Anthropic published an economic scenario this week. The headlines took one number: 15% annual U.S. GDP growth. The follow-up took another: the economy doubling every 4.5 years. Between those two figures sits an arithmetic problem almost nobody checked.
Run the compound growth. If U.S. nominal GDP sits near $29 trillion in 2024 and the scenario lands at $44.4 trillion by 2030, the implied compound annual rate is roughly 7.3%. Not 15%. The 15% is almost certainly a peak along the path β one year's velocity at the steepest point of the curve β not the average across the window. That distinction isn't cosmetic. It is the difference between an aggressive extrapolation and a number that is simply wrong as a description of the period.
Anthropic attached a disclaimer calling the scenario conditional and extreme. Good. The distribution layer β the part where someone tells you the probability weight sitting on that tail β never reached the coverage. And the market is already trading the headline. That is the market's chaos. Nobody reads the appendix.
Context
The framing matters. Anthropic's scenario models an economy where digital knowledge work is substantially automated, and explicitly excludes superhuman humanoid robotics. Elon Musk anchors his own doubling thesis to physical robots β Optimus in particular β extending automation from screens into factories, logistics, and services. Two mechanisms, two constraint sets, presented in coverage as one voice.
Then there is the part of the study carrying the most information and receiving the least attention: a survey of 10,980 Americans. Their typical expectation was roughly 10% in additional growth. About one in ten came anywhere near the extreme scenario.
In late 2017 I audited twelve top-20 token launches. Three contained irreconcilable inconsistencies in their supply schedules β emissions modeled against a demand curve the whitepaper never defined. The lesson was mechanical: a model that publishes its output but not its assumptions is not a model. It is a press release with decimal points.
Core
Start with growth accounting. A 15% peak for a mature economy requires decomposition β total factor productivity, capital deepening, labor input β each carrying a share that sums to something coherent. U.S. real GDP has lived in the 2β3% band for decades; postwar expansions rarely cleared 5%. The scenario asks a $29 trillion economy to outrun the fastest industrializing economy in modern history, sustained, without specifying which input does the lifting.
Then ask what happens to the two inputs the scenario needs and doesn't price.
Compute. If knowledge work is automated at the implied scale, the inference load is not marginal β it is a step function, and it translates into advanced packaging and gigawatts. AI data center power procurement has already migrated from a facilities question to a grid question. Interconnection queues and nuclear agreements run on five-to-ten-year clocks. The scenario doubles in 4.5 years. Those timelines do not reconcile, and energy supply carries a medium-term rigidity no amount of model capability compresses.
Capital. Buildout at that scale consumes capital on the order of trillions, and capital consumed is capital not allocated elsewhere. Booking the productivity upside without booking the crowding-out on the other side of the ledger is not conservative. It is incomplete.
Diffusion. Every general-purpose technology on record β electrification, the internal combustion engine, the computer β took decades to penetrate, not years. The bottleneck was never the technology; it was organizational redesign, regulatory approval, and the slow accumulation of trust. Attitudes move slower than weights. Diffusion is the most load-bearing and least disclosed parameter in the document. Move it from five years to twenty-five and the terminal number shifts by an order of magnitude.
The spread between Anthropic's tail and the surveyed median is the actual finding. Ten thousand nine hundred and eighty respondents, roughly 10% expected lift, and a professional scenario five points above them. That gap is a sentiment measurement disguised as a growth forecast β and it is measurable, which is more than can be said for the 15%.
Then the measurement problem, which is where this crosses my beat. GDP is a market-transaction instrument, and AI output lands disproportionately where that instrument cannot see: free-tier inference, open weights, household automation of tasks that were never invoiced. When a research memo takes four hours instead of forty, the productivity gain is real and the GDP line item is zero. That 32.4% excess-GDP figure is not a floor. It may not be a measurement.
I spent six months in 2026 dissecting autonomous agent-to-contract economics, and the conclusion transfers: value settles at the verification layer, not at the headline productivity number. Agents transacting without a verification primitive impose a trust tax that eats the efficiency gain. The same holds for macro scenarios. A projection with no probability weights, no sensitivity bands, and no disclosed production function is an unverified claim wearing a number's clothing.
Which brings in the crypto channel. This scenario was syndicated through a blockchain-adjacent outlet, not an AI vertical. That is distribution strategy, not accident. The AI Γ Crypto and productive-asset narratives need a macro ceiling to sell against, and a 15% scenario supplies one. Expect the number in token pitches within a quarter, stripped of its conditional framing and cited as a forecast. Watch the decentralized inference plays and the agent tokens with no verification primitive.
There is a second-order effect worth naming in crypto specifically. Tokenized compute, GPUs as yield-bearing collateral, decentralized inference markets β all of these need a demand terminal narrative. The 15% scenario functions as the top of that narrative stack. When the demand story gets priced into the asset rather than into the operator's cash flow, the unwind is not gradual. It is a liquidity event.
Contrarian
The consensus being sold here does not exist.
Anthropic and Musk are not aligned. They are adjacent. Anthropic's model has no robots, so its growth path runs entirely through software β near-zero physical constraint, heavy organizational friction. Musk's thesis is a manufacturing ramp with hard physical bottlenecks and a much larger addressable GDP share once cleared. Coverage stitched them into a single AI-doubles-the-economy chorus. That is a media artifact, not an analytical finding.
Anthropic's disclaimer is genuinely responsible. The problem is structural: the headline takes the tail, the caveat sits in the final paragraph, and the reader retains 15%. Honest work in a format engineered to defeat it.
The asymmetry is unflattering to the publishing side. Anthropic is an unlisted lab whose narrative supports its valuation, its lobbying position, and its recruiting. Musk's timeline binds to Optimus commercialization. Both have motive to round upward. Neither supplied a probability weight. The thesis held firm when the charts turned red β but nobody has shown me the chart.
Takeaway
Hold this as an upside option, not a base case. The tradeable signal is not 15% β it is the gap between 15% and the public's 10%, and the fact that the infrastructure to deliver either number sits in a power queue longer than the scenario's doubling period. Track three things: Anthropic's raw model documentation if it ships, U.S. quarterly GDP for any deviation outside the 2β3% band, and data center power contracts. Whitepaper versus technical reality has a familiar answer β a model is only as honest as its most-disclosed assumption. The only fully disclosed number in this story is 10,980.
