t allowed to talk business without a chaperone on the line — two of the highest-paid people in Manhattan needing a babysitter on the phone in case they said something a bit too useful to each other. The whole thing was championed by the New York Attorney General at the time, Eliot Spitzer — better known as client number nine of the Emperors Club VIP, a man with a well-documented understanding of the value of keeping certain transactions off the books.
The SEC scrapped the settlement last December, citing the need for lower compliance friction, which is the regulatory way of saying that the rule had become a hassle to enforce.
Enforcement falling apart
And it’s not an isolated decision. By almost any measure, American enforcement of financial crime has been falling apart for years. White-collar prosecutions have been drifting down over the last 30 years; they’re now running at about half the level of 20 years ago. According to The Economist, the Justice Department has cut its lawyers by a fifth and moved investigators onto immigration and drugs. The SEC brought a grand total of 10 enforcement actions against auditors last year — a fifth of its usual rate — running down the exact oversight that was built after Enron. And since most of these crimes carry a five-year statute of limitations, the trick is often just to keep the plate spinning long enough that the clock runs out. One study reckons only about a third of corporate fraud is ever caught at all.
The real question: does disclosure still fool people?
But here’s the turn, and it’s really the whole point of this video. For the companies we’re actually talking about, none of this matters. The Metas, the Oracles, the hyperscalers — they aren’t committing fraud. They don’t need to. Everything they’re doing is legal, and nearly all of it is disclosed. Which is a far more interesting situation than “will they get caught,” because they won’t.
The real question is this: if it’s all sitting there in the open — the adjusted earnings, the $50 billion in leases, the stock-based compensation added back — does it actually work? Does dressing up numbers that anyone is technically free to read still fool people? Are investors really being taken in?
It turns out that this is one of the most heavily studied questions in all of finance. And the answer is a deeply unsatisfying “yes, a bit.” And here’s exactly how.
Damodaran: investors anchor on the number in front of them
Take Damodaran, who we heard from earlier. His view is that markets are roughly efficient over time, but that presentation still matters, because most investors anchor onto whatever number is put in front of them. Show them an adjusted figure and that’s the only one they’ll use. His objection to adding back stock-based compensation is exactly this: it’s a real cost dressed up as a non-cost, and a lot of people simply accept it at face value.
Sloan (1996): cash beats accruals
Then the harder evidence. In 1996, an accounting professor named Richard Sloan published one of the most famous papers in the field. He split company earnings into two parts: the cash the business actually took in, and the accruals — the softer, judgment-based part that depends on management’s assumptions. And he found something the market apparently hadn’t: the accrual part is much less reliable than the cash part. It tends not to last. Companies whose profits leaned on accruals went on to disappoint; companies whose profits were backed by real cash went on to do better. But the market was treating both kinds of profit as if they were identical — which meant that you could earn excess returns for years simply by betting that it would eventually notice the difference. In plain terms: cleaner earnings beat dressed-up earnings. The polish doesn’t hold.
Bloomfield: the incomplete revelation hypothesis
So why doesn’t the market just spot this immediately? That’s the second idea. A professor named Robert Bloomfield has a nice explanation that he calls the incomplete revelation hypothesis. In theory, a market instantly prices in all public information. But Bloomfield’s point is that “public” and “usable” are not the same thing. The number you need is technically out there — it’s on page 83 of a 200-page filing, split across four footnotes, in a form you have to reassemble yourself. And extracting it costs time, effort, and attention, which aren’t free. So the harder a fact is to dig out, the less completely it shows up in the price. It isn’t that the information is hidden; it’s that reading it is annoying and most people don’t bother. Which is the entire game. The debt isn’t hidden — it’s just filed somewhere tedious enough that you won’t look.
The limits of arbitrage
At which point you might reasonably ask: fine, but if there’s money to be made reading the footnotes, why don’t the professionals just do it? Read the filing, short the overpriced stock, and collect. And that’s the third idea — the limits of arbitrage, a field that the economists Mitchell and Pulvino did much of the foundational work on.
The problem with being right about a footnote is that being right isn’t enough. You also have to stay solvent long enough for everyone else to catch up. If you short a beloved narrative stock because you found an ugly commitment buried in the accounts, and a few million people buy it anyway because they like the founder, that stock can keep climbing for a very long time — and your short position can wipe you out well before the market ever gets round to caring. The smart money is aware of this. So a lot of the time it simply doesn’t bother getting involved in hype stocks at all. The mispricing survives not because nobody can see it, but because the people who can see it can’t afford to bet against the people who can’t.
Bringing it all together
Which pulls the whole thing together. The people not reading the accounts outnumber the people who are. The people who are reading them can’t move the price on their own. And so a company that buries an inconvenient number in a footnote is making a perfectly rational bet: that the crowd won’t read it, and the professionals who do won’t be able to do much about it. It’s disclosed, it’s legal, and it works often enough to maybe be worth doing.
So, to bring this all together: the current panic over hidden tech debt isn’t the discovery of the next Enron. It’s people finally reading the footnotes and realizing how much money is actually being spent. As Damodaran puts it, valuation is a story disciplined by numbers. Right now, AI is a very expensive story, told by companies lending each other the money to buy their own products, and reported through adjusted figures that leave out the most expensive parts. The debt isn’t hidden — it’s just filed somewhere tedious enough that most people won’t look.
The problem isn’t that AI is a fraud. It’s obviously useful, and the businesses getting the most out of it seem to be the small ones — the solo founders and the one-person businesses paying their $20 a month. Which is a wonderful deal if you’re the one renting. It’s a rather worse deal if you’re the one who spent billions building it, and is still waiting for that $20 a month to add up to the $2.5 trillion a year it would take to pay the thing off.
Until it does, the companies building all of this will keep borrowing. They’ll keep filing the commitments in the footnotes, and the market will keep not reading them — right up until one day it decides to. Jamie Dimon put it well recently: will AI pay off? Probably, the way the internet did. Will it pay off the way you expect, on the timeline you expect? Definitely not. Which means that I should probably get back to the mirror and keep practicing my shocked face. I have a feeling I’m going to need it.
Aswath Damodaran on CNBC: peak AI, hyperscaler capital intensity, and whether Micron is “different this time”
On his call that AI has peaked: You’ve said your bet is that we hit peak AI a few months ago, with more consolidation and correction ahead — are you talking about the whole space?
Yes, and I’m including the smaller players as well. The Mag 7 are in many ways the most protected part of the space. They’ve invested tens of billions of dollars, but they have the cash flows and the capacity to carry debt, so they’re not in financial trouble.
When you see a shakeout in the AI space, it’s not so much the Mag 7 we should be watching, but the lesser companies. The recent stories about Situational Awareness in particular point to what can happen very quickly in the rest of the AI space as the correction and the cleaning up continues.
On the whiplash from the Situational Awareness scare to the earnings recovery: Given the many markets and notable blowups you’ve watched over the years, how do you put this rebound into perspective?
I think the facile reason is that it tells you how much investors want to be in this space. There are portfolio managers who are convinced they’re underinvested in AI, and they’re looking for any chance to jump in. So every time there’s a correction, rather than the correction continuing, you see funds flowing into companies — people finally say, now I have a chance to have an AI company in the space.
I wouldn’t put it deeper than that. It’s not like people are rethinking the fundamentals and coming back to investing. They just don’t want to be left out of this party.
On whether the fundamentals actually look better: Heading into these prints, the worries were return on investment and a pullback in spending — neither happened, so aren’t the fundamentals genuinely better?
The second one hasn’t happened — companies are continuing to invest in capex like it’s going out of style. The first one is still a concern. I was looking even at the Mag Five, outside of Tesla and Nvidia, at the marginal return on invested capital — the change in operating income over the change in invested capital. You look at Meta, at Alphabet, at Microsoft, and the drop-off in returns on capital is pretty amazing given how big these companies are.
Unless they start delivering earnings commensurate with the tens of billions invested in capex, you’re going to see a very different kind of company emerging from the mix — more capital-intensive, lower return on invested capital. Nothing wrong with that, but I think investors in these companies are not used to the risks of investing in a more capital-intensive company.
On whether Micron is “different this time”: It trades at five or six times forward earnings, and some argue its earnings momentum is so extraordinary that chip names can no longer be judged as cyclical — how do you assess that?
It’s been that battle between history and the AI shift for the last few years. History in chips has always been that earnings go up and earnings go down, so you’ve got to price it on some normalized version of earnings. The optimists on Micron say that’s not true anymore, because this demand is not a cyclical demand — it’s a secular shift in demand.
If somebody makes a bullish case for Micron, I’m willing to listen. But I think it’s still a richly priced stock if you think about it in terms of normalized earnings.
Boca Capital CIO, Kim Forrest, talks the AMD trade, shifting chip leadership, and the risk facing NVIDIA on CNBC
On a pivotal moment for the semiconductors: With AMD — her favorite — reporting after the bell, how is she thinking through the group and this afternoon’s earnings?
Well, it’s been quite a summer, hasn’t it? It’s been dangerous to have semiconductors in your portfolio — and now maybe it’s dangerous not to have them.
What we’ve been doing is getting nervous about: is this AI thing real? Will the spending continue? And who are the winners that are going to keep winning?
My guess is AMD is going to have a pretty good report tonight — maybe they’ll have sold out most of their products, and maybe they’re looking into the future saying, we’re going to have trouble meeting demand. Those kinds of sentences are going to drive all semiconductors higher.
On Intel’s bounce and the group more broadly: A great AMD report tonight — what does it mean for the market, and what’s driving names like Intel, up 10% again?
If you think back two years ago, we were all about NVIDIA — because NVIDIA was the only thing we could invest in and have it related to AI. NVIDIA theoretically won.
But as we go on in this AI cycle — and we’re at the end of the beginning, maybe — a lot of different things are going to happen, and there will be different winners in semiconductors. That’s because compute is going to change. The needs developers have are going to rest on different kinds of chips.
The most nimble companies offering them are probably your longer-term winners. I’d put both Intel and, to a larger extent, AMD up against that. They’ve come up the curve from being the second banana in x86 chips 10 or 15 years ago, and now they’re in leadership positions in a lot of these high-demand areas of today’s AI.
On who the shift disfavors: If changing compute needs favor Intel and AMD, who does it work against?
Anybody that’s overly stuck in their ways. And it might be NVIDIA.
NVIDIA has always been a great graphics processing unit creator — that’s what birthed them; they were the favorite of gamers. And they just kind of got lucky that AI needed that kind of compute. I think they really have to rethink what they’re doing, who their customer is, and branch off that whole GPU success to maintain leadership.
So anybody that’s stuck in the past, that’s not flexible, that isn’t talking to customers and asking where they want to go next — I think that’s really important.
Barnes & Noble CEO, James Daunt, on Barnes & Noble’s unexpected turnaround, hyperlocal curation, and a possible IPO
Remember how everyone thought Barnes & Noble was dying? Just a few years ago, America’s largest book retailer was in rough shape.
By 2020, Amazon captured over half of all U.S. print book sales and 80% of the ebook market, decimating traditional brick-and-mortar bookstores.
James Daunt: “There was a decade of declining store numbers, lots and lots of closures, no openings.”
Today, the chain is pulling off one of the most unexpected turnarounds in retail history.
James Daunt: “What we’re doing is opening up stores again in communities.”
Will their growth hit a wall?
As of July 2026, the book retailer had opened 20 new stores so far, with more planned this year. The footprint currently stands at 717 locations — 90 more than in 2019.
James Daunt: “We’ll open up 35,000-square-foot stores, we’ll open up less large stores. We’ll open an 8,000-square-foot store in a mall. We’re doing that because we’re putting bookstores back into places and back into communities that will appreciate a bookstore.”
During the 2010s, Barnes & Noble saw massive sales declines, 100 store closures, six different CEOs, and a $1 billion loss on its Nook reader — a device built to compete with Amazon’s Kindle. Then, in 2019, investment firm Elliott Advisors acquired the company for nearly $700 million, bringing in James Daunt as CEO.
James Daunt: “There are books here that you know are really good, and therefore you will read the others. So there is consistency of quality”
On taking over a company in rough shape: When you took the reins, the business was struggling — as you assessed it, what were you thinking?
It’s relatively straightforward for me, because I’d been a bookseller for 30 years at that point, longer. So I simply wanted to make these stores really good again, and one sets out on a series of very slow and incremental processes to do that.
Daunt started out as an investment banker in New York in the 1980s. He left banking to launch Daunt Books in 1990, building it into a highly successful and beloved 10-location chain in the UK.
On whether Barnes & Noble needs the indie “cool factor”: Daunt Books and many independents have that cool factor — does Barnes & Noble need some of it, or is it a different animal, a different concept entirely?
Barnes & Noble can be extremely cool, but it’s definitely a different sort of cool. And the challenge for us is: how do we do it appropriately in each place in which we are? And that, I think, we’re able to do increasingly well at the moment.
Every display in their Union Square store in Manhattan feels thought out to the smallest detail.
Shannon De Vito, Manager at Barnes & Noble: “These are a lot of conversation starters — economics, politics. For this local section, there’s a lot of gift books. There’s cookbooks.” — “I’ve bought that for somebody, it’s so cute.”
The inventory is hyper-curated for New York customers, reflecting social trends, local culture, and real-time demand.
Shannon De Vito: “A lot of our tables are smaller. They’re meant to be little stops along your way to points of interest as you weave your way through the store. We love when people spend time in the stacks and pull things out and find something that may be really old but is new to them.”
A 2025 study showed foot traffic in brick-and-mortar bookstores climbed 12% year over year, while another study showed independent-bookstore accounts surged by nearly 90% that same year.
Shannon De Vito: “Romance has been a booming category over the last few years. It really is landing on that escapism and happily-ever-after that’s working across a lot of different genres.”
For Barnes & Noble, the wins have come from trying to feel like your neighborhood bookstore. Asked about the biggest difference in how stores are run today versus when she was a rookie, Shannon De Vito says:
“We were more a traditional retailer — we were really operating like a big box, a Target or a Walmart, and I was essentially told to kind of turn my brain off when I came to work: put this book here on this date at this time, and then replace it with this book. Whereas now we’ve moved to this model of being bookseller-led, being able to curate interesting tables — a local approach, but also some interesting deeper cuts. It takes a lot more brainpower, in the best possible way.”
Daunt’s strategy was simple: build hyperlocalization into the chain’s corporate playbook. He eliminated the pay-for-display model, empowering each store to curate based on local demand rather than publisher co-op dollars. Stores now control their own visual merchandising — such as handwritten staff recommendations — and run their own social media accounts.
On scaling while staying flexible: How do you get the benefits of economies of scale while keeping this flexibility of size?
We used to have a model that involved a home office dictating very precisely what every store did, attached to a lot of commercial arrangements with publishers to have particular books in particular places — a very uniform presentation, for which you needed a uniform size of store. We’ve abandoned that 100%.
So now each store curates its own assortment, reorders its own books, does its own thing — and we encourage and develop and promote and train against how you do that well. But that also means we don’t really care what size it is, we don’t care what shape it is, because each store is doing its own thing. And that’s liberated us now to have different-size stores doing different things.
We can see what they’re trying to do here — you then want to pick it up, and it’s in your pile, and off you go. Now he’s applying the Daunt playbook to the U.S. market on a much larger scale.
On how American book buyers differ: You’ve now experienced American consumers as well as British — how are they different?
American consumers are wonderful, because they buy enormous amounts of books. Individually, they buy more than a UK customer does. The American consumer, of course, is world-famous in all forms of retail — but it applies also to bookselling.
On where the growth comes from: So the growth for you comes from getting more people to simply buy more books?
The book market is not a finite one. There are retail sectors where, if one person sells plus one, another person has to sell minus one — it’s a fixed market. That isn’t the case with books at all, particularly the way in which people buy books in a bookstore. If you come into a bookstore, you’re going to be inspired. You come in for one book, you walk out with three. That’s the way it works. The better our bookstores, the more books are sold, the larger the market expands. But it’s not eating the share of somebody else — it’s not a zero-sum game at all.
And now Wall Street is taking notice. There have been media reports of a possible IPO.
On a possible IPO: How do you think the company would be managed differently if it were public?
We are trying to run really good bookstores. Most of that effort is being done at the local level. I think what goes on in New York, any which way, doesn’t make much difference. What matters to us is what’s going on at the local level — how are we able to support them, how are they able to improve things. And that’s a steady, incremental process.
I’m a great believer in physical bookstores and have my own bookstores — the equivalent in the United Kingdom of Barnes & Noble — and we sorted that out very quickly and made those really nice bookstores, to the benefit of the whole book trade and bookselling as a whole. And then, obviously, with Barnes & Noble being the last man effectively standing here in the United States, it was very, very important for bookselling as a whole. But we’re just really bringing these principles — first a small bookstore, to a large chain, to an even larger chain — and it’s all based and founded on the core principle of what makes a great bookstore.
Portfolio Manager at Simplify Asset Management, Michael Green, on how leverage blew up Aschenbrenner’s fund and the danger in leveraged ETFs
On the implosion of Situational Awareness: Leopold Aschenbrenner’s fund reported a 439% first-half return, then lost roughly $35 billion in six days and was forced into a fire sale to Citadel — how did it happen, and what’s the lesson?
The quick answer is that when you look at somebody engaged in the behaviors Leo has, there’s really no mechanism for him to have learned not to do this. He had a very strong thesis, and he expressed it with the extraordinary use of leverage.
His initial exposure was largely to non-public entities, and he grew his business under that framework — which has a much lower-volatility framing, because non-public entities don’t reprice themselves in the same manner. But when you start running strategies with that much leverage against that much volatility in the individual securities, a blowup becomes inevitable.
What looks like happened is that he created conditions under which a small decline in prices would force him to sell to reduce his leverage, which caused prices to fall further, which forced him to sell more. Ultimately that cascaded into an event that sent both his longs and his shorts against him.
In particular, he had the thesis that traditional software companies would be heavily disintermediated by AI, and his own selling actually contributed to the underperformance of that sector. As he was forced to unwind, that pushed prices in the opposite direction of his positioning and created the conditions for the rapid collapse.
Nobody in their right mind should give a 25-year-old $20 billion at 4x leverage — but you can’t really blame the 24-year-old. He had a very strong conviction, and everything in his experience base up to that point told him this was the right strategy. Once you become that large, the street actually identifies you as a target. You effectively become a wounded shark, and a feeding frenzy emerges.
On how leveraged ETFs are warping the semiconductors: In a recent piece he tied the volatility in names like Nebius, SanDisk, and Micron to the rise of leveraged ETFs — how is that reshaping the sector, and why should investors care?
A levered ETF carries the same characteristics as Leo’s portfolio running at 4x leverage. The difference is that a levered ETF, because its prospectus requires it to maintain that levered exposure, has to rebalance every day. And that’s where volatility creates a phenomenon called volatility drag.
Imagine I make 10% today and lose 10% tomorrow. Many people assume that’s a zero return. But I start with $1, I’m up to $1.10, then I lose 10% and I’m at $0.99 — I’ve lost a penny. Add four times leverage and you get a 2-to-the-fourth-power impact on that drag.
Say I put in $1 of equity and $3 of borrowing, so $4 of exposure. I’m up 10%, the position is worth $4.40, and I still owe $3 — so my equity has risen to $1.40, a 40% gain on a 10% move, exactly as you’d anticipate. But do the same math on the way down, and the compounding effect of that leverage and the need to rebalance is what creates catastrophic losses.
Because the ETF has to rebalance every day, after a gain it doesn’t just sit there — it has to lever the new equity back up to 4x, forcing it to increase position sizes by nearly 50% to maintain the leverage promised to investors. That creates what’s called endogenous flow: it forces buying even without new investors adding money, and contributes to the run-up unless investors harvest those gains.
My piece is called “a semi theory of everything.” Most professional investors would run these as a volatility-harvesting strategy — you short both sides of the trade and harvest the loss created by the volatility drag, which gives a very stable return profile as long as the volatility characteristics hold.
Unfortunately, in the excitement of the recovery in early April, retail investors stepped into these products seeking a Leopold-like experience. They bought 3x levered ETFs or 2x single-stock ETFs and held them, rather than harvesting, and even tried to dollar-cost average in. At 3x leverage, running the volatility we saw in the semiconductor space in April and May, you’d need a return in excess of 170% a year just to break even. To dollar-cost average into something with a 170% break-even is absolutely absurd — a byproduct of the lack of education and the tools we’ve put into the marketplace to attract people to shiny objects rather than thoughtful investment vehicles.
On South Korea’s crash and whether leveraged ETFs should be banned: With the KOSPI down 44% from its June highs, more than a million Koreans hit with margin calls, and hundreds of thousands liquidated to zero, is abolishing these products the answer?
It depends on how you use them. A hammer is a very dangerous tool if used improperly and a very productive one if used properly. As volatility-harvesting tools, these can be used by professional investors to short a realized-volatility framework and generate profit by providing the financing for those vehicles.
South Korea has already banned the levered ETFs. They’ve been forced to close their market multiple times over the past several weeks — literally doubling the number of closures in a three-week period versus their entire history since around 1990. They’ve recognized these products create almost no social utility.
In the United States, we’re still trapped by market fundamentalism — you see it in everything from Kevin Warsh’s recent Fed testimony to the general regulatory view of “just let the market decide.” There’s a reason we don’t do that. There’s a reason we now have labeling on drugs telling us their addictive contents, and labeling on food listing the ingredients. We used to say “buyer beware,” but that’s extremely difficult for, say, an illiterate immigrant trying to figure out whether the sausage is filled with potato flour or meat. We recognized that and took steps.
I think we often go too far with this, and there is a role for experimentation and the utility of these tools — we could ban hammers because somebody hurt someone with one, and that would be a mistake. But in this case we’ve created a gambling environment in which people are increasingly nihilistic about prices, assuming governments will step in to support them or print money to paper over our problems. That’s simply untrue. And I’d emphasize for the younger audience: never substitute conspiracy when incompetence will suffice. We have regulators who have largely abandoned their role, and we’re left with products I’d describe as half-boiled spaghetti thrown at the wall to see what sticks and attracts investor dollars.
On whether the US is next: Assets in US leveraged ETFs have hit a record $218 billion, up 60% since the end of March — is America headed for a South Korea-style implosion?
Unfortunately, this is one of the key concerns, and it’s tied to my work on market structure and the growth of passive and price-insensitive money. Here we’re talking about leverage vehicles that don’t consider whether what they’re buying is a good thing or a bad thing — they’re simply fulfilling an investment mandate, what I call systematic portfolio rebalancing. That creates the conditions under which these feedback loops play out.
And I’d point out to South Korea that much of its problem wasn’t generated in South Korea. We imposed these conditions through the introduction of an unlevered, memory-centric ETF — a DRAM product — in the United States, which exploded in size to almost as large as Leo Aschenbrenner’s portfolio and was sending roughly half of its dollars in Korean-won-hedged terms, selling the currency and buying the stock in Korea.
It brings to mind 1971, when US Treasury Under Secretary John Connally told the emerging markets, “It’s our currency, but your problem.” This was our ETF and their problem. And I think you’re going to see the regulatory environment begin to recognize that — and it very well may be forced to change.
If you found these summaries helpful, please leave a comment or send us a message. We appreciate your feedback!