Anthropic Says $30 Trillion. Only Nvidia Showed the Money.

THE NUMBER: $30 trillion. That’s the total addressable market Anthropic is preparing to put in front of IPO investors, according to the Wall Street Journal — bigger than the entire economy of China, roughly the whole annual output of the United States, about a quarter of the planet. It edges out the $28.5 trillion SpaceX floated in its own filing in May and takes the record for the largest number anyone has ever printed on a pitch deck. Hold it, because on the very same afternoon that dream hit the wires, the same company signed a $45 billion bill to rent the electricity to chase it, a free model out of China matched its flagship on code, and the one company actually banking cash off the whole boom reported a quarter for the ages. The dream and the invoice, same day. One of them is already in the bank.

A market the size of a country

Start with what $30 trillion actually claims, because it’s not a forecast — it’s a redefinition.

For thirty years, a technology company sized its market by the budget it was chasing. Salesforce fought for CRM dollars. Snowflake went after the data-infrastructure line. Gartner puts all of global software spending around $1.47 trillion, and if you sold software, that ceiling was your world. Anthropic is doing something different, and it’s the whole game: it isn’t sizing the software budget, it’s sizing the labor. The pitch, per the Journal, bases the number on the full scope of work its models could theoretically perform — every memo drafted, every contract reviewed, every line of code shipped. “With the way Claude writes code, you could argue it’s replacing the work humans do end-to-end,” Alex Brunicki of Backed VC told Fortune, “and so the TAM is essentially the labor market for that work output.”

That’s the trick that gets you to $30 trillion. Not “how much do companies spend on AI,” but “what is all the knowledge work on earth worth, and what if we ate it.” Frame it that way and no number is too big — legal, accounting, engineering, the entire back office of the global economy folds into the denominator. It’s audacious, and it’s not stupid. It’s also, by Anthropic’s own hand, theology.

Because the confession is right there in the same filing. Anthropic’s own near-term projection is roughly $190 to $200 billion of revenue by 2028. Put the two numbers next to each other and they disagree by a factor of a hundred and fifty. The company is showing retail investors a $30 trillion sky and telling sophisticated ones, in the fine print, to model $200 billion. Brunicki said exactly that: the pros “are going to build their own cash-flow models” off contracts and roadmap, while the number that grabs headlines is the one for the traders who won’t. Fred Hickey, who has been writing The High-Tech Strategist since the last two bubbles, was blunter on X: “annual U.S. GDP is currently $32.5 trillion, and yet this nonsense is allowed to continue so that Wall Street and Silicon Valley can extract as much money from unwitting investors as possible, before the inevitable stock market bubble collapses.”

We’ve seen this movie. In 2019 Uber told the market its TAM was $6 trillion by counting every mile every human travels. In October 1999, the 199 internet stocks Mary Meeker tracked carried a combined $450 billion of market value against $21 billion of sales and $6.2 billion of losses. The TAM slide is the oldest instrument in the IPO band. What’s new isn’t the size of the number. It’s what the number is a claim on.

Ep 16 – Google’s Dream Team Just Quit, and Nobody Can Find the AI Bear Case

Four top Google AI researchers walked out the same day. Anthony Batt and Harry DeMott on what that exodus actually signals, and why the industry’s doom talk might be more marketing than warning.

The invoice came the same afternoon

Here’s the part the eye-roll misses. While one wire carried the $30 trillion dream, another carried the bill.

The same Wednesday, Anthropic agreed to spend $45 billion over six years renting compute from Nscale, a two-year-old London company, at a data-center campus in West Virginia. Four hundred and sixty megawatts of power — what Bloomberg reckons is the draw of about 345,000 homes — filled with Nvidia’s next-generation Vera Rubin chips. The full campus, with its own on-site power plant, will run about $71 billion to build, $47 billion of that in silicon. And here’s the tell inside the tell: Microsoft signed the letter of intent for that exact site in the spring, then walked away over the summer. The disciplined hyperscaler looked at the economics and passed. The pre-IPO lab that needs a capacity story for its prospectus took the lease Microsoft dropped. As The Next Web put it, secured compute is “a story for the prospectus as much as for the engineers.”

This is not a one-off. It’s the pattern of Anthropic’s entire year. Fifty billion to Fluidstack. Forty-five billion to rent SpaceX’s Colossus, at $1.25 billion a month through May 2029. Tens of billions of servers from AMD, which is also putting up to $5 billion into the company. Anthropic does not own the machines that make its product. It rents them, on take-or-pay contracts that run for years, from landlords and chipmakers who get paid whether or not the $30 trillion ever shows up. We told you this in There Can Be Only One on August 17 — Dario’s own argument that AI “structurally concentrates power” is, read as an allocator instead of a philosopher, a confession that the power concentrates around whoever owns the compute, and he leases his. Now it’s not an argument. It’s a signed contract.

And look at who’s on the other side of it. Nscale is two years old. It booked something like $100 million of revenue in a recent quarter. Its contracted backlog — future money that counts in full the moment the ink dries — just vaulted to $51 billion, most of it this single Anthropic deal, and it’s taking that backlog public in New York next month. A dream, financing a dream, going public on the strength of a dream. If that arrangement rings a faint bell, it should: it’s the same booked-it-before-you-earned-it accounting that has ended every capital cycle since the railroads.

You can’t own what’s racing to free

Now the second problem with a $30 trillion market: the product inside it is getting cheaper by the week, and the people making it cheaper aren’t sending Anthropic an invoice — they’re giving it away.

The same day Anthropic floated its number, Z.ai shipped GLM-5.3-Flash under an MIT license — the most permissive license there is — a 320-billion-parameter model that, on the company’s own coding benchmark, comes within reach of Claude at roughly a tenth the price, and runs “entirely on Chinese AI chips.” Alibaba’s Qwen team put out an open-weight model that hobbyists are running on a Mac with 75 gigabytes of memory. Perplexity and Nvidia rolled out a fully local agent with, in their words, zero token cost. And OpenAI unveiled its first in-house inference chip, Jalapeño, which by SemiAnalysis‘s Dylan Patel — no easy grader — is “beating Nvidia Blackwell and even Rubin. This is huge news.” The cost of running a model just fell through the floor from four directions at once.

This is Clayton Christensen’s nightmare, and it’s worth being precise about the mechanism. A TAM measures a market before competition arrives. Anthropic is showing investors the value of a market on the same afternoon four separate competitors proved the thing sold inside it is deflating toward the price of electricity. A $30 trillion market on a commoditizing good is not a prize. It’s a measure of how much value is about to be competed away. The bigger the number, the more capital sprints through the gate, the faster the price collapses. You don’t own a $30 trillion market. You rent a thinning slice of it, and the rent your customers are willing to pay falls every quarter. Brunicki, the same VC who defined the TAM, said the quiet part: Anthropic’s capture “could easily come under pressure as more companies adopt industry-specific models built on cheaper open-source systems.” He also said the plain thing about the cohort raising at these valuations — many “are going to go to zero.”

Compute is revenue

So if the model is racing to free, where is the money actually landing? Jensen Huang answered the question on his own earnings call, and he didn’t dress it up.

Nvidia reported $96.2 billion of revenue for the quarter, up 106% from a year ago. Data center alone was $89 billion, up 117%. Seventy-five points of gross margin. It guided to $108 billion for the current quarter and told the Street to expect roughly 70% growth into next year — a forecast, as CNBC noted, far above what analysts dared to model. And then Huang, holding the biggest cash-printing quarter in the history of the semiconductor business, said it: “AI has reached its inflection point. It’s doing useful work. Its tokens are productive and profitable. Now, compute is revenue.”

Read that sentence next to Anthropic’s. Anthropic is showing you the value of all the work AI could someday do. Huang is showing you $96 billion that already cleared. One of those is a promise about the next decade; the other is a wire transfer that settled last week. This is the argument we made in Gotta Dance on August 12 — that compute had quietly become the real asset class, the productive, revenue-generating, fungible thing Jensen kept calling it — now confirmed by the man himself in seven words. The dreamers pitch the labor market. The shovel-seller books the cash.

And the shovel-seller is careful. Nvidia now backstops a chunk of its own customers’ financing — a $500 billion program with six of the largest allocators on earth, residual-value guarantees on the chips, Jensen effectively co-signing the buyers’ debt. Morgan Stanley opened credit coverage of Nvidia this week and, even liking the equity, wouldn’t touch the tail: “too early-stage, opaque, and sizable to step in.” Which tells you something. The house that’s winning every hand is also quietly insuring the players it’s winning from. You do that when you’ve noticed the table is more fragile than the pot suggests.

The most expensive rate in twenty years

Here’s the macro fact that makes the timing almost unfair, and nobody in the AI press is naming it: the 30-year Treasury just hit a 20-year high, around 5.3%.

A $30 trillion TAM is the longest-duration asset anyone has ever pitched. All of the payoff sits years out, in the fog, past the horizon of any spreadsheet. And the way you value a payoff that far away is to discount it back against the risk-free rate — and the risk-free rate has not been this expensive since before the iPhone existed. Every dollar of that far-off $30 trillion is worth less today than it was worth six months ago, not because the dream got smaller, but because the money to wait for it got dearer. We called this The Perfect Storm on August 20: the AI buildout is being funded with debt and equity into the teeth of a $40 trillion Treasury refinancing, and the boom is bidding up the very rate the boom is priced against. The call is coming from inside the house. Anthropic is unveiling the most speculative long-duration asset of the cycle at the exact moment the cost of financing a dream is at a two-decade high. That’s not a coincidence you can wave off. That’s the tide going out.

Seven names holding up the room

Which brings it home to the only question that matters for anyone with money in the market: if the dream is priced against the most expensive rate in twenty years, and the product is racing to free, what are you actually long when you’re long “the market”?

Because the market isn’t the market anymore. The whole index’s earnings growth now comes from about seven names, and we’ve gotten comfortable calling that diversification. It isn’t. It’s one bet wearing a diversified costume. And when you take the costume off, the seven names aren’t all the same bet, which is the part worth getting right. Nvidia gets paid whether open source wins or the frontier does — it sells the shovel to every side of the war, and that’s the cleanest position in the whole business. Apple is the bet on the models coming home to your desk, the $899 box we wrote about yesterday, the open-source-on-device play. Amazon, Microsoft, SpaceX — those are compute, toll roads, fine on survival. On the names, the cohort is mostly defensible.

The trouble is the rate. The growth rate that makes the Magnificent Seven look bulletproof is, to a real degree, the frontier labs’ capex, recycled. Anthropic’s $45 billion of rent is somebody else’s revenue line — it lands on Nvidia’s income statement, on a cloud provider’s bookings, on a landlord’s backlog. So the second derivative isn’t whether Anthropic is worth $2 trillion. It’s that the day the labs slow their spend — because the dream deflated, or the funding door shut, or the 20-year-high discount rate finally bit — the growth rate of everything “safe” deflates with them. Nvidia still gets paid. It just stops getting paid more, faster. And its own out-year guide, the +70% that has the Street levitating, rests on labs paying take-or-pay leases they’re funding against a $30 trillion story. Jensen knows it; that’s what the $500 billion backstop is for. The canary and the coal mine share the same air.

We called it in Canary in a Coal Mine on Monday: hardware is engineered scarce, so it appreciates; intelligence is engineered abundant, so it depreciates — and when the canary stops singing, you don’t check the canary, you check the air. The air in this market is one dream funding everybody’s revenue. Sooner or later that canary keels over, and the morning it does, you find out how few hands were actually holding up the room.

The tool went to zero. So did the middle.

Yesterday, in The Calculator Defense, we said the tool went to zero and judgment is the only alpha left — that when language becomes free, authorship migrates from producing the words to owning the thought. This is the same truth, moved up from the writer’s desk to the balance sheet. When the product is free, the money isn’t in making the intelligence. It’s in owning the road the intelligence runs on, or being the human who signs the answer. Everything in the middle — the frontier lab that rents the road and gives the judgment away, priced at $2 trillion on a $30 trillion dream — is the part that gets competed to the bone.

So here’s where a person with thirty years of watching these cycles ends up. I’m not buying the tenant at $2 trillion when the tenant rents its factory, gives away its product, and prices itself off a number a hundred and fifty times its own revenue projection. And I’m not calling seven stocks a diversified market, not when the growth holding them up is one correlated bet on the labs spending forever, into the most expensive money in twenty years. Own the shovel if you want the boom — Nvidia gets paid either way. Own the judgment if you want the part no glut can reach. Just don’t be the middle, and don’t mistake the size of the dream for the safety of the trade.

They’ll show you a $30 trillion market. Ask to see the money. The shovel and the judgment are the only ones still standing when the canary keels over — everyone in the middle is paying rent on somebody else’s future. The market is slower than the mob. But it always learns the difference.