Dan Niles — "Bearish into November. Room to run after"

0:02 The set-up — a 30-year tech investor, and a semis call already half-paid

• Justin Carbonneau and Jack Forehand introduce Niles as founder and portfolio manager of Niles Investment Management, "over 30 years experience in researching and investing in the technology business."
• The opening question re-tests his own published call: that we are "100% in a bubble" and that semiconductor stocks would see "somewhere in the range of 30 to 50% pullback" — and semis have since "pulled back quite a bit from June."

1:27 Every industrial revolution overinvests — that is the definition, not the exception

• The frame is historical rather than technological: canals in the late 1700s, railroads in the early 1800s, radio, TV, electricity, the internet. "If you believe you're in one of those revolutions… then by definition, you're going to have a lot of people chasing that opportunity because they know if they're the last company left, they're going to make an inordinate amount of money."
• The inference he draws is counter-intuitive and is the hinge of the whole hour: "that overinvestment doesn't necessarily mean it's a bad time to invest. And in fact, it means it's a great time to invest. The problem is when that eventually breaks."
• The reason he expects a shallower break than 2000: "Cisco… was sort of the Nvidia of its day, it was trading at a PE valuation of over a 100 times of earnings. Nvidia, in contrast, is at 15 times." Hence a hoped-for 30–50% semiconductor drawdown rather than the March-2000-to-October-2002 78% NASDAQ collapse.

4:03 The revenue test reduces to two numbers: price per token and tokens produced

• "There's number of tokens being produced and what you can charge for each of those tokens." Tokens are "sort of the unit of currency for AI."
• The price side has already broken: "since end of May… what you can charge for those tokens has gone down by about 50%. Why? Because you have this thing called open source or open weights."
• The volume side has more than compensated: "the number of tokens being actually produced is up two and a half times since the end of May." He is watching whether that stays true — "is the dropping in costs being offset by a more than offsetting increase in tokens."

4:57 The profitability test — the big three clouds accelerated and expanded margins

• "You may be generating revenues, but if you're losing tons of money, it really doesn't matter." So revenue growth alone is not evidence.
• AWS, Azure and Google Cloud together "saw accelerating revenue growth from the March to the June quarter where revenues went from increasing 35%… to up 43%."
• The half he ranks higher: "more importantly, the operating margins also expanded about 2 percentage points from March to June. And so not only were revenues accelerating, but the profitability of those revenues was also going up."

6:45 Situational Awareness — how 4x leverage manufactures a bottom

• The mechanism, stated as arithmetic: "the problem wasn't necessarily the positions they had. The problem was they were running it at 4x leverage… if you have a 25% drawdown, but you have 4x leverage on it, then you're down 100%. At which point your fund is liquidated."
• Both sides of the book worked against them — "software was rallying and your longs going down. In this case, semiconductors were getting crushed" — which is the same shape as Long-Term Capital Management in the late 1990s.
• Why the blow-up was bullish: with fundamentals intact, the forced sale removed the overhang. "Situational awareness being pushed into that for sale with Citadel helped solve some of those issues because the market obviously ripped the next day." He had published the call on July 29th "thinking we reached a short-term bottom because I had been hearing about funds in trouble."
• The generalisable claim underneath it: "there's only two things that drive stocks… earnings [and] the multiple on those earnings. The multiple… gets over inflated because you have individuals as a group or funds as a group doing stupid things. And stupid things usually means you've levered up."

9:40 The thing that actually changed: hyperscaler free cash flow went negative

• "Your cash flows, which had been hugely positive, like for Google, are now negative for the first time since they went public." That, not growth, is "the part that people are wrestling with."
• How the cycle ends, mechanically: "eventually somebody will have an issue and then they won't be able to live up to their commitments… you're going to see a slowdown in capex spending by one of the hyperscalers, at which point they'll all probably go ahead and slam the brakes on at the same time because they all kind of know they're overinvesting."
• The industry says so itself — the Microsoft CEO has talked about overinvestment, and "Larry Page [said] something along the lines of, 'I'd rather go bankrupt than lose this race.'"

10:09 The 78% precedent — the internet kept doubling all the way down

• "The NASDAQ went down 78% from its peak in March of 2000 to the lows in October of 2002… a punishing, grinding beatdown over 2 and 1/2 years."
• The technology never faltered: "the internet kept growing over that whole period of time. In fact, it doubled each of those years." The error was extrapolation — "people were thinking, 'Oh, well, it's going to double every 3 months or 6 months.'"
• The lesson he carries into AI is that being right on the technology gives no protection at all on the multiple.

12:22 The surprise negative: the data-center backlash

• Asked whether physical constraints act as a governor on the bubble, he calls it "a great point" and produces the one development that surprised him: "data centers are getting pushed back on really hard."
• The number: "in the latest Gallup poll… when people were asked about having a data center in their backyard or nuclear, it was 71% against data centers and only 53% against nuclear." Being less popular than a nuclear plant is the shock.
• It is bipartisan and pre-electoral: "what they all seem to be agreeing on going into the midterms, which surprised me, especially from the red states like Texas, is data centers. They're all pitching running against data centers." He blames the industry's own PR — the community benefits (jobs, tax revenue, their own electricity, minimal water) went unsold.

13:54 "Between now and the midterms on November 3rd, I'm pretty negative"

• The core two-part position of the episode: bearish into the vote, constructive after it — the same land constraint that hurts sentiment now "should help prolong this buildout for even longer because you don't have the land to actually put these data centers on."
• He treats the politics as a short-term risk that education should defuse "especially after the midterms are done."
• The tail he will not dismiss is ownership, not permitting: a landslide for "socialists that want to make these companies forcibly give part of the share ownership to the government."

15:55 Where in the stack the money lands — ask how many railroads are in the top 10

• The screen is a series of questions with the same answer: "How many railroads are in the top 10 market cap in the US? What about electricity companies? What about canal companies? What about fracking companies?"
• The single most compressed statistic in the hour: Cisco, "which used to be the most valuable company in the world, just recently got back above its market cap from 2000. That's a pretty long wait."
• Building the infrastructure and owning the returns are different jobs, and history says the builders usually do not keep them.

16:50 A Ford will do — value moves from model providers to infrastructure

• "You don't need a Ferrari to go to the corner store to get milk, right? A Ford will work just fine… people aren't going to be using Anthropic to go summarize their emails."
• The conclusion: "the value is going to move from the model providers to the infrastructure providers and then ultimately the companies sitting on top of that that create great businesses."
• The precedent is the fibre glut: the bandwidth stranded by the dot-com bust "gave you the infrastructure necessary to have companies like Google and Amazon and Facebook, Netflix all show up." He expects the same from cheap tokens — "you're going to have some really great companies coming around because of that."

18:23 Winner-take-most: Anthropic and Google squeeze OpenAI

• The lineage argument: "ChatGPT was launched for consumers. Anthropic started focusing on enterprise" — so "Anthropic is going to win on the enterprise side and then Google on the consumer side are going to squeeze OpenAI between them."
• Consumers are the weak flank because the reference price is zero: "we've been all trained that hey, we get answers from Google for free. So why on earth would you pay for it?"
• The base rate: e-commerce is "really only Amazon," social "only really… Meta," streaming "really only… Netflix." So "do you think you're going to have 10 models that are all thriving?… the history of technology tells you that's probably not likely."
• And the humility clause — AOL, Yahoo, Lycos, Netscape were all obvious winners once. "Just because ChatGPT is what started all this doesn't mean that OpenAI is going to be ultimately the biggest winner."

21:38 Japan, then Korea, then China — the state-sponsored memory pattern

• Asked whether semis are still cyclical, the answer is a flat "No," and then something stronger: "I'm actually starting to wonder if the semiconductor bubble breaks before" the AI trade does.
• The template: "Intel at one point had 75% market share in the DRAM industry… by the early 1980s, Intel was facing bankruptcy and they switched to this thing called microprocessors and Japan took over the DRAM market." Then Korea, "also state sponsored," entered in the 1990s "and by the 2000s you had the Japanese companies being driven out of business."
• Why the third run is bigger: China dwarfs Japan or Korea "in global GDP… the population size, land mass size, ability to put up these massive factories" — and has an extra motive, because being "cut off from chips from the US" makes domestic supply "as important as having an aircraft carrier or nuclear weapons… more of a defense technology."

22:38 CXMT and YMTC — the numbers behind the China memory call

• "CXMT, which is China's DRAM company, [went] public. And then YMTC, which is their NAND company, is going to go public very soon."
• Capacity: "CXMT plans on increasing their wafer starts from 300,000 to 500,000 by the end of next year"; YMTC "plan to be bigger than Samsung or SK Hynix in the NAND business by the end of next year."
• Share today is small but the slope is not: "about 8% of the global DRAM market which is nothing, but it is ramping incredibly fast." And quality is closing — the same day, "CXMT said they had produced a high bandwidth memory that was actually yielding very very well… only a generation behind."
• The direct hit is on the consensus story: the argument that "there's going to be shortages through 2030" has "very minimal chance" of surviving if China hits its own published targets.

25:32 The Cisco investor-relations exercise — an antidote to "smart people are spending for a reason"

• His prescription is a specific, repeatable piece of homework: "go on to Cisco's investor relations website, pull the earnings release from May of 2021, if memory serves me correctly" — the 2001 release — "where the CEO of Cisco Systems says we've gone from 70% year-over-year bookings growth in several months to negative 30%."
• The point is not that the operators are stupid: "Are these really, really smart companies? Absolutely. Are they the best companies that we have right now? 100%. But do massively smart big companies get it wrong? Yes, all the time."
• Which leaves the investor with a timing problem that cannot be delegated: "you never know, well, is this the actual break or is this just a pause that refreshes?"

27:02 The agentic phase — 10 to 100 times more tokens, and only seven months old

• Phases, not one continuous ramp: "You had the training phase initially. Then you had the inference phase and then on January 30th of this year, you had the agentic phase."
• The consumption step-change is the reason he stays constructive on the volume half of his test: "the agentic phase uses 10 to 100 times more tokens than the chat-based AI phase."
• The novelty has no dot-com analogue — "we didn't have that with the internet… there wasn't a new thing that came out for the internet" — hence "at least another year for stocks to go higher," paired with "watch the data like a hawk."

28:34 The Nvidia print — bullish, with one thing he did not like

• Did the blowout teach him anything new about AI? "Not really. I thought it would be bullish."
• The objection is process, not numbers: "the only thing in that report I didn't like is when they gave long-term revenue guidance, because Nvidia for its entire history has guided out one quarter in advance." He understands the motive — frustration with the multiple — but distrusts the instrument.
• The precedent he reaches for immediately: "Cisco Systems in 2000 was saying… our revenues can grow 30 to 50% sustainably in the future, and then two years after they made that proclamation their fiscal year revenues were down for two consecutive years in a row."
• The valuation defence, unchanged: "trading at a 15 times PE off of calendar 27 numbers with the S&P up at 19 times, you go, this isn't Cisco in 2000. Even if they're wrong, it's not Cisco going 20,000 in terms of the downside."

30:48 Software: $105B of AI revenue has to come out of somebody's budget

• The arithmetic that starts the section: OpenAI plus Anthropic combined annual run-rate revenue "was 29 billion to start the year and now 7 months later it's 105 billion. That amount of money that is being spent with those two has to be made up somewhere else."
• Denominator choice decides the answer. Against software alone, "a little over a trillion dollars," $105B is a real problem. Against software plus IT services it is smaller. Against knowledge-worker pay — "more like 35 to 50 trillion a year in spending" — "you're not even going to notice it."
• He notes the bull argument turns on itself: if agentic AI hits point-solution tools "10 to 100 times more than a human being," then software spending goes up — "in which case, who's going to get hurt?"

32:38 Three safe spaces in software — and three prints that challenged the bear case

• The prior: "within software there are three safe spaces… security, system of record companies, and then a much smaller grouping… of entertainment video game companies." His own caveat: "obviously they didn't act safe until very recently."
• Three contrary datapoints in quick succession — Atlassian, "supposed to be one of the names getting disrupted," up 35% the next day; Silver Lake circling Workday at "like 50 billion," which requires believing in terminal value 7 to 10 years out with 30-year treasuries at 2007 highs; and Salesforce up over 20% on the Anthropic deal.
• The relative-value case is now doing the work: software "was up something like 4% year to date and the semiconductor index was up over 60% year to date… that at least makes you feel like maybe there's some margin of safety." He is not yet committed — "it's making me wonder if I can get more broadly involved in the group" — and notes security's safety is now consensus and priced.

39:13 Issuance, credit, and the most important price in the world

• Asked whether the market can absorb the SpaceX / Anthropic / OpenAI issuance wave, he reframes to the cost of money: "credit is the lifeblood of the economy… you can argue that's the most important price in the world is the 10-year Treasury yields."
• Three pressures at once: government debt at $40 trillion against $33 trillion of GDP; deficits at 6% of GDP, "the highest levels outside of a major war that we've ever had"; and a new competitor for capital — "in the past, you didn't need to compete with AI debt spending by some of the biggest companies in the world, which in the past were massively cash flow generative like a Google."
• Result: "the 30-year is up at the highest level since 2007, that's a problem."

40:48 The chart he checks every morning: hyperscaler CDS versus the IG index

• "One of the charts I look at every morning is the credit default swaps for the big guys, Microsoft and CoreWeave and all the rest of them, Nvidia, to see what they're doing. What's the cost to insuring that debt?"
• The reading is already amber: "when you've got some of the biggest, most cash generative companies in the world that have credit default swaps that are trading at premiums to the North American investment grade credit default swaps, that tells you something."
• Why he weights credit above equity: payoff asymmetry. "If Google goes from a hundred to a billion dollars per share, the credit investor still only gets whatever they were lending out money to Google at. So they have to feel darn sure that they're going to get paid back." Equity and venture can carry nine zeros out of ten; "that's not the way it's going to work for a credit investor."
• The precedent is explicit: "credit default swaps were something that everybody was looking at during the global financial crisis to see, okay, which companies are potentially in trouble."

43:25 Dot-com was funded by free cash flow; this one is debt and circular deals

• Confirmation of the funding shift, and his answer to whether it worries him: "Absolutely."
• Current levels: CDS for these companies are "at the highest levels they've ever been" though "still at very low levels relatively" — but "even Nvidia, I think when I looked this morning, their credit default swaps were actually higher than the average North American investment grade credit default swap for 5 years."
• And the cost of being early: in 1999 "the NASDAQ was up 86%," then "at the beginning of 2000, NASDAQ was up another 24%." Knowing the bubble breaks within 18 months would still have cost you that. "You can know we're in a bubble but still have a lot of opportunity to make money before that bubble breaks."
• Which is precisely why the work shifts to selection: "that's what makes stock picking become much more important at this point."

45:54 Process — start from downside protection, then flex the short book

• The premise: "stocks can go up faster than you ever imagined possible and then… they can go down more than you ever imagined humanly possible. And so I don't want to lose money. If you go down 100%, doesn't matter if you're up a billion percent, you still are wiped out."
• Evidence he is not theorising: "when I downgraded everything I covered when I was a sellside analyst at Lehman Brothers in 2000, I would have never in a million years thought NASDAQ would be going down until October of 2002."
• Position sizing is the first decision, not the last: "if I think things are massively oversold, I may have almost close to no shorts. And then if I think we're in a period of time like we are now… I may have more shorts on than longs."
• Longs and shorts come off one research process — "as you're looking for longs sometimes I'll run into names and I'll go, well I thought this was a long but actually it's probably better off as a short because it is doing AI but it's actually probably one of the guys that's going to get displaced."
• And the layer most fundamental investors skip: "what's the big picture? What's the credit picture? If the cost of money is getting more expensive and you're having to fund a data center buildout with high-risk loans and your credit default swaps are going up, that's an issue."

50:21 Don't fight the Fed — and he reads Warsh as having pre-announced a hike

• The rule first: "my simple rule is don't fight the Fed, starts there" — also the title of his Friday note on Warsh's speech.
• The evidence: "Kevin Warsh, the chairman of the Federal Reserve, came out at Jackson Hole… and said we've had 65 months of sustained elevated inflation and that sits squarely with the central bank. So in my mind he pretty much told me that he's going to raise rates."
• The timing is deduced from the calendar, not the data: the meetings are September 16th and October 28th, the midterms November 3rd, and "there's no way in my mind that he's going to want to irk the White House by raising rates… less than a week before the election. So I firmly believe… there's a rate hike coming on September 16th. They'll probably pause on October 28th."
• The consensus he is fading is a political one: "a lot of people are still saying… 'Warsh got appointed by Trump because he wanted to lower rates, so there's no chance he's going to raise rates.' And… I'm not in that camp."
• His justification for the hike is distributional: over five years of above-target inflation is "killing" the "40-some% of the population that… isn't owning stocks and that don't own their own homes" — and that is "how you end up with people saying, 'Yeah, big business is terrible.'"

53:18 September, midterms and Iran — stacking the odds against the tape

• Seasonality: "since 1957, which is the inception of the modern S&P 500, September is the only month of the entire year that on average is down. And it's the only month… that the probability of it being down is more than the probability of it being up."
• The midterm overlay, from his own work: "peak to trough losses… are 10% from July 31st through November 9th since 1990… The normal is five in non-midterm election years."
• Geopolitics on the same clock: he expects Iranian "flare-ups right up until the midterms," reasoning from the 1980 hostage release timed to Reagan's swearing-in — "if you end up with a change in the political climate, that's much better for Iran."
• The method is explicitly cumulative, not single-factor: "it's not the Fed, it's not oil prices, it's not Iran, it's not September seasonality, it's not data centers… It's not any one of those things. But you put them all together with the valuations where they are." The poker framing: "when you know you're going to lose seven out of eight hands, it becomes very crucial on focusing on how much do you lose on the seven hands."

55:52 Closing — "strong conviction but loosely held," and no leverage

• The contrarian belief: "saying there's some stocks you just need to buy and hold is completely moronic, because you don't know that. You have a lot of survivability bias." The survivors get the podcasts; "you never have people on that say AOL was a buy and hold, Yahoo was a buy and hold, Nokia was a buy and hold, Cisco was a buy and hold" — or IBM. Microsoft, great "through three different decades," is "one company."
• He names his own live counter-example rather than a comfortable one: "one of my big disasters this year… is Nike. It used to be a market share leader… They had years of mismanagement" — with Disney as the companion ("you've got to put it away for your grandkids. Well, that hasn't worked out so well").
• The one lesson for the average investor: "I don't think there's enough emphasis placed on downside protection… you don't know if you're holding the next Google or… the next Yahoo… the next Apple or… the next Nokia… Snapchat or… Facebook."
• And the specific prohibition: "that also means don't play with leverage, because if you get it wrong, as millions of retail accounts in Korea figured out — that could not only go on margin but then invest in levered ETFs on margin — bad things can happen and you never want to put your family at risk." He closes on Goldman's "be greedy long term" and Einstein on compounding.