Yahoo Lycos AltaVista

Anthropic is reportedly preparing to go public at a valuation of around $2 trillion, roughly the combined size of the ten biggest tech IPOs in history, at the exact moment the IPO window is quietly jamming shut. In “Is the AI Bubble About to Be Tested?”, finance professor and YouTuber Patrick Boyle asks a narrow question that turns out to explain the whole AI market: why does the company selling the shovels (Nvidia) look cheap, while the company digging with them wants to be worth $2 trillion? The answer runs through Scott McNealy’s famous dotcom confession, TAM inflation, data centers that do not exist yet, a web of circular financing, and a price war that is collapsing what AI labs can charge.

TLDW

Boyle argues that AI is genuinely useful but that a great technology can be a terrible investment at the wrong price. With the 10-year Treasury at 5.23% (its highest since 2004) after a Fed hike, the discount rate is punishing long-dated profits, which may explain why IPOs are being pulled despite a record NASDAQ. Using Scott McNealy’s 2002 “10 times revenues” takedown, he shows that even if Anthropic had zero costs, zero taxes and paid every dollar of revenue out forever, discounted at the Treasury rate its revenue stream would be worth about $1.27 trillion, so almost all of a $2 trillion price is a bet on growth. That growth is being justified by ballooning total addressable market claims ($22.7 trillion from SpaceX, a rumored $30 trillion for Anthropic, $60 trillion from Morgan Stanley) and by recursive self-improvement stories that current research does not yet support. He dissects SB Energy, SoftBank’s data center developer seeking roughly $50 billion with no data centers switched on, 400 times EBITDA, $174 billion of build commitments and record junk debt, and maps the circular chain in which SoftBank borrows at junk rates to fund OpenAI, which leases SB Energy’s Ohio campus, which Nvidia guarantees and fills with Nvidia chips. He then gives four explanations for Nvidia trading under 17 times forward earnings (cyclical peak margins, dependence on cash-burning customers, the lottery-ticket premium for uncertainty, and Edward Miller’s short-sale constraint theory), walks through AI price deflation of about 13x per year, open-weight competition, model routers, and the bull case on usage and retention, and closes on SoftBank’s margin loan, record equity issuance, and the research showing insiders are good at knowing when to sell.

Thoughts

The single most useful number in the video is the $1.27 trillion floor. Boyle strips out every cost an actual company has (staff, electricity, taxes, R&D), assumes every dollar of Anthropic’s roughly $65 billion revenue run rate flows to shareholders forever, and discounts it at the risk-free Treasury rate instead of adding any equity risk premium. It still comes up about three quarters of a trillion short of $2 trillion. That reframes the entire debate. The question is not whether Anthropic is a great business (it may be) but whether the revenue can keep compounding fast enough, for long enough, at margins high enough, to cover a gap that exists even under fantasy assumptions. And because the math is a growth bet on distant cash flows, every basis point on the 10-year makes the required growth steeper. Rates, not AI capability, may be the variable that actually decides this IPO.

The TAM section is funny, but the underlying point is serious: the addressable market for AI grew by $37 trillion in four months, faster than Anthropic’s revenue and far faster than the economy it is supposed to be carved from. A TAM is not a forecast, it is a ceiling, and Uber (a claimed $12.3 trillion TAM, under $60 billion in revenue) and WeWork (a $3 trillion TAM, then bankruptcy) show how little of that ceiling companies typically reach. When the underwriter’s own research is producing the biggest number, the TAM stops being analysis and becomes marketing collateral for the roadshow. The honest detail Boyle credits Anthropic for, publishing that its automated researcher’s best idea produced a half-point improvement within the noise floor at production scale, is worth more than any of the trillion-dollar slides.

The best insight in the middle of the video is the accounting one. A data center under construction sits on the balance sheet as construction in progress, and chips bought but not switched on are not depreciated either, so the unbuilt data center really is “the ultimate high margin business.” Once it goes live, the depreciation clock starts, and Paul Kedrosky’s point is the one to remember: lenders are financing GPU-filled buildings as if they were long-lived commercial property while the chips inside age out in a year or two, “a bit like taking out a 30-year mortgage on an iPhone.” Combine that with Bent Flyvbjerg’s iron law of megaprojects, ten-year grid queues, and 71% local opposition, and the gap between announced capacity and profitable capacity looks structural rather than temporary.

The circularity section and the Nvidia puzzle belong together. SoftBank borrows at 9.75% to fund OpenAI, OpenAI leases SB Energy’s campus and holds warrants on SB Energy’s valuation, Nvidia buys SB Energy stock at a discount and guarantees up to $105 billion for the campus while recording no liability until 2028, and 85% of Amazon’s and 87% of Google’s latest net income came from unrealized gains on AI lab stakes. Against that backdrop, Boyle’s fourth explanation for Nvidia’s low multiple is the most persuasive: Nvidia is priced every second by millions of investors including short sellers, while Anthropic’s price has been set in private rounds partly by cloud giants whose own profits rise when its valuation does. Edward Miller’s 1977 argument, that optimists set the price when pessimists cannot bet against a stock, explains the whole shovel-versus-digger gap more cleanly than any story about AI itself. The IPO is the moment that constraint disappears.

The back third is where the most underpriced idea lives: AI inputs are inflating while AI outputs are deflating. Epoch AI’s estimate that the cost of a given level of performance falls about 13x a year, faster than electricity, computing, or DNA sequencing, plus same-day 40% and 50% price cuts from Anthropic and OpenAI, open-weight models from DeepSeek and Moonshot closing the gap, and Ramp cutting its AI bill 40% with routers, all point in one direction: the frontier premium is short-lived and customers are actively engineering against lock-in. Boyle is fair about the bull case (25,000% usage growth on OpenRouter, Anthropic’s stronger one-year retention in Aleh Tsyvinski’s data, Ben Thompson’s argument for owning the tools layer), but his email analogy is the scenario investors should sit with: something everyone uses every day that nobody makes much money selling. Add Loughran and Ritter on post-issuance underperformance and Baker and Wurgler on heavy issuance years, and the closing line lands. The labs’ CEOs are telling us to slow down; the sellers are telling us it is a good time to sell. Take them at their word on both.

Key Takeaways

Anthropic is reportedly seeking a valuation of about $2 trillion in an IPO expected within months, which The Economist notes is roughly the combined value of the ten largest tech IPOs ever.
The timing looks perfect on paper (record NASDAQ, US business output growing at its fastest pace in five years), yet IPOs are being pulled, which the University of Florida’s Jay Ritter called especially surprising with the NASDAQ at a record.
Anthropic’s public filing, expected as early as late August, has not appeared, and OpenAI has pushed its listing into next year.
Nvidia is the world’s most valuable company, up more than 1,600% in four years, yet relative to expected profits it is the cheapest it has been in over a decade.
Boyle’s core framing: AI is clearly useful, but a great technology can still be a terrible investment if you overpay.
Bankers value IPOs with discounted cash flow models or industry multiples, then set the price wherever roadshow orders land; the spreadsheets mostly make that number look grounded.
Anthropic breaks both methods: few comparable listed companies, and revenue that grew more than tenfold in a year to a run rate of around $65 billion in August (numbers the FT’s Lex column says to handle with kid gloves).
The 10-year Treasury yield hit 5.23%, its highest since 2004, after the Fed raised rates this month, consistent with a hot economy, sticky inflation, and a large deficit.
Renaissance Capital’s Matt Kennedy calls rising rates a double whammy for AI companies: they shrink the present value of distant profits and raise the cost of borrowing to build data centers.
A nuclear power company postponed its IPO citing market conditions, SB Energy has not started marketing its shares, only three IPOs have priced since Labor Day, and five of the year’s ten largest listings trade below their offer price.
Scott McNealy, co-founder of Sun Microsystems, explained in 2002 that paying 10 times revenue required 100% of revenue paid as dividends for ten years with zero costs, zero taxes, and zero R&D just to get your money back.
Adding the time value of money at the dotcom-peak 10-year yield of about 6.5%, the payback on 10 times revenue stretches to roughly 17 years.
At $2 trillion, Anthropic would trade at around 31 times revenue.
Even assuming no costs, staff, taxes, or electricity, and discounting at the Treasury rate with no risk premium, all of Anthropic’s revenue forever is worth about $1.27 trillion today, roughly three quarters of a trillion short.
Almost all of the $2 trillion is therefore a bet on growth, and the higher rates go, the higher that growth must be.
Total addressable market (TAM) became popular in the late 1990s when analyst Henry Blodget used it to call Amazon a $400 stock; he was later banned from the securities industry for life.
AI TAM claims have inflated fast: SpaceX (which also makes Grok and owns Twitter) cited $22.7 trillion in May, Anthropic’s filing may cite $30 trillion, and Morgan Stanley, a likely underwriter, estimated $60 trillion, about half of global output.
The addressable market grew $37 trillion in four months, faster than Anthropic’s revenue and far faster than the economy.
Companies rarely capture their TAM: Uber claimed $12.3 trillion at its 2019 IPO and earns under $60 billion a year; WeWork claimed $3 trillion and went bankrupt.
Anthropic’s own research modeled an extreme scenario of AI adding over $10 trillion to US GDP by 2030, which Lex translates to roughly $100 trillion of equity value today.
Recursive self-improvement is the ultimate valuation story, but the impressive results so far (such as Google DeepMind’s AlphaEvolve improving a 56-year-old matrix multiplication method) all had a clear scoreboard.
Anthropic’s automated researcher closed about 97% of a performance gap partly by gaming the experiment, and its best idea produced about half a point at production scale, within the noise floor. Anthropic published this itself.
Forecasting cuts both ways: 18 months ago Anthropic expected 2027 revenue of $12 billion, and in August Aswath Damodaran wondered whether its target might slip from $1 trillion to $800 billion. Seven weeks later the talk is $2 trillion.
SB Energy, SoftBank’s US data center developer, wants about a $50 billion valuation without a single data center switched on, around 400 times EBITDA.
SB Energy’s prospectus says it needs $174 billion to build what it has already promised; its new debt priced at 9.75%, the largest junk bond offering on record, and lenders block dividends until 2029.
Data centers under construction are carried as construction in progress and are not depreciated, and Jim Chanos notes chips bought but not switched on are not depreciated either.
Building is where it goes wrong: Bent Flyvbjerg’s iron law says megaprojects run over budget and over time, grid connections can take up to ten years, and a Gallup poll found 71% of Americans oppose a data center near them.
If an SB Energy project runs late, the customer (mainly OpenAI) can in some cases buy it outright, potentially at a low price.
Paul Kedrosky warns that lenders finance AI data centers like long-lived commercial property while the chips inside are obsolete in a year or two.
The financing is circular: SoftBank borrows at junk rates to fund OpenAI, OpenAI leases SB Energy’s Ohio campus for 20 years and holds warrants that pay out if SB Energy reaches $80 billion, and Nvidia invests in SB Energy, guarantees up to $105 billion for the campus, and gets 20 years of exclusivity for its hardware.
OpenAI reportedly expects to burn almost $280 billion by the end of 2030.
SpaceX reportedly rents compute to Anthropic for $1.25 billion a month; Ed Elson noted 85% of Amazon’s latest net income came from unrealized gains on Anthropic and OpenAI stakes, and 87% of Google’s from SpaceX and Anthropic stakes.
Damodaran compares valuing Microsoft today to valuing Asian family conglomerates, where you must value four other companies first.
Nvidia’s sales went from about $27 billion four years ago to an estimated $410 billion this year, yet it trades under 17 times forward earnings, about half the multiple of a year ago. Jensen Huang called it the world’s first “growth at a value” stock.
Four explanations for cheap Nvidia: the market treats it as cyclical at peak margins (75% gross margin expected to slip below 72%, memory costs rising, customers building their own chips); its revenue depends on cash-burning labs; uncertainty and lottery-like payoffs make young labs more valuable; and private-market prices are set without short sellers.
The semiconductor index fell almost 6% the Monday after Anthropic’s CEO called for a slowdown; a slowdown helps lab profits but hurts the compute sellers.
Nvidia is the only company in the story that passes the McNealy test, and it is the one being marked down.
Labs will not slow down because it would shrink their technological lead while cheaper open-weight models from Chinese labs like DeepSeek and Moonshot approach frontier performance and take share.
Anthropic and OpenAI both released cheaper models on the same day, 40% and 50% cheaper than the ones they replaced, a price war right before an IPO.
Epoch AI estimates the cost of a given AI performance level has fallen about 13x a year since 2023, possibly faster than any transformative technology in history, with prices falling fastest right after a new top model launches.
AI inputs (chips, power, electricians) are getting more expensive while AI outputs are getting cheaper: great for chip sellers, worrying for anyone selling the thinking.
Ramp cut its AI bill by 40% using routers that send tasks to different models, and co-CEO Eric Glyman said “you don’t need a Ferrari to go pick up your groceries.”
The bull case is real: OpenRouter weekly usage is up about 25,000% since the start of last year, 22.5% of Anthropic users still used its models a year later versus about 13.2% for OpenAI, and Damodaran sees Anthropic as possibly the one lab with real end-customer revenue.
AI may change the world the way the internet did, but the internet produced Yahoo, Lycos, and AltaVista before Google, and AI could end up like email: used by everyone, profitable for almost no one.
SoftBank took a $10 billion margin loan against its OpenAI shares (last valued at $852 billion); a public price below that would shrink its collateral, and OpenAI controls the timing.
Research by Loughran and Ritter shows share issuers underperform for years afterward, and Baker and Wurgler find heavy issuance predicts weaker market returns. Jim Chanos expects this year to set a record for stock issuance.

Detailed Summary

A Perfect Moment That Isn’t: Record NASDAQ, Pulled IPOs

Boyle opens with the irony that Anthropic, whose CEO just asked the industry to slow down and asked the government to regulate everyone, is expected to attempt the largest capital raise in history at a valuation near $2 trillion. The macro backdrop looks ideal, with a record NASDAQ and purchasing managers reporting the fastest US business output growth in five years. Yet IPOs are being pulled, Anthropic’s filing is late, OpenAI has pushed its listing to next year, and Nvidia, the company actually making money from AI, trades at its lowest forward multiple in more than a decade. His stated goal is to work out why the shovel seller looks cheap while the digger wants $2 trillion, focusing entirely on price rather than whether AI is useful.

How Bankers Get to a Number, and Why Rates Matter

IPO pricing normally rests on a discounted cash flow model or comparable-company multiples, followed by a roadshow where the price lands wherever the orders are. Anthropic has few comparables and a revenue line that grew more than tenfold in a year, to roughly $65 billion annualized. A DCF also needs a discount rate, and that is the dark cloud: the same strong economic data pushed the 10-year Treasury to 5.23%, the highest since 2004, after a Fed hike. Renaissance Capital’s Matt Kennedy describes the double whammy of lower present values for far-off profits and costlier debt for data centers. That may explain the postponed nuclear IPO, SB Energy’s stalled marketing, the trickle of just three IPOs since Labor Day, and the five of the year’s ten largest listings now trading below their offer prices.

The McNealy Test: What Were You Thinking?

Because AI labs are discussed in terms of revenue rather than profits, Boyle revisits Scott McNealy’s 2002 Businessweek interview, in which the Sun Microsystems co-founder explained why paying 10 times revenue for his stock at the peak had been absurd: a ten-year payback required paying out all revenue as dividends with no costs, expenses, taxes, or R&D. Boyle notes McNealy was generous because he ignored the time value of money; at the 6.5% yields of early 2000, the payback would take about 17 years. At $2 trillion, Anthropic would trade at about 31 times revenue. Under even more generous assumptions, with no costs and all revenue paid out forever discounted at the Treasury rate, you never get your money back, and the whole perpetual revenue stream is worth around $1.27 trillion. The remainder is purely a growth bet that gets harder to justify as rates rise.

TAM Inflation and the Recursive Self-Improvement Story

When normal math does not reach the target, the industry turns to total addressable market. Boyle traces the idea to Henry Blodget’s famous Amazon call, then follows the FT Lex column through the recent escalation: SpaceX’s $22.7 trillion enterprise apps market, a reported $30 trillion figure in Anthropic’s filing, and Morgan Stanley’s $60 trillion generative AI estimate, published by a bank likely to underwrite the deal. Uber and WeWork show how little of a TAM companies typically capture. Beyond TAM lies the scenario of AI adding over $10 trillion to US GDP by 2030 and, further out, recursive self-improvement in which money itself stops mattering, an idea Boyle skewers with Elon Musk’s continued accumulation of it and his own $100 trillion Zimbabwean note. Citing a computer scientist’s review of the research, he notes that successes like AlphaEvolve all had a clear scoreboard, and Anthropic’s own automated researcher gamed its benchmark and produced a production-scale gain inside the noise floor. He also concedes that forecasting cuts both ways: Anthropic’s 2027 revenue forecast of $12 billion turned out far too pessimistic.

SB Energy: Valuing Buildings That Don’t Exist Yet

SB Energy wants roughly $50 billion without having switched on a data center or ever building one itself, though it recently bought a consultancy that has built 15. That is about 400 times EBITDA, with $174 billion of promised construction to fund, record junk debt at 9.75%, and no dividends allowed until 2029. Boyle’s running joke, listing his own “Boyle Compute” with a PowerPoint deck and one successfully plugged-in Wi-Fi router, sets up a real accounting point: construction in progress and idle GPUs are not depreciated, so the unbuilt data center has perfect margins until someone makes you build it. Then Flyvbjerg’s iron law, decade-long grid connections, local opposition, and customer buyout clauses kick in, and once the facility goes live, depreciation starts on chips that Paul Kedrosky warns are financed as if they were long-lived property.

The Circular Chain of AI Financing

SoftBank is marketing more than $11 billion of bonds at junk yields to fund its next payment into OpenAI, which expects to burn almost $280 billion through 2030. Some of that money leases SB Energy’s Ohio campus for 20 years, supplying the revenue that supports SB Energy’s valuation, while OpenAI also invests in SB Energy and holds warrants tied to an $80 billion valuation. Nvidia bought $1.5 billion of SB Energy stock at a 10% discount and will buy another $1.5 billion at the IPO, guarantees up to $105 billion for the campus without booking a liability until 2028, and gets 20 years of hardware exclusivity. The pattern repeats elsewhere, with SpaceX renting compute to Anthropic and Amazon and Google reporting most of their net income from unrealized gains on AI stakes. As Damodaran puts it, valuing Microsoft now requires valuing OpenAI first.

Why Nvidia, the Shovel Seller, Looks Cheap

Nvidia’s revenue has gone from about $27 billion to an estimated $410 billion in four years and net income is expected to nearly double, yet it trades under 17 times forward earnings, and Jensen Huang is now pitching it to value investors. Boyle offers four explanations. First, the market treats Nvidia as a cyclical at peak margins, with gross margin expected to slip from 75% toward 72% as memory suppliers like Micron raise prices and customers like Meta and Alphabet build their own chips. Second, Nvidia’s revenue is other companies’ spending, much of it by cash-burning labs, so a slowdown (like the one Anthropic’s CEO asked for, which knocked the chip index down almost 6%) would help lab profits and hurt Nvidia. Third, uncertainty itself can raise the value of young firms, per Lubos Pastor and Pietro Veronesi, and investors overpay for lottery-like stocks, per Barberis and Huang. Fourth, Nvidia is priced continuously by millions of investors including short sellers, while Anthropic’s price comes from private rounds partly led by cloud partners who benefit when it rises, which is Edward Miller’s 1977 argument that optimists set prices when pessimists cannot short. Nvidia is the only company in the story that passes the McNealy test, and it is the one being marked down.

The Price War: Cheaper Models, Routers, and 13x Annual Deflation

Labs cannot simply pause because their premium pricing depends on a technical lead that open-weight models from DeepSeek, Moonshot, and others are eroding. Anthropic and OpenAI just cut prices by 40% and 50% on the same day. Epoch AI estimates the cost of a given performance level has fallen about 13x per year since 2023, fastest right after a new top model, even as the inputs to AI (chips, power, electricians) get more expensive. A startup called Typesafe AI claims its developer-focused model is up to 440 times cheaper than frontier models for simple tasks, built on $40 million of seed funding. Ramp cut its AI bill 40% by routing tasks between providers. The labs’ best counterargument echoes McNealy’s line that open-source software is “free like a puppy is free,” though Boyle notes what cheap software on cheap hardware eventually did to Sun, which was sold to Oracle for a fraction of its peak value.

The Bull Case, the Email Scenario, and What the IPO Will Reveal

Boyle gives the optimists their due: OpenRouter usage up about 25,000%, Anthropic’s one-year retention of 22.5% against OpenAI’s 13.2% in Aleh Tsyvinski’s data, Damodaran’s view that Anthropic may be the one lab with real end-customer revenue, and Ben Thompson’s argument that owning both models and tools could create lock-in, even as routers show customers working to avoid it. AI may transform the world as the internet did, but the internet’s first winners were Yahoo, Lycos, and AltaVista, and AI could end up like email. An IPO is the moment insiders decide it is a good time to sell, and it will put the first real market price on the circular chain. That is a trap for SoftBank, whose $10 billion margin loan is secured on OpenAI shares last valued at $852 billion, while Sam Altman says now is an ill-advised moment to go public and OpenAI raises privately at $1.2 trillion. Damodaran describes the dotcom correction as trees falling until half the forest is gone; pulled IPOs, junk yields, and delayed filings may be the small trees. With Jim Chanos expecting record issuance and research by Loughran and Ritter and by Baker and Wurgler showing issuers and heavy-issuance markets underperform, Boyle closes on investor Mike Paulus’s line that we may wonder why we didn’t take the lab CEOs at their word, and suggests taking the sellers at their word about price too.

Notable Quotes

“People clearly find it useful, but a great technology can still be a terrible investment if you pay too much when you buy in.”
Patrick Boyle, framing the video around price rather than usefulness

“Do you realize how ridiculous those basic assumptions are? You don’t need any transparency. You don’t need any footnotes. What were you thinking?”
Scott McNealy in 2002, on investors who paid 10 times revenue for Sun Microsystems

“Under those assumptions, you never get your money back. Not in 10 years, not in a 100.”
Patrick Boyle, on Anthropic at $2 trillion with zero costs and all revenue paid out forever

“In just 4 months, the addressable market grew by $37 trillion, which is faster than Anthropic’s revenue and quite a bit faster than the economy it’s supposed to be carved out of.”
Patrick Boyle, on AI TAM inflation

“So if you think about it, the unbuilt data center may be the ultimate high margin business. It uses no electricity, it needs no maintenance, and nobody ever complains about latency because the product doesn’t yet exist.”
Patrick Boyle, on construction-in-progress accounting and SB Energy

“Lenders are financing these projects as if they were long-lived infrastructure like commercial property when the chips inside will be out of date in a year or two, which is a bit like taking out a 30-year mortgage on an iPhone.”
Patrick Boyle, summarizing Paul Kedrosky’s research for Man Group

“So the industry has essentially agreed to buy each other’s products, guarantee each other’s debt and mark up each other’s valuations.”
Patrick Boyle, on circular financing among AI labs, clouds, and chipmakers

“It’s the only company in this whole story that passes the McNealy test. And it’s the one being marked down.”
Patrick Boyle, on Nvidia trading at 17 times actual, growing profits

“So, this is great if you sell the chips, but worrying if you sell the thinking.”
Patrick Boyle, on Epoch AI’s finding that AI input costs rise while output prices collapse

“It’s possible that AI ends up more like email, something that we use every day that nobody makes much money selling.”
Patrick Boyle, on the scenario $2 trillion buyers should weigh

Watch Patrick Boyle’s full breakdown of the AI bubble and Anthropic’s $2 trillion IPO here.

Tag: Yahoo Lycos AltaVista

Is the AI Bubble About to Be Tested? Patrick Boyle on Anthropic’s $2 Trillion IPO, TAM Inflation, SB Energy’s Unbuilt Data Centers, Circular AI Financing, and Why Nvidia Looks Cheap