Musk Says AI Doubles the Economy. The Company Pricing the Biggest AI IPO Says Otherwise

In the first days of September, the two statements that best frame AI's investment story landed within hours of each other. On September 9 Elon Musk posted on X that "AI + robots will more than double the global economy in less than 10 years," a claim he had been refining at the G20 — AI alone up 20% to 30%, robots "many multiples" more once robots can build robots. Almost the same day, Anthropic, the lab behind Claude, published an interactive model of what AI does to the U.S. economy through 2030. Its most optimistic scenario has GDP growth reaching 15.4% a year by 2030 — more than seven times today's pace. Its most conservative has AI mattering about as much as the internet did, lifting GDP 1.6% above baseline. The company assigns no probability to any path.

That spread — from "another internet" to "the economy grows seven times faster" — is not a curiosity. It is the range investors are quietly picking inside whenever they hold an AI stock. Nvidia, the backbone of this build-out, is worth about $5.26 trillion and still grew revenue 83% over the last year. Anthropic is on the eve of what bankers size as a debut near $1.8 to $2 trillion, one of the largest IPOs ever attempted. Such a price is not a bet that AI pays off in some scenarios. It is a bet on the top of Anthropic's own range. The doubling case and the cautious case collided in one news week because they are the two poles of the same valuation.

Where capability and value split

Here is where I think the signal sits, and it is not in either headline. Musk describes capability. Anthropic, when it runs the numbers, is describing diffusion — how fast AI spreads into the economy. Investment returns follow diffusion, not capability.

Capability is not in dispute. Anthropic itself says the number of tasks AI can complete independently doubles roughly every four months, and its extreme scenario assumes half of cognitive work is touched. The dispute is whether that capability reaches the economy, and who is left holding the margin when it does. That is a unit-economics question, and Anthropic's own books are the clearest answer available.

Anthropic's revenue is real and growing in a way that excuses hyperbole: about $9 billion at the end of 2025, a $47 billion run rate by May, on track for roughly $100 billion annualized by year-end. But its gross margin sits around 40% — and Anthropic lowered it in late 2025 specifically because inference costs rose. Compute is the dominant cost of the models it sells. In plain terms: at the very moment AI is supposed to double the world economy, the company best placed to sell that AI keeps only about $0.40 of every revenue dollar before its compute bills and remaining operating spend.

Now hold that against Nvidia, whose gross margin is about 74%. The layer that makes the chips captures the value; the layer that delivers the intelligence runs nearly even. This is the hardware-to-software value migration in its early, inverted phase. In a mature cycle the application layer takes an ever larger share of profit as hardware costs fall. Here the hardware is so expensive and demand so great that the frontier lab looks less like a compounder than a toll road — spectacular revenue, thin capture, and a margin that fell, not rose, exactly when inference demand surged. That is the training-to-inference transition landing in an income statement.

The book that prices the bet

Now the part that changes the judgment: Anthropic is about to sell itself at the top of this range. Its bankers are reportedly valuing it on a bet that it books roughly $190 to $200 billion of revenue in 2028 — more than twenty times its 2025 base — a multiple only paid if the extreme scenario comes to pass. Even Musk concedes the binding constraint is not model quality but power: AI chip production is growing faster than the electricity supply, the next AI race "may be a race to build power," and labs including Anthropic are renting compute from SpaceX. A doubling of the economy gated by how fast the world can build power stations, at a 40% gross margin, is a back-half-weighted proposition, not a near-term one.

So what does an investor do with two headlines that cancel each other out? Treat each as expectation, not evidence. Musk's doubling is a vision statement. Anthropic's model is unusually honest about the fact that it does not know — and that honesty from the very company about to price the largest AI IPO is worth more than the hype either way. Both outcomes remain on the table. Until diffusion shows up in the income statement — in Anthropic's reported gross margin, in the cost per token as inference scales — the doubling story is a price the market is bidding, not a fact it has been shown.