The AI Bubble Is Real — But Only in One Layer
There is no single AI bubble — just a re-pricing of the exposed pure-model labs, and plenty of bell-ringers with an OpenAI IPO to price.
Executive summary
There is no such thing as "the AI bubble." The phrase does the thinking for you, and it thinks wrong. What is actually being re-priced in late 2026 is one thin, exposed layer — the pure-model frontier labs, the companies whose only product is a model — while the compounding, full-stack, revenue-real core beneath them looks far sturdier than the panic implies. Treat AI as a single balloon waiting to pop and you will sell the wrong thing or buy the wrong thing.
This is a re-pricing, not a reckoning. The leaders everyone is nervous about are nothing like the dot-com casualties the comparison keeps reaching for. OpenAI is running at roughly a $40 billion annualised revenue rate as of August 2026, up from about $25 billion in March; Anthropic has climbed from a $47 billion run-rate in May toward an estimated $100 billion-plus by year-end. Pets.com had a sock puppet and no revenue. These companies have close to a billion weekly users and contracts that compound. That does not make them cheap — OpenAI is still on track to lose about $14 billion this year — but it makes "it is 1999 all over again" a bad map for the ground you are standing on.
And notice who is ringing the bell, and when. Much of the loudest bubble talk is landing at the exact moment two of these labs try to walk into public markets: Anthropic toward a roughly $2 trillion listing that investors expect in October, and OpenAI toward a listing valued as high as $1 trillion that it has now pushed to 2027 while it bleeds executives and fights a trust problem in open court. A world convinced that all of AI is a bubble prices those two floats very differently from a world that sees the risk concentrated in exactly the layer they occupy. Understanding the bubble is useful. Calling its top is a different, worse bet — and this is firmly about the first.
To understand this bubble you have to do the one thing the word "bubble" prevents: take AI apart into its layers and ask which layer you are actually exposed to. That question — not "when does it pop" — is the one that matters for anyone trying to make sense of where the AI revolution goes from here, because the answer decides whether a correction is a healthy re-pricing of a few over-valued specialists or a system-wide unwind that takes the whole build-out down with it. The rest of this is that dissection: where the dot-com analogy earns its keep and where it misleads, which layer is fragile and which is fortified, who benefits from the fear, and the honest case for why the optimists could still be wrong.
The dot-com map fits the mood, not the money
Reach for 1999 and you get the vibes exactly right. Retail euphoria, a capex frenzy, "this changes everything" on every slide, a handful of eventual winners and a coming graveyard of losers — the pattern-match is real, and the people making it are not cranks. OpenAI's own Sam Altman said in August that investors as a whole are "overexcited about AI" and that a bubble is a genuine possibility. Ray Dalio, writing in early August, put the AI boom at roughly 80% of the way to the euphoria that preceded the 1929 crash and the 2000 dot-com peak. A draft report circulating inside the US Treasury war-games a dot-com-style unwind. JPMorgan's technical desk says the tape shows the same pre-2000 divergence. When the loudest optimist and the most decorated pessimists agree the mood rhymes with 1999, the mood rhymes with 1999.
Where the analogy breaks is the money — and, underneath the money, the compounding. The dot-coms sold a story with no revenue and no moat beneath it; by most counts only around 14% of them were even profitable at their peak. The measuring stick makes the gap concrete. Cisco was the Nvidia of that cycle, the pick-and-shovel seller everyone had to buy from, and at the March 2000 top it traded at a forward price-to-earnings multiple of about 131 — on a company whose market value had swelled past half a trillion dollars. Nvidia today trades around 45 times forward earnings while posting a net margin north of 50% on more than $215 billion of trailing revenue. That is expensive. It is not delusional in the way 131-times-earnings-on-hope was delusional.
The lesson of the dot-com map is not "relax." It is "read the map for the money, not the mood." The mood tells you sentiment is stretched; only the money tells you whether the thing underneath can go to zero. And on the money, the two cycles diverge hard.
Why the core compounds
Here is the difference the analogy misses entirely, and it is the whole game. A dot-com was a website in front of a business that mostly did not exist. Today's leaders sell a product that close to a billion people use every week and that a growing share of them pay for — ChatGPT crossed a billion weekly users in August 2026 — and the technology underneath it compounds. Each model generation makes the next one cheaper to train and more useful to buy; the capability does not plateau politely between releases, it climbs. That is the thread we pulled in our compounding episode, number 45, a few weeks back: AI improvement is not a straight line you can extrapolate and short, it is a curve that bends from within.
Compounding shows up in the revenue, which is the part a bubble-caller has to explain away. OpenAI's annualised run-rate roughly doubled inside the year, from about $25 billion, where it sat from February, to around $40 billion by August 2026 — a business adding the equivalent of a large public software company's entire revenue in months. Anthropic's did something even steeper. A thing that compounds and earns can absolutely still be over-priced against its earnings. It is much harder to make it worth nothing.
The compounding is not a slogan; it is a cost curve. The price of running a fixed unit of intelligence — a million tokens through a model of a given capability — has fallen by more than an order of magnitude in two years, as better architectures, cheaper inference and open-weight competition all push the same direction. That collapse is what turns a research demo into a business: falling unit costs let a lab serve a billion weekly users without the bill scaling one-for-one, and let capability that cost a fortune last year ship inside a cheap product this year. A dot-com's economics got worse as it grew, because it was subsidising eyeballs that never paid. A frontier model's economics, on the serving side, get better as the technology matures — which is the opposite of a Ponzi, even when the valuation on top is frothy.
None of that says the price is right — OpenAI is still burning roughly $14 billion a year to earn that revenue, and we will come back to how it funds the gap. It says the failure mode is different. Dot-coms failed by evaporating; there was nothing under the ticker. If today's leaders fail, they fail by being worth less than their hype, not worth nothing — which is a correction, not an extinction. So the useful question stops being "is AI a bubble" and becomes "which parts of it are priced for a correction, and which for extinction." To answer that, you have to stop treating AI as one asset.
AI is not one asset — segment by exposure
This is the move almost everyone skips, and it is where the whole argument turns. "AI" is not a single thing you can be long or short. Split it by who owns the stack and the risk stops being uniform — it clusters, hard, in one place.
The exposed layer is the pure-model lab: a company whose entire product is a frontier model. It rents its compute from someone else's cloud, reaches customers through someone else's channels, and prices its output against open-weight models that are very nearly free. It has real revenue and real losses, and it funds the gap with outside cash on terms that assume the growth never stops. That is a genuinely fragile place to defend a valuation with eleven zeros in it — OpenAI and, to a large degree, Anthropic live here.
The full-stack platform is a different animal entirely. Google, Microsoft, Amazon and Meta own the chips or the cloud, the data, the distribution and the products the AI rides inside. For them a model is a feature of Search, or Office, or AWS, or the feed — not the whole company. When frontier model prices fall, the specialist loses pricing power and the platform gets a cheaper input for a product it already sells at scale. The commoditisation that threatens the exposed layer is a tailwind for the fortified one.
So the "bubble," to the extent there is one, is concentrated risk in one layer, not a system-wide con. That single distinction changes what a correction would mean: a re-pricing of the pure-model labs is survivable and probably overdue; a collapse of the full stack would require the whole AI thesis to be wrong, which is a much bigger and much less supported claim. Keep the two apart and most of the doom discourse resolves into a specific, tractable worry about a specific, small set of companies.
The exposed layer up close
To see why the pure-model layer is fragile, follow its unit economics. A frontier lab's cost of goods is compute it does not own, bought from a cloud it does not control, to serve a model whose closest substitute keeps getting cheaper and, increasingly, free to download and run yourself. There is no distribution moat — customers reach the model through an API anyone can swap out in an afternoon — and there is no data moat that a well-funded rival cannot rebuild. The only durable asset is being a few months ahead on capability, and the lead time keeps shrinking. Two years ago the gap between the best closed model and the best open one was measured in a year or more; by mid-2026 it is measured in months, sometimes weeks, and the open side is the one accelerating. A moat that is time is a moat that evaporates on a schedule you do not control.
Price is where the squeeze becomes visible. A closed frontier model sells its intelligence by the token at a premium that assumes scarcity. But the scarcity is going away. DeepSeek's V4-Pro, an open-weight model you can run yourself, reached general availability in August 2026 at roughly $0.44 per million input tokens after a 50% price cut — and it is one of several. When a capable, openly licensed model costs a fraction of the closed frontier and can be self-hosted, the closed lab cannot hold a premium without a capability gap wide enough to justify it. That gap is the entire business, and it is being competed away.
The implication is uncomfortable for the exposed layer and only for the exposed layer: the thing being commoditised is the exact thing a pure-model lab sells, and the thing doing the commoditising sits, increasingly, in China.
The squeeze from below: China's open-weight wave
The pressure on frontier pricing is not abstract or future — it arrived this summer, in a rush, from a set of Chinese labs releasing capable models with open weights faster than the Western frontier can cut prices. Inside an eight-week window, three flagship open-weight models shipped, and none of them came from OpenAI, Anthropic or Google. Zhipu AI, operating internationally as Z.ai, shipped GLM-5.2 in mid-June. Moonshot released Kimi K3 in July, a 2.8-trillion-parameter model it billed as the world's largest open-source model, and which now tops the independent rankings of Chinese systems. Alibaba followed in early August with Qwen3.8-Max, its first "Max-class" model released with open weights. DeepSeek's V4-Pro reached general availability days later at cut-rate pricing.
For a full-stack platform, cheap and capable open weights are a gift: a lower-cost engine to drop inside a product it already owns and distributes. For a pure-model lab charging a premium for closed intelligence, they are the tide coming in. There is a strategic edge to it, too. Releasing a frontier model with open weights is not charity; it is a way to deny a rival the pricing power its valuation depends on. If your competitor's whole business is selling closed intelligence at a premium, giving away something nearly as good for free is a direct attack on that premium — and China's labs have every incentive to keep pressing it. This is the mechanism of the re-pricing — not sentiment, not a Treasury memo, but a real and accelerating collapse in the price of the exact commodity one layer of the market depends on selling dear. Which is exactly why the layer above it looks so much safer.
The full-stack fortress
Watch what "owning the stack" buys you in practice. Google is the cleanest example: it designs its own AI chips, runs its own global cloud, holds one of the deepest proprietary data estates on earth, and distributes through Search, Android, Chrome, YouTube and Workspace — surfaces with billions of users it does not have to acquire. Its Gemini 3.1 Pro led 13 of 16 major benchmarks at its February 2026 launch, and Gemini reached hundreds of millions of monthly users largely by being wired into products people already open. That every-layer positioning is not an accident of the moment; as we covered in our episode on Google's founders, number 51, it was the bet the company was built on decades ago.
Microsoft shows the distribution moat from the revenue side: Microsoft 365 Copilot passed 20 million paid enterprise seats by April 2026, with a majority of the Fortune 500 running large deployments — seats sold into an installed base it already billed for email and documents. Amazon monetises the boom as the landlord, renting the compute the whole industry runs on. Meta folds AI into a feed that already prints cash. For each of them the model is a component, not the company, and the capex is (still, mostly) funded out of a profitable core rather than a fragile outside cash loop.
The fortress is not invulnerable — its valuations still assume enormous AI returns, and its capex is climbing faster than its cash flow, which we will get to. But its downside is a haircut on a profitable business, not a zero. So if the exposed layer is where the fragility lives, the interesting question becomes: why is the fear being broadcast as if the whole edifice were about to fall? Who gains from that framing, and why has it gone loudest right now?
Who is ringing the bell — and what they gain
Start with the people saying "bubble" loudest and ask what each of them gets from the word. Not because they are lying — several are making a fair point — but because a bubble narrative is never neutral, and in a market this concentrated the framing itself moves money. The bell-ringers fall into distinct camps with distinct incentives.
Read the board and a pattern appears. The macro investors gain from caution, the regulators are paid to warn, the researchers found real friction that is easily inflated into doom, and the traders profit from the swing regardless of direction. All of that can be true and useful — the mood really is stretched — without adding up to "the whole thing is about to collapse." But there is one camp whose incentive is sharper and more specific than any of these, and it is the one that explains the timing. The bell got loudest exactly as the most exposed lab in the market tried to sell itself to the public.
Why now: the OpenAI IPO and the trust problem
Timing is the tell. The bubble chorus swelled in the same months OpenAI moved toward the public markets — it filed a confidential draft registration in June 2026, was reported to be eyeing a listing valued as high as $1 trillion, and has since told staff, through CFO Sarah Friar, that it now expects to go public in 2027 rather than 2026 unless growth accelerates. An IPO is, above everything, a story-selling exercise. And OpenAI's story has developed cracks that a bubble narrative presses on precisely where it is weakest.
The first crack is trust, and it is playing out in a courtroom. The Musk v. OpenAI trial opened in Oakland in late April 2026, narrowed to claims of breach of charitable trust and unjust enrichment over the company's turn from non-profit to for-profit. Musk's lawyer opened his cross-examination of Sam Altman with a single question — "Are you completely trustworthy?" — and current and former insiders testified about their qualms with him, including a described pattern of resistance to board oversight and of being less than candid with senior leadership. That is the reputational fault line we mapped in our Sam Altman episode, number 28, last month: the persuasion that built the most important company in AI is also the thing its own board flagged as a risk. A trust problem is survivable for a private lab, where a handful of aligned investors can look past it. It is expensive for one asking thousands of public shareholders — and the analysts who advise them — to underwrite a trillion-dollar valuation, because public markets price governance risk explicitly, in the discount they demand and the questions they ask before they will buy.
Here is the honest core of the "who's ringing the bell" question. A world that believes all of AI is a bubble prices OpenAI's float very differently from a world that sees the risk sitting in exactly OpenAI's layer — and the second framing is both more accurate and much worse for OpenAI specifically. The narrative and the biggest, shakiest pending listing are not unrelated. Some of the noise is genuine analysis; some of it is the sound of a story being contested at the worst possible moment for the company that most needs it to go well.
The exec exodus
If the trust problem is the visible crack, the executive exodus is the structural one, and it is unusually large. Business Insider tallied 13 senior departures from OpenAI in 2026 — the kind of turnover that is hard to wave away as normal churn when a company is preparing to face public-market scrutiny. The names are not junior.
Some of this is a normal pre-IPO reshuffle — companies do clean house and pull founders like Greg Brockman back into operational roles before a listing. But the sheer count, across operations, revenue, safety and infrastructure at once, reads to the market as instability, and instability is exactly what a bubble narrative feeds on. Every senior exit is a data point the bell-ringers can point to, and a reason for public investors to demand a discount. There is also a knowledge problem hiding in the turnover. When the people who built the revenue engine, ran the safety programme and planned the data-centre footprint all leave inside a single year, the institutional memory that a prospectus is supposed to certify walks out with them — and the replacements, however capable, have to relearn the business in public. The exodus does not prove OpenAI is over-valued. It makes it much harder for OpenAI to argue that it is not.
Two IPOs, two stories
The cleanest way to see that the market is re-pricing exposure rather than condemning AI wholesale is to put the two big pending listings side by side. They occupy the same layer, and they are being received in opposite ways — which is the whole thesis in miniature.
The two are not equally fragile, and that is the point. Anthropic's investors are talking about a roughly $2 trillion valuation that would make it the largest IPO in history, past even SpaceX's $1.77 trillion float in June — and, tellingly, that figure is coming from investors, not from Anthropic executives, who have not set a public target. OpenAI, in the same layer, with comparable revenue and a far larger user base, is the one delaying and absorbing the trust hit. If this were a monolithic AI bubble, the market would be discounting both. It is not. It is discriminating — rewarding the lab with the cleaner story and pressuring the one whose story is contested. That is a re-pricing of exposure and credibility, not a verdict on the technology. Both are still pure-model labs, though, which means both still carry the layer's underlying fragility — and that fragility runs through a piece of financial plumbing worth understanding before you conclude the optimists have won.
The honest counter-case
The bullish case here is strong, but it is not a free pass, and pretending otherwise would be its own kind of hype. There are three real reasons the pain might not stay politely inside one layer.
The first is circular financing. A striking share of the AI build-out is funded by its own suppliers investing in their own customers. Nvidia has committed around $100 billion to OpenAI; OpenAI signs enormous cloud contracts, including a roughly $300 billion commitment with Oracle; Oracle spends that revenue on Nvidia chips. Cash leaves Nvidia's balance sheet as an "investment" and returns to its income statement as "revenue," having passed through the customers in between — and analysts now count more than $800 billion of such arrangements across the industry. It flatters everyone's growth numbers, and it means a stumble at the exposed layer would not stay there: it would ripple straight back into the revenue of the chipmaker the entire market treats as bedrock.
The second is concentration. The top ten stocks now make up roughly 35% of the S&P 500 — higher than the 25% they reached at the dot-com peak — so even a re-pricing confined to a few AI names would drag the whole index, and everyone's retirement account, with it. A concentrated market does not need a broad collapse to hurt; it needs a few giants to wobble. And because those giants are held by nearly every index fund, the exposure is not confined to people who chose to bet on AI — it reaches anyone with a passive pension, which is most people. That is how a re-pricing that is, in company terms, narrow becomes, in economic terms, wide.
The third is the cash-flow gap. Combined hyperscaler capital spending is running past $600 billion in 2026, up more than a third on the prior year, and for the first time it is outpacing the cash these companies generate — so they are turning to debt, with well over $100 billion raised recently and far more projected. The "funded from a profitable core" defence of the full stack is true today and getting thinner by the quarter. If AI demand merely disappoints rather than collapses, that debt-funded capex is where the strain shows up first, and it sits inside the fortress, not the exposed layer. The counter-case does not overturn the thesis. It sharpens it: the fragility is concentrated, but the plumbing that connects the layers means "concentrated" is not the same as "contained."
Bottom line
Hold two things at once, because both are true. This is not the dot-com bubble reborn, because the core compounds and earns — a correction here is a re-pricing, not an evaporation. And it is genuinely fragile where a company's only asset is a model that everyone else, increasingly in China, is learning to make for nearly free. The market is not condemning AI; it is discriminating within it, rewarding the labs and platforms with defensible stories and pressuring the ones without — which is why Anthropic can chase a record float in the same season OpenAI delays and bleeds executives.
So do not ask "when does AI pop." That question flatters the doom and answers nothing. Ask "which layer am I actually exposed to," and understand that the answer is not the same for Google as it is for a pure-model lab with an IPO to sell and a trust problem to manage. The bell-ringers are mostly right about the mood and mostly wrong about the unit of analysis. The honest hinges to watch are the ones that would let the fragility escape its layer: the circular-financing loop unwinding, the concentration snapping back, or the debt-funded capex meeting demand that merely disappoints. Watch those, and you will understand the bubble. Do not try to call it — that is a different, worse bet, and it is not the one worth making.
Sources
- Anthropic targets a ~$2T IPO valuation in October 2026 — investors (not execs) eye a record October float; last raised at a $965B valuation; as of Aug 2026.
- OpenAI's confidential filing, ~$1T target and 2027 timeline — CFO Sarah Friar signals a 2027 listing after a June 2026 confidential filing; as of Aug 2026.
- OpenAI revenue run-rate ~$40B and ~$14B 2026 loss — run-rate ~$25B (Mar) to ~$40B (Aug 2026); deep losses; Anthropic's steeper climb.
- ChatGPT passes ~1 billion weekly users — ~900M weekly (Feb) rising past 1B (Aug 2026); the usage the dot-coms never had.
- OpenAI's 2026 executive exodus — COO Brad Lightcap, CRO Denise Dresser, the No. 2 Fidji Simo and a top data-centre exec among a tallied 13 exits.
- DeepSeek V4-Pro pricing and China's price war — open-weight V4-Pro GA at ~$0.44/1M input after a 50% cut, Aug 2026.
- China's summer of open-weight frontier models — GLM-5.2, Kimi K3 (2.8T params), Qwen3.8-Max shipped in an eight-week window; none Western.
- Hyperscaler 2026 capex tops $600B, funded increasingly by debt — >$600B, ~36% higher than 2025, now outpacing free cash flow.
- AI's circular-financing loop — Nvidia's ~$100B into OpenAI, Oracle's ~$300B commitment, and >$800B of round-tripping industry-wide.
- The bubble's bell-ringers — a Treasury draft, Ray Dalio's ~80% call, Altman's "overexcited," the MIT 95% pilots study and JPMorgan's tape.
- Dot-com vs today, by the numbers — Cisco's ~131× peak multiple vs Nvidia's ~45×; the S&P's top-10 concentration at ~35% vs ~25% in 2000.
- Sam Altman testifies in the Musk trial — "Are you completely trustworthy?"; insiders testify to oversight and candour concerns; trial opened April 2026.
Transcript
Sam: Everyone keeps saying it like it's one thing. "The AI bubble." Like there's a single balloon, and one day it pops.
Alex: And that one little word — bubble — is quietly making you dumber about the most important market on earth. Because there isn't one bubble. There's a re-pricing, and it's happening in exactly one thin layer.
Sam: One layer. Not the whole thing.
Alex: One layer. And if you can't tell that layer apart from the rest, you'll sell the wrong thing, or you'll buy the wrong thing, right at the worst possible moment.
Sam: Okay. So the bubble is real — but only in one layer.
Alex: That's the whole episode.
Sam: Welcome back to Dan's AI Intel — the show where we try to make honest sense of the fastest, most consequential shift most of us are ever going to live through. I'm Sam.
Alex: And I'm Alex. And today we are walking straight into the scariest word in tech right now: bubble. It's on every front page, it's in every group chat, and there's a very good reason it's getting louder this exact month.
Sam: So here's what actually kicked this off, Alex. Late 2026, and suddenly everyone from the loudest AI optimist to the most decorated market doom-sayers is using the same word. A bubble. At the same time, two of the biggest AI labs on the planet are trying to walk out onto the public markets and sell shares to the rest of us.
Alex: And that collision is the thing I couldn't stop pulling on. Because the interesting question isn't the one everyone's asking — it isn't "when does it pop." The question underneath is much better: is "the AI bubble" even one thing? And once you take it apart — really take it apart, layer by layer — which layer are you actually standing on?
Sam: So where are we going today?
Alex: We're going to hold the 1999 dot-com map up against today and see exactly where it fits and where it lies to you. We're going to split "AI" into its real layers and find the one where the danger actually lives. We're going to look at a wave of models coming out of China this summer that is squeezing that layer from below. And then — this is the part I find almost funny — we're going to ask who is ringing the bubble bell the loudest right now, and what, very specifically, each of them gets out of you being scared.
Sam: And I'm told there's a counter-case. A reason the optimists might still be wrong.
Alex: There is, and it's an honest one — I'm not going to pretend it away. There's a piece of financial plumbing connecting these layers that could let the danger escape the one room it's supposed to stay in. We'll get there.
Sam: Before we dig in — if you're the kind of person who wants this stuff explained properly, do one tiny thing: hit follow on the show, wherever you're listening. It's free, and it means the next one just shows up for you.
Alex: So let's start with the map everyone's reaching for. 1999.
Sam: The dot-com comparison is everywhere. And honestly? On the vibes, it feels bang on. The euphoria, the "this changes everything" on every slide, everyone's cousin suddenly an investor. Is that comparison just lazy, or is there something to it?
Alex: No, that's the thing — it's not lazy. The mood really does rhyme with 1999. And the people saying so are not cranks. Here's what got my attention. Sam Altman — who runs OpenAI, the single most hyped company in this whole story — said in August that investors as a whole are, his word, "overexcited" about AI, and that yes, a bubble is a genuine possibility.
Sam: Wait, the guy whose entire company depends on the hype is saying it might be a bubble?
Alex: The chief optimist himself. And on the other side you've got Ray Dalio — legendary macro investor — writing in early August that the AI boom is about eighty percent of the way to the kind of euphoria we saw right before the 1929 crash and the 2000 dot-com peak.
Sam: Eighty percent of the way to nineteen twenty-nine. That's not a subtle number.
Alex: It's not. And it gets thicker. There's a draft report circulating inside the US Treasury that literally war-games a dot-com-style unwind. JPMorgan's technical desk says the tape — the actual price charts — shows the same divergence you saw right before 2000. So when the loudest optimist and the most decorated pessimists all agree the mood feels like 1999…
Sam: …then the mood feels like 1999. Fine. So the comparison holds.
Alex: The mood holds. The money does not. And that's the whole trick. Because 1999 tells you two totally different stories depending on whether you look at the sentiment or the balance sheets. On sentiment — twins. On money — they're not even the same species.
Sam: Okay, make that concrete for me, because "the money's different" is exactly the kind of thing a bull always says right before the crash.
Alex: Fair. So let me give you the cleanest number in the whole episode. Back in the dot-com era, the pick-and-shovel company — the one everyone had to buy from, the one selling the gear the whole boom ran on — was Cisco. Cisco was the Nvidia of that cycle. At the very top, March 2000, Cisco traded at a forward price-to-earnings multiple of about a hundred and thirty-one.
Sam: A hundred and thirty-one. And just so I'm clear — that multiple, that's basically how many years of profit you're paying up front for the stock.
Alex: Exactly right. A hundred and thirty-one times earnings means you're paying a hundred and thirty-one dollars for every one dollar a year the company actually makes. That's a price built almost entirely on hope. Now — Nvidia today, the pick-and-shovel seller of this cycle. Trades around forty-five times forward earnings.
Sam: So a third of Cisco's.
Alex: About a third. But here's the part that matters more than the ratio. Cisco's hundred-and-thirty-one was riding on a thin, fragile profit. Nvidia's forty-five is riding on a net margin north of fifty percent — on more than two hundred and fifteen billion dollars of revenue.
Sam: Hang on. Fifty percent net margin means for every dollar that comes in the door, more than fifty cents drops to the bottom line as actual profit?
Alex: More than fifty cents. On two hundred billion-plus. That is one of the most profitable large businesses that has ever existed. So is Nvidia expensive? Absolutely. Forty-five times earnings is not cheap. But it is not delusional the way a hundred-and-thirty-one-times-earnings-on-hope was delusional. Expensive and delusional are different words, and the whole bubble debate blurs them into one.
Sam: So the takeaway isn't "relax, everything's fine."
Alex: No — and I want to be careful here, because that's the lazy bull version. The takeaway is: read the map for the money, not the mood. The mood tells you sentiment is stretched. Only the money tells you whether the thing underneath can actually go to zero. And on the money, these two cycles pull apart hard.
Sam: So you keep hinting the money's different in some deep way. What's the actual difference? Because plenty of dot-coms had revenue too.
Alex: They had revenue. They didn't have this. Here's the difference the analogy misses completely, and it's genuinely the whole game. A dot-com was a website sitting in front of a business that mostly didn't exist. Pets.com — the famous one — had a sock puppet mascot and basically no revenue underneath it.
Sam: The sock puppet. I remember the sock puppet.
Alex: Everyone remembers the sock puppet and nobody remembers the profit, because there wasn't one. Now compare that to today's leaders. They sell a product that close to a billion people use every single week. ChatGPT crossed a billion weekly users in August 2026.
Sam: A billion. Weekly.
Alex: Weekly. And a growing share of them pay for it. But here's the part I actually want to plant in your head, because it's the mechanism under everything else. The technology itself compounds.
Sam: Compounds how? Give me the picture.
Alex: Okay. Think about interest in a savings account versus a straight line. A straight line just adds the same amount every year. Compounding bends upward — each year's gain sits on top of the last one and the curve steepens. AI is doing that. Each model generation makes the next one cheaper to train and more useful to sell. The capability doesn't politely flatten out between releases — it climbs, and the climb feeds the next climb.
Sam: And that's the thread you pulled on a couple weeks back, right?
Alex: It is — that was our compounding episode, number 45, from a couple of weeks ago. And the one-line version is: AI improvement is not a straight line you can just extend on a chart and bet against. It's a curve that bends from the inside. And that changes what "bubble" even means here.
Sam: So how does the compounding show up in the numbers — the part a bubble-caller actually has to explain away?
Alex: In the revenue, and it's dramatic. OpenAI's annualised revenue run-rate roughly doubled inside one year. It sat at about twenty-five billion dollars back in the spring, and by August 2026 it's around forty billion.
Sam: From twenty-five to forty. In months.
Alex: In months. That's a company adding the entire revenue of a large public software firm — in a matter of months. And Anthropic's climb was even steeper. So could a company that's growing and earning like that still be over-priced against its earnings? Yes, completely. But it is a lot harder to make a thing like that worth nothing.
Sam: And that's the distinction — over-priced versus worthless.
Alex: That's the distinction, and let me nail the mechanism so it sticks, because this is the counterintuitive bit. The compounding isn't a slogan, it's a cost curve. The price of running a fixed unit of intelligence — say, a million words of text pushed through a model of a given quality — has fallen by more than ten-fold in two years. More than an order of magnitude.
Sam: So the exact same amount of "thinking" costs a tenth of what it did two years ago.
Alex: Less than a tenth. And here's why that's the opposite of a Ponzi scheme. A dot-com's economics got worse as it grew, because it was subsidising eyeballs that never paid — every new user cost it money. A frontier model, on the serving side, gets better as the technology matures. Falling costs let a lab serve a billion people a week without the bill scaling one-for-one. The thing gets more efficient underneath you as it gets bigger.
Sam: And that collapsing cost — that's also what turns a lab experiment into an actual product, isn't it?
Alex: That's the quiet magic of it. The capability that cost a fortune to run last year can ship inside a cheap consumer product this year, precisely because the cost fell out from under it. So the frontier and the everyday product aren't two different things — the frontier just becomes the product about twelve months later, at a tenth of the price. That's a flywheel a dot-com never had. And it's why Anthropic's climb was even steeper than OpenAI's — it went from a forty-seven-billion-dollar run-rate in May toward an estimated hundred billion-plus by the end of the year.
Sam: Forty-seven to a hundred billion. In one year. That's not a company, that's a rocket.
Alex: It's a rocket with a real engine under it — that's the whole point. You can argue about the price of the ticket. You can't argue that there's no rocket.
Sam: But that can't be the whole story, or nobody would be nervous at all.
Alex: Right, and I'm not going to hand-wave it. OpenAI is still burning roughly fourteen billion dollars a year to earn that revenue. The price on top can absolutely be too high. But notice how the failure mode changed. A dot-com failed by evaporating — you lifted the ticker and there was nothing underneath. If today's leaders fail, they fail by being worth less than the hype. Not worth nothing.
Sam: A correction, not an extinction.
Alex: A correction, not an extinction. Which means the useful question quietly changes shape. It stops being "is AI a bubble," and becomes "which parts of AI are priced for a correction, and which are priced for extinction." And to answer that, you have to do the one thing the word "bubble" refuses to let you do.
Sam: Which is?
Alex: Stop treating AI as one single thing.
Sam: So this feels like the hinge of the whole argument. Take AI apart. What are the layers?
Alex: This is the move almost everyone skips, and it's exactly where the whole thing turns. "AI" is not one asset you can be long or short, like a single stock. Split it by who actually owns the stack, and the risk stops being spread evenly across everything. It clusters. Hard. In one place.
Sam: Okay, so what's the fragile layer?
Alex: The fragile layer is what I'd call the pure-model lab. A company whose entire product is a frontier model. That's it. That's the whole company.
Sam: And what's fragile about that? On paper it sounds like the crown jewel — you make the smartest model.
Alex: It sounds like the crown jewel. Look closer. That company rents its computing power from somebody else's cloud. It reaches its customers through somebody else's channels. And it has to price its output against open models that are very nearly free. It has real revenue and real losses, and it plugs the gap with outside cash — raised on terms that quietly assume the growth just never, ever stops. That is a genuinely nervous place to be defending a valuation with eleven zeros on the end of it. And that is where OpenAI lives, and to a large degree, Anthropic too.
Sam: Eleven zeros. That's hundreds of billions.
Alex: Hundreds of billions. Now — the other layer. The full-stack platform. Completely different animal. That's Google, Microsoft, Amazon, Meta. They own the chips, or the cloud, or the data, or the distribution — usually several of those. And crucially, for them, the model is a feature. It's a feature of Search, or Office, or the AWS cloud, or the feed. It is not the whole company.
Sam: So let me test whether I've got this. When the price of frontier models falls — which you just said is happening fast — the pure-model lab is in trouble because that's the only thing it sells. But the platform...
Alex: …the platform just got a cheaper ingredient for a product it already sells at massive scale. Same event. Falling model prices. For one layer it's the tide coming in over your ankles. For the other, it's a discount on your raw materials. The exact thing that threatens the exposed layer is a tailwind for the fortified one.
Sam: That is such a cleaner way to hold it. So "the AI bubble" isn't wrong so much as… it's the wrong unit.
Alex: It's the wrong unit of analysis. That's the sentence I'd tattoo on this whole discourse. To the extent there's a bubble, it's concentrated risk in one layer — not a system-wide con. And that single distinction totally changes what a correction would even mean. Re-pricing the pure-model labs? Survivable, and honestly probably overdue. A collapse of the full stack? That would require the entire idea of AI to be wrong — which is a much bigger, much less supported claim.
Sam: Keep the two apart, and most of the doomscrolling just… resolves.
Alex: It resolves into a specific, tractable worry about a specific, pretty small set of companies. So let's go stare right at the fragile layer and figure out why, exactly, it's so exposed.
Sam: Right. So the pure-model lab. Walk me through why it's fragile. What's actually weak about it?
Alex: Follow its unit economics — just follow the money through the business. Its cost of goods is compute it doesn't own, bought from a cloud it doesn't control, to serve a model whose closest competitor keeps getting cheaper and, increasingly, is free to just download and run yourself.
Sam: So no moat on the cost side.
Alex: And no moat anywhere else either. There's no distribution moat — customers reach the model through an interface, an API, that anyone can swap out in an afternoon. There's no data moat that a well-funded rival can't rebuild. The one durable asset the pure-model lab has is being a few months ahead on raw capability. And that lead keeps shrinking.
Sam: How much has it shrunk?
Alex: Two years ago, the gap between the best closed model and the best open one you could grab for free was measured in a year or more. By the middle of 2026 it's measured in months. Sometimes weeks. And the open side is the one accelerating.
Sam: Okay, so here's the analogy forming in my head, tell me if it's wrong. If your only asset is a head start, and the head start is shrinking…
Alex: …then your moat is time. And a moat made of time evaporates on a schedule you don't control. That's exactly it. You wake up every morning with a little less of it than you had yesterday, and there's nothing you can buy to refill it.
Sam: That's genuinely unsettling if you're the one holding the valuation. So where does that show up first — where does the squeeze become visible?
Alex: Price. A closed frontier model sells its intelligence by the token — by the word, basically — at a premium that quietly assumes it's scarce. But the scarcity is going away. Let me give you the number that makes it real. DeepSeek — a Chinese lab — put out an open-weight model called V4-Pro in August 2026. Open-weight means you can download it and run it yourself. And it came in at roughly forty-four cents per million input tokens, after a fifty percent price cut.
Sam: Forty-four cents per million words-ish. And what does the premium closed model charge?
Alex: A representative closed-frontier tier is around five dollars for the same million. So you've got roughly five dollars versus forty-four cents — and the forty-four-cent one you can host yourself, no permission needed. When a capable open model costs a fraction of the closed one and you can run it on your own machines, the closed lab simply cannot hold its premium — unless its model is so much better that the gap is worth ten times the price.
Sam: And that capability gap is the whole business.
Alex: The capability gap is the entire business. And it's being competed away, in public, month by month. Which brings us to the uncomfortable part — the thing doing the commoditising. Because it increasingly sits in one place.
Sam: Let me guess. China.
Alex: China.
Sam: So this isn't some future threat. This is now.
Alex: This is this summer. And it arrived in a rush. Inside one eight-week window, three flagship open-weight models shipped — and not one of them came from OpenAI, or Anthropic, or Google.
Sam: Give me the roll call.
Alex: Mid-June: a lab called Zhipu, which operates internationally as Z-A-I, shipped GLM-5.2 — big gains on coding and on agentic tasks, the stuff where the model actually does things for you. Then mid-July, Moonshot released Kimi K3. And Kimi K3 is a monster — two-point-eight trillion parameters, which they billed as the world's largest open-source model, and it now tops the independent rankings of Chinese systems.
Sam: Two-point-eight trillion parameters. I don't fully know what that means but it sounds enormous.
Alex: Think of parameters as the number of tunable dials inside the model's brain — more dials, roughly, more capacity to capture patterns. Two-point-eight trillion of them is at the absolute frontier of scale, and they gave it away with open weights. Then early August, Alibaba followed with Qwen3.8-Max — its first top-tier, "Max-class" model released open-weight. And then, days later, DeepSeek's V4-Pro at that cut-rate forty-four cents.
Sam: So four frontier-class open models, one summer, all Chinese.
Alex: All Chinese. And now sit with what that does to each layer, because it's not the same. For a full-stack platform, cheap and capable open weights are a gift. It's a lower-cost engine you can drop straight into a product you already own and distribute. But for a pure-model lab charging a premium for closed intelligence — it's the tide coming in.
Sam: Okay but why would China just… give these away? That's the part that confuses me. That's expensive to build and they're releasing it for free.
Alex: Great question, and it's not charity. Think about it strategically. If your competitor's whole business is selling closed intelligence at a premium, what's the most damaging thing you can do to them?
Sam: Give away something almost as good for free.
Alex: Give away something nearly as good, for free. It's a direct strike on the exact thing their valuation depends on — their pricing power. Releasing a frontier model with open weights is a way to deny a rival the premium it needs to survive. And China's labs have every incentive in the world to keep pressing on that.
Sam: So this isn't sentiment. This isn't a Treasury memo or a scary chart.
Alex: This is the actual mechanism of the re-pricing. It's a real, accelerating collapse in the price of the exact commodity that one specific layer of the market depends on selling dear. It's physics, not mood. And it is precisely why the layer sitting above it looks so much safer.
Sam: So let's go up a floor. The full-stack platforms. You keep calling them a fortress. What does "owning the stack" actually buy you when the pressure hits?
Alex: Let me make it concrete with the cleanest example, which is Google. Google designs its own AI chips. It runs its own global cloud. It sits on one of the deepest private data estates on the entire planet. And it distributes through Search, Android, Chrome, YouTube, Workspace — surfaces with billions of users it does not have to go out and buy.
Sam: Right, it already has everyone.
Alex: It already has everyone. And on capability it's not coasting either — its Gemini 3.1 Pro model led thirteen of sixteen major benchmarks at its February 2026 launch, and Gemini got to hundreds of millions of monthly users largely by being wired into products people already open every day.
Sam: So it didn't have to win an audience. It just switched the AI on inside the doors everyone was already walking through.
Alex: That's exactly the move. Now watch Microsoft, because Microsoft shows you the distribution moat from the revenue side. Its Microsoft 365 Copilot — the AI baked into Office — passed twenty million paid enterprise seats by April 2026, with a majority of the Fortune 500 running big deployments.
Sam: Twenty million paid seats. And those are companies it was already billing anyway.
Alex: Already billing them for email and documents. The AI is just another line on an invoice they already pay. Amazon plays the landlord — it rents the computing power the whole industry runs on, through AWS. And Meta folds AI into a feed that already prints money. For every one of them, the model is a component. Not the company.
Sam: So the same falling model prices we said were a disaster one floor down…
Alex: …are a discount up here. Same weather, opposite outcome. And the capex — the enormous spending on data centres — is still, mostly, funded out of a profitable core, rather than the fragile outside-cash loop the pure-model labs rely on.
Sam: I want to push on one thing though, because "fortress" makes it sound bulletproof, and nothing's bulletproof.
Alex: Good, push, because it isn't. The fortress is not invulnerable. Its valuations still assume enormous AI returns land, and — this is important, we're coming back to it hard later — its capex is now climbing faster than its cash flow. But here's the key difference in the failure mode: the downside for a Google or a Microsoft is a haircut on a hugely profitable business. It's a bad year. It is not a zero.
Sam: A haircut versus a funeral.
Alex: A haircut versus a funeral. So if that's where the fragility actually lives — down in the exposed layer — it raises this genuinely weird question. Why is the fear being broadcast as if the whole building is about to come down? Who benefits from that framing? And why has it suddenly gotten so loud right now?
Sam: Follow the incentive.
Alex: Follow the incentive.
Sam: Okay, this is my favourite part. Who's shouting "bubble" the loudest, and what does each of them get out of it?
Alex: And I want to be fair here — not everyone yelling bubble is lying. Several of them are making a genuinely fair point. But a bubble narrative is never neutral. In a market this concentrated, the framing itself moves money. So let's just walk the board, and for each one, ask: what do you gain from the word?
Sam: Start me off.
Alex: Ray Dalio — the macro investor, the one with the eighty-percent-to-1929 call. Real, serious analyst. Also talking his book, which leans bearish. Caution is good for him. Next: Altman — the incumbent — granting that yeah, a bubble's possible, people are "overexcited." That's honest. It also happens to cool down his rivals and make him look like the sober adult in the room.
Sam: Sneaky. The optimist gets to sound humble and kneecap the competition in the same breath.
Alex: Both at once. Then the US Treasury, with that draft memo war-gaming an unwind. That's the regulator — and regulators are paid to warn. That's their job; a warning costs them nothing and covers them if it goes wrong. Then there's an MIT study that found ninety-five percent of generative-AI pilots inside companies aren't yet profitable.
Sam: Ninety-five percent not profitable — okay, that one sounds genuinely bad.
Alex: It's real friction, and it's worth knowing. But notice how easily "ninety-five percent of pilots aren't profitable yet" gets inflated into "AI doesn't work." Early pilots are supposed to be unprofitable — that's what a pilot is. And last: JPMorgan's trading desk, pointing at the tape. But a trading desk makes money on the swing. Up or down, volatility is their product. They profit either way.
Sam: So read the whole board at once and what do you see?
Alex: You see a pattern. The macro investors gain from caution. The regulators are paid to warn. The researchers found real friction that's easy to inflate into doom. The traders profit from the swing regardless of direction. And all of that can be true and useful — the mood really is stretched — without it adding up to "the whole thing is about to collapse."
Sam: But you paused. There's a "but."
Alex: There's one camp whose incentive is sharper and more specific than any of those. And it's the one that actually explains the timing. Because the bell got loudest at the exact moment the most exposed lab in the entire market tried to sell itself to the public.
Sam: The timing's the tell.
Alex: The timing is the tell. The bubble chorus swelled in the very same months OpenAI started moving toward the public markets. It filed a confidential draft registration in June 2026. It was reported to be eyeing a listing valued as high as one trillion dollars. And it's since told its own staff, through its CFO Sarah Friar, that it now expects to go public in 2027 rather than 2026 — unless growth accelerates.
Sam: So they pushed it back.
Alex: They pushed it back. And here's the thing about an IPO — above everything else, it is a story-selling exercise. You are standing in front of the public asking them to believe your story enough to buy it. And OpenAI's story has developed some cracks, and a bubble narrative presses on them exactly where they're weakest.
Sam: What's the first crack?
Alex: Trust. And it's playing out in a courtroom. The Musk versus OpenAI trial opened in Oakland in late April 2026. It narrowed down to claims of breach of charitable trust and unjust enrichment — basically, over the company's turn from a non-profit into a for-profit. And Musk's lawyer opened his cross-examination of Altman with a single, brutal question: "Are you completely trustworthy?"
Sam: Oof. As an opening line.
Alex: As the opening line. And then current and former insiders testified about their own qualms with him — a described pattern of resistance to board oversight, and of being less than candid with senior leadership.
Sam: And this connects back to something we've covered.
Alex: It's the exact fault line we mapped in our Sam Altman episode — number 28 — last month. The short version: the persuasion, the sheer force of personality that built the most important company in AI, is also the thing his own board flagged as a risk. And here's why that matters so much more now than it did a year ago. A trust problem is survivable for a private lab — a handful of aligned investors can look past it. It is expensive for a company asking thousands of public shareholders, and the analysts advising them, to underwrite a trillion-dollar valuation. Because public markets price governance risk explicitly. It shows up in the discount they demand and the questions they ask before they'll buy a single share.
Sam: So say the honest version of the "who's ringing the bell" question out loud.
Alex: Here it is. A world that believes all of AI is a bubble prices OpenAI's float very differently from a world that sees the risk sitting in exactly OpenAI's layer. And the second framing — the accurate one — is much worse for OpenAI specifically. So the loud narrative and the biggest, shakiest pending listing are not unrelated. Some of that noise is genuine analysis. And some of it is the sound of a story being fought over at the worst possible moment for the one company that most needs it to go well.
Sam: You said trust was the visible crack. Is there a structural one?
Alex: There is, and it's unusually large. If trust is the crack you can see, the executive exodus is the one in the foundations. Business Insider tallied thirteen senior departures from OpenAI in 2026.
Sam: Thirteen. In one year. That's not people leaving for a better coffee machine.
Alex: And the names are not junior. Brad Lightcap, the Chief Operating Officer, who'd been there since 2018 — leaving to "start something new." Fidji Simo, who ran Applications, effectively the number two — departed. Denise Dresser, the Chief Revenue Officer — out after about nine months. The heads of safety, of ethics, the chief futurist — gone. And in August, a top data-centre executive walked too.
Sam: Operations, revenue, safety, infrastructure — that's basically the whole spine of the company leaving at once.
Alex: All four at once. Now, in fairness — some of this is a normal pre-IPO reshuffle. Companies do clean house before a listing; they pull founders like Greg Brockman back into hands-on operational roles. That's routine. But the sheer count, hitting every function simultaneously, reads to the market as instability. And instability is precisely what a bubble narrative feeds on. Every senior exit is a data point the bell-ringers can point at, and a reason for public investors to demand a discount.
Sam: And there's a subtler cost too, right? It's not just optics.
Alex: There's a knowledge problem hiding inside it, and I think it's the real one. When the people who built the revenue engine, ran the safety programme, and planned the data-centre footprint all leave inside a single year, the institutional memory that a prospectus is supposed to certify — that "here's exactly how this business works" — walks out the door with them. And the replacements, however good they are, have to relearn the business in public, under a spotlight.
Sam: So the exodus doesn't prove OpenAI's over-valued.
Alex: It doesn't prove it. It just makes it a lot harder for OpenAI to argue that it isn't. Which sets up the cleanest single piece of evidence in this whole story.
Sam: And that's the two IPOs side by side.
Alex: Put the two big pending listings right next to each other, because it's the whole thesis in miniature. Same layer — both pure-model labs. And they're being received in completely opposite ways.
Sam: Start with Anthropic.
Alex: Anthropic's investors are talking about a valuation of roughly two trillion dollars — a float that would make it the largest IPO in history. Bigger even than SpaceX's one-point-seven-seven trillion listing back in June. Its run-rate went from about forty-seven billion in May toward an estimated hundred billion-plus by year-end. Trust intact. Stable senior bench.
Sam: And the tell you flagged — where's the two-trillion number coming from?
Alex: That's the beautiful detail. The two-trillion figure is coming from investors. Not from Anthropic's executives, who haven't set a public target at all — the last round they actually raised valued them at about nine hundred and sixty-five billion. So the market is voluntarily reaching up past that, to two trillion, to hand them a number they didn't even ask for.
Sam: So the buyers are more excited than the company. That's the opposite of a hard sell.
Alex: It's the opposite of a hard sell, and it's the opposite of what you'd see in a blanket bubble. Now put OpenAI right next to it. Same layer. Comparable revenue, around forty billion. A far bigger user base, actually. And it's the one delaying — valued at up to one trillion, absorbing the trust hit in open court, thirteen executives out the door.
Sam: So if this were one monolithic AI bubble —
Alex: — the market would be discounting both of them. It's not. It is discriminating. Rewarding the lab with the cleaner story, pressuring the one whose story is being contested. That is a re-pricing of exposure and credibility. It is not a verdict on the technology. Two labs, same room, opposite receptions.
Sam: But you'll notice you said they're both still in the same room.
Alex: Both still pure-model labs. So both still carry the layer's underlying fragility. And that fragility runs through one piece of financial plumbing that I think you have to understand before you let yourself conclude the optimists have simply won.
Sam: Okay. So here's where I want you to argue against yourself. The bull case has been strong all episode. What's the honest reason it might be wrong?
Alex: And I mean it — pretending the bull case is a free pass would be its own kind of hype, and I'm not doing that. There are three real reasons the pain might not stay politely inside that one exposed layer. Three ways it escapes the room.
Sam: Give me the first.
Alex: Circular financing. And this one is genuinely wild once you see it. A striking share of the entire AI build-out is being funded by its own suppliers investing in their own customers. Follow the loop. Nvidia has committed around a hundred billion dollars to OpenAI. OpenAI then signs enormous cloud contracts — including a roughly three-hundred-billion-dollar commitment with Oracle. And Oracle spends that revenue on… Nvidia chips.
Sam: Wait. So the money leaves Nvidia, goes to OpenAI, goes to Oracle, and comes back to Nvidia?
Alex: It leaves Nvidia's balance sheet labelled "investment" and returns to its income statement labelled "revenue" — having just done a lap through the customers in the middle. And analysts now count more than eight hundred billion dollars of these arrangements across the whole industry.
Sam: Okay, give me the analogy, because that sounds like a shell game.
Alex: Picture three shops on a street who agree to keep buying from each other with the same fifty-dollar bill, passing it in a circle. Every one of them can now report "sales are up!" — but there's really only one fifty-dollar bill, going around and around. The revenue looks bigger than the actual demand underneath it. It flatters everyone's growth numbers. And critically — it means a stumble down at the exposed layer wouldn't stay there. It'd ripple straight back into the revenue of Nvidia, the one company the entire market treats as bedrock.
Sam: So the fragile layer is wired to the bedrock. That's the opposite of contained.
Alex: That's the first way it escapes. Here's the second: concentration. The top ten stocks now make up roughly thirty-five percent of the entire S&P 500. That's higher than the twenty-five percent they hit at the dot-com peak.
Sam: So the index is more top-heavy now than it was in 2000?
Alex: More top-heavy than 2000. And that means even a re-pricing confined to a few AI names would drag the whole index — and everyone's retirement account — down with it. A concentrated market doesn't need a broad collapse to hurt. It just needs a few giants to wobble. And because those giants are inside nearly every index fund, the exposure isn't limited to people who chose to bet on AI. It reaches anyone with a passive pension. Which is most people.
Sam: So a re-pricing that's narrow in company terms —
Alex: — becomes wide in economic terms. That's how "it's only one layer" still ends up in your retirement statement. And the third way is the cash-flow gap — and this one sits inside the fortress, which is what makes it sting. Remember I said the platforms fund their capex from a profitable core? That's getting thinner by the quarter. Combined hyperscaler capital spending is running past six hundred billion dollars in 2026 — up more than a third on the year before. And for the first time, it's outpacing the cash these companies actually generate.
Sam: So they're spending more than they're making — on data centres.
Alex: On the build-out. And to cover the gap, they're turning to debt — well over a hundred billion dollars raised recently, and far more projected. So the "don't worry, it's funded from profits" defence of the full stack is true today, and it is getting weaker every quarter. If AI demand merely disappoints — doesn't collapse, just underwhelms — that debt-funded capex is where the strain shows up first. And it's inside the fortress, not the exposed layer.
Sam: So none of these overturn your thesis. But they complicate it.
Alex: They don't overturn it — they sharpen it. The fragility is concentrated. But the plumbing connecting the layers means "concentrated" is not the same as "contained." And holding both of those at once is the entire skill here.
Sam: Okay. Bring it home for me. If I take one thing away, what is it?
Alex: Hold two things at once, because both are true at the same time. One: this is not the dot-com bubble reborn — because the core compounds and it earns. A correction here is a re-pricing, not an evaporation. And two: it is genuinely fragile in exactly the spot where a company's only asset is a model that everyone else — increasingly in China — is learning to make for nearly free.
Sam: And the market itself is telling us which is which.
Alex: The market is not condemning AI. It's discriminating within it — rewarding the labs and platforms with defensible stories, pressuring the ones without. That's why Anthropic can chase a record float in the very same season OpenAI delays and bleeds executives. Same layer, opposite receptions. That's not a bubble popping. That's a market getting more precise.
Sam: So the question everyone's asking —
Alex: — "when does AI pop" — is the wrong question. It flatters the doom and it answers nothing. Ask the better one: which layer am I actually exposed to? Because the answer is completely different for Google than it is for a pure-model lab with an IPO to sell and a trust problem to manage.
Sam: And if I want to actually watch this play out — what do I keep my eye on?
Alex: The honest hinges. The three ways the fragility could escape its layer. Watch whether that circular-financing loop starts to unwind. Watch whether the market concentration snaps back. And watch whether that debt-funded capex meets demand that merely disappoints. Watch those three, and you'll genuinely understand this bubble. But do not try to call its top. That's a different, and a much worse, bet — and it is not the one worth making.
Sam: The bell-ringers are mostly right about the mood.
Alex: And mostly wrong about the unit. That's the whole thing.
Sam: And that's where we'll leave it for today — thank you so much for spending this time with us. I hope you came away seeing a bit more clearly where all of this is actually heading. It is a genuinely complex, fast-moving picture, with a brutally short shelf-life on what you think you know — and honestly, that's exactly what makes it worth following this closely.
Alex: One honest note on how this show is made. It's AI-generated. Dan builds a custom stack of AI tools to research, analyse, verify and illustrate the questions worth understanding — mostly to learn them himself, and then he shares what he finds. It's AI-assisted, fact-checked, and always worth a second look. And one thing specific to today: everything we just talked through is general analysis of the AI market and the companies in it — it is not investment advice. For decisions about your own money, please talk to a licensed financial adviser.
Sam: And before you go, one genuinely useful thing you can do for us: follow the show. Whatever app you're listening in right now, there's a follow button, or a little plus — and it's one tap, and it's free.
Alex: And it does two real things. You'll get each new episode the moment it lands, so you're never scrambling to catch up. And honestly — for a small, independent show like this one, a follow is the single biggest lever there is for helping it reach other people trying to make sense of all this. So if this was worth your time today, go ahead and hit follow. That's how it grows.
Sam: One last thing, and it matters to us more than a follow. We want to make this show better, and the fastest way we do that is you. If there's something in here you'd push back on — a number you'd question, or a thread you want us to pull harder on next time — tell us. The address is podcast at connectiveshift dot com.
Alex: We read every single message. And it genuinely shapes what we go deep on next — so if you want a say in what we dig into, that's how you get it.
Sam: Until next time — keep asking the better question.
Alex: Not "when does it pop." Which layer am I standing on. See you next episode.