Innovation Has a "Historical Track" Record:
The House Always Wins
There is no shortage of people arguing that today's AI infrastructure spending is a bubble. Michael Burry questions the accounting. Ed Zitron questions the revenue. IBM's Arvind Krishna questions the economics. Yann LeCun questions whether the underlying technology can ever deliver what's being paid for. Sridhar Mahadevan is proposing a different mathematical foundation to fix AI's hallucination problem entirely. And just this week, Ludovic Subran looked at SpaceX issuing a $25 billion bond and called it bubble territory, out loud. At a summit. They disagree about exactly what's wrong, but each one is questioning a different assumption the whole boom depends on.
I don't think any of them are wrong, exactly. I think they're answering a different question than the one I'm actually interested in.
The question I keep coming back to isn't whether AI is overhyped. Of course it's overhyped — name me a transformative technology that wasn't, in its buildout phase. The real question is whether this is still a justified investment even if the hype deflates, even if the most optimistic revenue projections turn out fantasy, even if three quarters of the companies placing these bets lose every dollar they put in. I think the answer is yes, and the case isn't obvious — I don't think it should be taken on faith, mine included.
This isn't a forecast. I don't know how this plays out, and anyone who tells you they do is selling something. It's a case for a possibility I think is underweighted in the current conversation: that the bubble and the bet can both be real at the same time, without one canceling out the other.
The infrastructure underneath all of this — the chips and the data centers — is what I actually want to make the case for here, not the apps or the labs sitting on top of them. I'll do that through three lenses: whether the actual math behind the spending holds up, where the bubble everyone's worried about actually lives, and why building makes sense even for the companies that won't end up winning. That last one matters because every wave like this produces far more losers than winners, and the case for building anyway points back at a previous time it was done. The obvious choice there is railroad track laid across the country in the 19th century, which is also where this piece's title comes from. But there's a more recent example that hits harder, and that's the one I'll use instead.
Does the math actually work?
Start with the most conservative number on the table, not the most exciting one. Aswath Damodaran recently revised his estimate of the AI services market down to $3–4 trillion over the next decade — not Altman's $10–15 trillion, not the $25 trillion Musk put in SpaceX's own IPO prospectus.
| Source | Estimate | What has to be true |
|---|---|---|
| Damodaran (conservative) | $3–4T | AI stays mostly a tool; margins stay compressed by infrastructure cost |
| Altman / bull case | $10–15T | AI meaningfully augments and replaces some labor across white-collar work |
| Musk / SpaceX S-1 | $25–26T | AI substantially replaces human labor at scale |
I'm building this piece on the low end deliberately — it's the number that stress-tests the thesis rather than flatters it, and the one Damodaran himself seems most willing to defend in public. Here's what surprised me: even there, the infrastructure layer's piece of the pie looks fine.
AI services don't carry software-style margins today — every additional query still costs real electricity and real silicon, so costs don't collapse the way software distribution costs do, at least not yet (more on whether that holds when we get to the Jevons paradox below). Damodaran's own numbers imply 65 to 80 cents of every AI dollar goes to infrastructure, not profit; Dwarkesh Patel has put a finer point on it, sizing chips at roughly 70% of a data center's lifetime cost and power at 10–15% — a figure Musk has independently confirmed.
Run that against a $3 trillion ceiling and the chip-specific revenue pool, at maturity, comes to roughly $1.4 to $2 trillion a year — well above what the actual chip and memory companies (Nvidia, AMD, Micron, SK Hynix, Broadcom, Marvell) are making today.
| Horizon | Chip-specific revenue pool needed | Annual growth required to get there |
|---|---|---|
| Today (2026) | ~$557B (actual) | — |
| 2030 | ~$1.4–2T | ~25–37% |
| 2035 | ~$1.4–2T | ~10–15% |
Every major chip and memory company is currently growing faster than even the more demanding of those two numbers — Nvidia's data center revenue up 92% last quarter, AMD's 57%, the memory makers well into triple digits. That's no guarantee — sprinting growth can stop sprinting — but the bar ahead is a deceleration from where things already are, not a stretch.
One objection worth confronting head-on rather than skating past: if chips keep getting more efficient, doesn't that mean fewer chips need to get bought? Not necessarily — and there's a name for why. Economists call it the Jevons paradox: historically, making a resource more efficient to use has tended to increase total consumption of it, not shrink it, because the lower cost unlocks demand that didn't exist at the higher price. Whether that holds here too is worth unpacking properly rather than asserting.
Does the Jevons paradox actually save the chip math?
The original case is 19th-century coal: more efficient steam engines were supposed to cut Britain's coal consumption, and instead the country burned more coal than ever, because cheap, efficient engines made coal-powered everything worth doing that wasn't economical before. Computing's whole history runs on the same logic — every leap in cost-per-compute has historically grown total consumption, not shrunk it.
So far, the evidence says this is exactly what's happening in AI too, not just in theory. Inference prices have been falling sharply — the open-weight price wars discussed elsewhere in this piece are the most visible example — and over that same window, Nvidia's Data Center revenue grew 92% year-over-year. Falling prices and exploding chip revenue have been happening at the same time, which means demand has been elastic enough that cheaper compute is pulling in more than enough new buyers to offset the lower price per unit.
The honest limit on this: elastic demand is a property of an expanding market, with real unmet demand still waiting to be unlocked by lower prices. Once a market truly saturates, the same efficiency gain stops unlocking new buyers and starts just competing away existing revenue instead.
What I genuinely don't know is how far away that limit is, or whether compute demand has one in any practical sense at all. Every time the industry has more or less declared demand satisfied, a new wave has shown up needing more of it than the last one did: mainframes gave way to a "does anyone really need a computer at home" era that gave way to PCs, then laptops, then smartphones, then crypto mining, and now AI, with robotics looking like a plausible next wave after that. Compute demand has behaved less like a market with a visible ceiling and more like an appetite that keeps finding new things to want. I wouldn't bet against that pattern continuing — but I also wouldn't claim it's guaranteed to, since "it always has before" is a description of history, not a law of economics. Whether that boundary is close, far away, or simply keeps moving is really the same question as how much of the $3–4 trillion ceiling is still unclaimed, which is what the companion piece on where this revenue actually comes from is built to answer.
What if someone solves Nvidia's power problem first?
Nvidia's real weak point is power consumption per unit of compute — Cerebras's wafer-scale design needs roughly a fifth the power of an equivalent GPU cluster for the same work, a documented gap, not a hypothetical one. But solving that doesn't drain money out of the chip-specific bucket either way it plays out. If an efficient challenger grows independently, that's revenue currently undercounted here, not lost — Cerebras isn't one of the five companies in the tally above. If an efficient challenger gets absorbed by an incumbent instead — the way Nvidia paid $20 billion for Groq's assets in December 2025 — the revenue still lands inside a company already being counted, just via a different technology. Either way, this is a concentration question (which company, which stock) rather than a leak out of the aggregate. Nvidia's Vera Rubin platform, which reportedly folds Groq's acquired technology into a heterogeneous rack design, is its own attempt to close this gap before an outside challenger forces the issue — though that's a more direct power-efficiency answer than its separate CPU-bundling story, and there isn't confident public data yet on whether it actually closes the gap, so I'd hold that as unresolved rather than settled.
If you want the company-by-company numbers behind this
Nvidia's data center segment alone is running at roughly $300 billion annualized. AMD's data center business, including both EPYC server CPUs and Instinct accelerators, sits around $23 billion annualized — and it's worth noting AMD's own leadership now frames the CPU side as an AI product too, since agentic workloads are pushing toward something closer to a 1:1 ratio of CPU to GPU compute, not a separate legacy line riding alongside the AI story. Micron's cloud and AI-relevant memory segment is around $31 billion annualized, though a cross-check from SemiAnalysis's own reporting suggests the industry-wide memory number, once SK Hynix and Samsung are properly counted, is closer to $180 billion a year just from the four largest hyperscaler customers — meaningfully higher than a Micron-only estimate would suggest. Broadcom's AI-specific revenue, mostly custom silicon design fees for Google's TPU and similar programs, is annualizing around $43 billion and accelerating, and Marvell — the second-largest custom-silicon design partner, working with Amazon's Trainium and Microsoft's Maia — adds roughly another $10 billion. None of this is a clean, audited single number — it's stitched from quarterly filings, earnings calls, and one credible industry tracker's public commentary — but it triangulates to roughly the same place from several independent directions, which is the most I'd ask of any number this far out.
A few smaller names round this out without changing the picture much: Intel's Gaudi/AI accelerator line (~$2B), MediaTek's AI data-center silicon (~$2B), and Alchip's custom ASIC design business (~$1B, though down sharply from the year before) add perhaps another $5 billion combined — pushing the true total closer to $560–565 billion than $557 billion. A few more interesting smaller players are worth a specific mention since they don't change the math but might seem like they should: Cerebras (a real, audited $510 million in 2025 revenue, simply too small at that scale to move this number), Groq (no longer independent — Nvidia absorbed its assets in a $20 billion deal in December 2025, so whatever it generates now lives inside Nvidia's own figures), and Tenstorrent (private, with no disclosed revenue I could verify).
Worth a word, too, on how reliable this growth pattern has actually been — though the clearest historical precedent here is more oblique than it looks at first. Nvidia's total revenue did fall 13% year-over-year in early 2023, and Micron posted a net loss that fiscal year — but neither was an AI or data-center story. Nvidia's dip came from gaming GPU demand collapsing after the crypto-mining bust and a broader post-pandemic pullback in PC sales; Micron's loss came from that same general memory glut, not anything specific to AI. Data Center revenue at Nvidia was actually growing the entire time, driven by cloud providers, while gaming cratered. So the honest takeaway isn't "this has happened before and the business survived it" — it's that these companies' non-AI revenue lines have proven genuinely volatile, while the AI-specific revenue itself hasn't yet been tested by a real downturn of its own. That's a real unknown, not a resolved one.
Worth being precise about what kind of bet this actually is, rather than letting it sound more settled than it should: today's chip spending is already running ahead of what today's actual AI services revenue alone would justify — that revenue, frontier labs included, is probably in the low hundreds of billions right now, not low trillions yet. That's not a flaw unique to this cycle, though. It's the same shape every infrastructure bet in this piece takes, by definition — the interstate system was sized in 1956 for traffic that wouldn't exist until 1975, the fiber went into the ground years ahead of the demand that eventually filled it. Spending ahead of revenue isn't a warning sign here; it's what building infrastructure means. The part that's still genuinely open is whether revenue closes that gap on the timeline the rest of this piece is built around — that's the actual bet, and it's worth tracking closely rather than assuming it resolves itself just because the historical pattern usually works out.
The honest caveat to all of this: this chip math rests on a demand-side number I haven't fully defended here — how much real AI services revenue actually exists, and how fast it's growing toward Damodaran's $3–4 trillion baseline. That baseline has the advantage of coming from an independent valuation source with no stake in the outcome, rather than from the AI labs' own self-reported numbers — which is exactly where it gets more complicated, since those self-reported numbers are what the labs themselves are growing from. I'm setting that question aside here and taking it up properly in the companion piece on where this revenue actually comes from.
Where the bubble actually lives
Here's where I think most of the doom commentary gets the diagnosis right and the conclusion wrong. There is, without question, a bubble. I just don't think it's where most people are pointing.
Broadcom posted AI revenue up 143% year over year, beat its own quarter on revenue and earnings, and gave guidance that was merely excellent rather than the number some analysts had quietly decided to expect. The stock — and the chip sector with it — shed over $1.3 trillion in a single session, twice, within a few weeks. That's not a company's fundamentals breaking. That's a multiple — the price the market will pay for a dollar of those fundamentals — getting too excited and then sobering up. Nvidia, the company actually doing the revenue heavy lifting in this story, trades at roughly 13 times forward revenue — a genuinely modest multiple for the most dominant, fastest-growing company in the stack. If that multiple corrects further, it's the market trimming hype out of the price, not a sign anything is structurally wrong underneath. The hype concentrates wherever the story is punchiest, not wherever the numbers are soundest — AMD, growing off a smaller base with a punchier narrative, carries a richer multiple and less of that cushion.
Everything so far has been about chip company stock prices — sentiment, not solvency; none of those companies is at any real risk of missing a payment. The bubble's second home is a level up the stack, in the data center operators and hyperscalers borrowing against future AI revenue — a solvency risk this time, and one that isn't evenly distributed either.
How this gets financed matters here — Morgan Stanley estimates roughly a $1.5 trillion gap between what hyperscalers can self-fund and what this buildout actually costs, filled by equity and debt very differently. Equity carries no repayment obligation; debt has to be serviced on schedule no matter how the bet performs — and where that debt sits determines who's actually exposed.
Google and Amazon are conspicuously missing from what follows, and it's worth saying why. Both are tapping debt markets too — Alphabet alone has reportedly worked on a $15 billion bond for AI investment — through ordinary corporate bonds on their own balance sheet, the same basic structure Oracle uses. What actually matters isn't structure, though — it's strength. Oracle carries that debt on an already highly leveraged company with a concentrated customer base; Google and Amazon carry it on some of the most cash-generative, diversified businesses in the world. Same kind of debt, very different ability to absorb it — which is why the table below focuses on the more structurally distinctive arrangements Microsoft and Meta built, not on company size.
| Company | Where the debt sits (structure and terms) | Backed by (what stands behind repayment) | The exposure |
|---|---|---|---|
| Microsoft (its AI Infrastructure Partnership, or AIP) | Ring-fenced in a separate fund | The fund's own assets and contracts, not Microsoft's balance sheet | A shortfall stays in the fund, not on Microsoft's books |
| Meta (its Hyperion data center project) | An off-balance-sheet joint venture, financed with long-dated, fixed-rate debt | The data center asset itself, backstopped by Meta's residual-value guarantee | A residual-value guarantee, not maturity risk |
| Oracle | On its own balance sheet | Oracle's whole company, which is already highly leveraged | Customer concentration plus leverage, already showing up in widening credit spreads |
| CoreWeave | On balance sheet, financed with floating-rate loans tied to specific GPU contracts | Those same GPU contracts | A 2026 debt maturity roughly equal to its entire cash position |
The companies whose core business can absorb a bad bet built that protection into how they financed it — whether that protection comes from an elaborate ring-fenced structure, like Microsoft's and Meta's, or simply from a balance sheet strong enough that ordinary bonds, like Google's and Amazon's, never become a real threat. CoreWeave doesn't have that cushion either way. None of this changes the aggregate math from Lens One — it's the reason "the bubble" isn't one evenly distributed risk; it's concentrated in whoever financed their seat at the table with debt that has nowhere to hide if the bet runs late.
The bond market, for what it's worth, has been less impressed by the story all along. SpaceX just had to pay a higher cost of borrowing than other similarly rated companies despite strong enough demand to upsize its bond from $20 billion to $25 billion — debt investors, as Allianz's chief investment officer put it, want their coupon regardless of how compelling the Mars story is. I'd take that as a healthy sign, not an alarming one.
If you want the full multiples, the second selloff, and the financing detail spelled out
A few weeks after the first $1.3 trillion semiconductor selloff, it happened again — a fresh $1.3 trillion erased from the sector in another session, with strategists openly describing it as a valuation reset rather than a change in the underlying story. Nvidia alone lost $740 billion of the first round on pure sentiment, despite having no real competitive overlap with Broadcom's business — that's how indiscriminately a sector-wide re-rating can move even an unrelated stock.
On the actual multiples: Nvidia's 13x figure sits against a $5.1 trillion market cap and an estimated $391 billion in fiscal 2027 revenue. AMD trades at 17 to 23 times, richer specifically because the market is pricing in a longer runway of share gains that haven't happened yet, not because its current numbers are weaker. Micron, posting margins that now rival Nvidia's after a decade of brutal commodity cycles, trades at the cheapest multiple of the three — the market still hasn't fully believed memory's good fortune is durable.
On the financing side: Microsoft and Meta's AI debt mostly sits behind an entire profitable company — search, cloud, retail — or is ring-fenced in a vehicle explicitly designed so a shortfall stays there and not on the parent's books. CoreWeave's revenue is concentrated in a handful of customers and its debt moves with interest rates in real time. Oracle sits in between, carrying real leverage on its own books rather than ring-fenced in a vehicle built to take the hit.
One more liquidity risk worth knowing, separate from whether any of this is fundamentally sound: SpaceX's float is expected to be followed by Anthropic's and OpenAI's IPOs in the same general window, and some investors are openly questioning whether the market can absorb that much new equity supply smoothly — the US market is reportedly close to tipping into net positive equity issuance for the first time in two decades, after years of buybacks shrinking the available float. A pile of fundamentally sound companies can still produce a disorderly few months if too much paper lands on the market at once. That's a plumbing problem, not a thesis problem, but worth knowing the difference before you mistake one for the other.
Why you build anyway
Why would any individual company keep pouring tens of billions into this, knowing that most of them — by almost everyone's admission, including some of the people building it — will not come out ahead? I think the honest answer is that not building is the worse bet — and the proof of it is something everyone reading this already uses.
In May 1999, Barron's ran a cover story called "Amazon.bomb," dismissing Jeff Bezos as "just another middleman" and predicting the company's collapse. The criticism wasn't crazy at the time: Amazon ran up nearly $3 billion in cumulative losses through 2001, didn't post a profitable quarter until that year's fourth, and watched its stock fall from $116 to $6 after the dot-com crash. Bezos kept reinvesting in infrastructure anyway — warehouses, then the servers and data centers Amazon needed to run its own retail business. By his own account, when Amazon started renting out the spare capacity on that infrastructure as what became AWS, "a lot of observers characterized [it] as a risky distraction... 'What does selling compute and storage have to do with selling books?' they wondered."
Today AWS isn't a side hustle — it generates the majority of Amazon's operating profit, on top of infrastructure that started as pure overhead nobody was praising in 1999. That's the sharper version of this thesis: physical infrastructure doesn't automatically reward whoever builds it, and it didn't save the dozens of other dot-com names that burned cash on warehouses and servers and still went to zero. It rewarded the one company that paired the infrastructure with real discipline and a long enough time horizon to still be standing when a use for it nobody had predicted showed up. The bet isn't "build infrastructure and you'll eventually win." It's "infrastructure gives you the option to capture a future nobody can see yet — but only if you're disciplined enough to still be holding it when that future arrives."
None of this verticalization transfers cleanly to the chip companies — Nvidia, AMD, and TSMC don't own a data center fleet or a customer-facing app, and they're not going to. But does that make a pure infrastructure business fragile? AWS already answers this from inside the same example: no app or storefront of its own, profitable regardless of which customer wins, and exactly where Amazon's most durable value ended up — closer to the house than to any one player at the table.
TSMC makes the same point from the chip layer — and it's not alone there, since Nvidia and AMD occupy the same kind of position one level up the stack. None of the three do anything beyond their own narrow specialty: TSMC doesn't design chips, Nvidia and AMD don't manufacture them or run the data centers that house them. What makes that valuable isn't the absence of competition — it's that staying at the bleeding edge takes a scale of continuous, compounding investment that's extraordinarily hard to match. TSMC holds roughly two-thirds of the foundry market and an estimated 90% of the world's most advanced chip production because it keeps reinvesting in the next process node before anyone else can close the gap. Nvidia and AMD are making the same bet one layer up: stay far enough ahead that being a step behind isn't a viable place to compete from. A pure-play infrastructure position isn't weaker for lacking an apps layer on top of it — it can be exactly where the durable value concentrates, as long as the investment in staying ahead never lets up.
More examples that make the same case
Every transformative infrastructure technology has been through some version of this. Railroads had 1869 — the transcontinental line completed with nowhere near enough freight to fill it — then 1873, when overbuilding triggered a panic that bankrupted a quarter of the country's railroads, set off partly by Jay Cooke's failed financing push for a second transcontinental line, the Northern Pacific, which dragged the broader economy down for years. (This is also, literally, where this piece's title comes from — a track record meant actual track, laid down in steel, long before it meant anything else.) Highways had 1956: Eisenhower's Federal-Aid Highway Act sized the interstate system for the traffic planners expected by 1975, two decades out, not the traffic that existed the day he signed it. The internet had its own version in the 1990s: the Telecommunications Act of 1996 set off a frenzy of fiber-optic buildout so far ahead of demand that, as Michael Burry has pointed out, less than 5% of U.S. fiber capacity was even in use by the time the telecom crash hit.
More recently, and closer to home: Ford put $2.7 billion into Argo AI and wrote off every dollar when it folded in 2022; GM spent over $9 billion on Cruise and walked away with, in one tracker's words, "nothing to show for autonomy"; Intel spent $15 billion on Mobileye, ending up with a real, profitable driver-assistance business instead of the robotaxi future it was sold on; and Volkswagen lost money alongside Ford in that same collapsed Argo AI venture. None of these four companies got back what they originally expected, and none of them went out of business over it either — their core businesses absorbed the loss the same way Detroit absorbed going "0 for 2" on the most important auto bet of the decade. Each was paying for a seat at a table that, for a while, looked like it might decide who got to exist in the next era of the industry — and for companies the size of Microsoft, Google, Amazon, and Meta, where AI isn't a side project but a bet on the next form of their actual core business, that's closer to paying a premium against becoming the next Kodak or Nokia than it is to losing a venture bet.
On the deeper architecture risk, if you want it
There's a version of the bear case that's more interesting than "the revenue won't show up" — it's "the infrastructure being built is making a bet on the wrong algorithm." Yann LeCun left Meta convinced that scaling today's large language models further will not reach anything resembling general intelligence, and raised over $1 billion to pursue a fundamentally different architecture built on world models rather than text prediction. Separately, and through a completely different mathematical toolkit, Mahadevan has been building a case from category theory that the entire foundation needs rebuilding to get past hallucination, not just more data. I find the fact that two serious people, in two unrelated fields, independently arrived at "the current path doesn't get there" more convincing than either argument alone.
The good news, on examination, is that this is a smaller problem for the infrastructure thesis than it first appears. Both alternative approaches still run on GPUs — just trained on different data with different objectives. The chip and power layer doesn't need the current algorithm to be the right one to remain a justified investment; it needs some viable path through AI to keep using massive parallel compute, which every serious contender for what comes next still does.
I'll say plainly what I'm leaving out, because pretending otherwise would be dishonest: none of this engages with what happens to the people whose jobs get automated along the way. That's a real, separate question for economists and policymakers to sit with on its own terms — not something this piece resolves, and I'd be suspicious of anyone who tells you the economics settle it.
What I think the case actually adds up to is this: the bubble is in the stock price. The bet is in the spend. Those are different moving parts, not one yes-or-no question — seeing them separately is the whole point of looking under the hood, which is what this piece has tried to do, and what most of the noise about AI right now skips entirely.
I don't know which lab wins, which app wins, or which number turns out true. Neither does anyone else. The chips and the data centers underneath all of it don't need to know either. They're not playing the game. They built the casino. They built the table everybody else plays on. The house always wins.
This is the first in a longer thread — I'll be linking out to deeper pieces on the energy layer, the Cerebras wafer-scale bet, and a few of the application-layer markets (medicine, robotics, cybersecurity, defense) that I think matter more to this story than they're currently getting credit for.
AI was used to assist in the writing of this article.