The number arrived without an appendix.
No model. No assumptions. No confidence interval. No publication date on the relay. Just a figure โ forty-four trillion dollars, the projected size of the United States economy in 2030, attributed to Jack Clark, co-founder and policy lead at Anthropic โ carried through a crypto news brief that ran a handful of paragraphs and spent one of them on the byline.
I have spent enough of my working life reading documents engineered to be skimmed and believed to recognize the shape of this one. In 2017 I audited more than fifty initial coin offering whitepapers in a single quarter. The fraudulent ones shared a signature: they were specific where specificity was free โ token names, advisory-board logos, roadmaps with quarters printed on them โ and vague exactly where the money lived. Supply schedules. Vesting cliffs. What the treasury would actually be spent on. The vagueness was never sloppiness. It was load-bearing.
This brief carries the same signature, inverted. It is vague where precision would cost something, which is the methodology, and specific where precision costs nothing: a large round number, a round decade, a famous company.
That is not grounds for dismissal. It is grounds for reading the artifact correctly. The $44 trillion figure is not primarily an economic forecast. It is a capital instrument, and the productive question is not whether it is true but what it is collateralizing.
The artifact and its chain of custody
Start with chain of custody, because in forensic work the custody chain tells you more than the exhibit.
The claim originated, as best the public record allows, in a spoken setting โ an interview or a podcast appearance, the kind of long-form conversation in which a policy executive is invited to sketch a future rather than defend a regression. It was then relayed by Crypto Briefing, a digital-asset outlet, not by a macroeconomic research desk and not by a technology-policy journal. The relay carried no direct quotation, no surrounding context, no definition of terms.
Three omissions matter more than the number itself.
First, there is no release date. Without one, I cannot tell whether this is a fresh claim or an older one resurfacing, and I cannot calibrate it against the baseline GDP that prevailed when it was made. A forecast anchored to a stale baseline is a different object than one anchored to the current one.
Second, there is no methodological disclosure. No model, no assumptions, no error band. In serious macroeconomic forecasting this is disqualifying. A point estimate presented without a distribution is not a forecast; it is a posture. For all anyone knows, the confidence interval spans ten trillion dollars in either direction โ which would make the figure compatible with almost every position in the debate, including the pessimistic ones.
Third, and most consequentially, the units are undefined. Nominal dollars or constant 2025 dollars? Full-year 2030 or the fourth quarter of 2030 annualized? The distinction is not pedantry. It is roughly a third of the entire claim.
Here is what that means arithmetically. If the figure is nominal, then ordinary inflation does an enormous amount of the work for free. Take a 2025 nominal base near $30.3 trillion and apply a 2.5% deflator for five years: you arrive mechanically at roughly $34.3 trillion without a single new unit of productivity. Roughly one-third of the path to $44 trillion can be walked by inflation alone. Only the remainder requires an actual economic story. The brief never tells us which story it is telling, and that silence is the single most revealing thing in it.
Reading the code that writes the culture means paying attention to what a document refuses to specify. The refusal is the message.
Where Anthropic actually stands
The claim did not emerge from a vacuum, and it did not emerge from a neutral party. Both facts are necessary context, and neither is disqualifying on its own.
Anthropic enters this narrative cycle with a valuation near $183 billion against annualized revenue of roughly $7 billion โ a multiple north of 26 times ARR. Traditional enterprise software has historically transacted between 8 and 12 times. That gap is not a scandal; it is a statement about expected future growth, and it is a statement that has to be continuously re-argued to the people holding the paper.
Alongside the valuation sits a financing event of roughly $13 billion and a set of compute commitments that would have been unimaginable outside this decade: TPU capacity at Google measured in millions of chips, an AWS relationship, and an Azure commitment involving Microsoft and Nvidia reported in the neighborhood of $30 billion. Those commitments have payment schedules. Payment schedules require revenue curves. Revenue curves require a story about how far into the economy this technology reaches.
That is the structural position. Now the doctrinal one.
In 2025, Clark published a long essay โ Technological Optimism and Appropriate Fear โ that established a two-track frame: an extremely bullish growth narrative on one rail, an extremely urgent risk narrative on the other. The $44 trillion belongs to the first rail. The accompanying language about worker displacement and "adaptive policy" belongs to the second. Their appearance in the same breath is not a coincidence of temperament. It is an architecture.
The optimistic rail purchases political and capital legitimacy for the industry. The risk rail purchases a seat at the table where the rules get written. A company that says only "this will be enormous" is a lobbyist. A company that says "this will be enormous and also dangerous, and we are the serious people to manage the danger" is a co-regulator in waiting. The second sentence is what converts the first from advocacy into governance.
And why did a crypto outlet carry it? Because the digital-asset press has spent this cycle desperately renting the AI narrative. Agent tokens, decentralized-compute tokens, verifiable-inference tokens โ the entire category trades on the premise that machine intelligence is the next macro wave and that crypto will capture some of its settlement layer. A bullish macro number about AI is, for that audience, a bullish number about themselves. There is an amplification bias baked into the relay, and it runs in exactly one direction.
The arithmetic, which is where forecasts go to die
The claim is falsifiable. That is its one genuine virtue, and I intend to use it.
Take the publicly available baselines. United States nominal GDP was approximately $29.2 trillion in 2024 and approximately $30.3 trillion in 2025. The Congressional Budget Office baseline for 2030 sits near $37 trillion to $38 trillion. That baseline implies a nominal compound annual growth rate in the neighborhood of 4.2%.
Now run the $44 trillion claim through the same machinery. Reaching $44 trillion from $30.3 trillion by 2030 requires a nominal CAGR of roughly 7.7%. Strip out an assumed 2.5% inflation and you are left with real growth north of 5% sustained for five consecutive years.
That is the number that should stop the conversation, and it is the number the brief never prints. The United States has run five consecutive years of 5%-plus real growth perhaps twice in the post-war record โ the mid-1960s expansion and the post-pandemic rebound โ and never once under the propulsion of a single technology inside a five-year window. Seven point seven percent nominal is not an optimistic forecast. It is a regime change in the national accounts.
Set that against what the serious modeling community has published.
PricewaterhouseCoopers estimates AI adds roughly $15.7 trillion to global GDP by 2030. Goldman Sachs estimates roughly $7 trillion in global incremental output over a decade, about 7% of world GDP. McKinsey puts generative AI's annualๅข้ โ annual increment โ at $2.6 trillion to $4.4 trillion globally. The United States accounts for roughly 26% of world GDP, so a mechanical allocation of these estimates lands American AI-attributable increment somewhere between $1.5 trillion and $4 trillion by 2030.
The $44 trillion claim requires an increment of roughly $6.5 trillion to $7 trillion above baseline โ an increase of 17% to 18% over the CBO path. That is 60% to 100% above the upper bound of the consensus range. It is not a heterodox estimate. It is a different genre.
And the strongest counter-evidence is not from skeptics with an ax to grind. It is from Daron Acemoglu at MIT, in The Simple Macroeconomics of AI (2024), who modeled a cumulative total factor productivity contribution of about 0.66% over ten years and a cumulative GDP uplift of roughly 1.1%, with an upper bound of 1.5% to 2.5%. His estimate rests on a specific, checkable finding: only about 4.6% of tasks are economically viable for automation given current cost structures and error rates.
The distance between 1.1% and the implied path of the $44 trillion claim is more than an order of magnitude. One of these is a model. The other is a sentence.
The accounting trap nobody wants to name
There is a subtler problem, and it is the one I would press hardest if I had Clark in a room with a whiteboard.
In 2025, American AI-related capital expenditure โ data centers, accelerators, power infrastructure โ ran at an annualized $400 billion to $500 billion, or roughly 1.3% to 1.6% of GDP. That spending contributed on the order of one full percentage point to first-half 2025 GDP growth.
Read that again, because it is the hinge of the entire debate. A full percentage point of measured growth came from building the capacity to do AI, not from doing AI.
Capital expenditure is counted in GDP when it is spent. It is a level shift, not a slope change. Once the data centers are built, that spending stops contributing to growth and begins contributing to depreciation โ a drag on future measured output. If the $44 trillion narrative quietly absorbs a capex boom and books it as a productivity dividend, then the forecast is not wrong so much as it is a category error dressed as a projection. It is measuring the construction of the factory and calling it the output of the factory.
This is precisely the kind of distinction that gets flattened in a four-paragraph relay, and it matters enormously for anyone allocating capital. A capex-driven GDP bump has a completely different duration profile from a TFP-driven one. The first decays. The second compounds. Confusing them is how you end up long the wrong asset at the wrong point in the cycle.
The diffusion problem, which is the oldest objection and still the best
General-purpose technologies do not deploy at the speed of their capability curves. They deploy at the speed of organizational reconstruction, and organizational reconstruction is slow for reasons that have nothing to do with silicon.
Robert Solow's productivity paradox โ you can see the computer age everywhere but in the productivity statistics โ was not a joke about measurement. It was a description of a lag. Paul Brynjolfsson's productivity J-curve formalizes the same observation: intangible complements such as process redesign, worker retraining, and management practice have to be built before the technology shows up in the aggregate data, and the building takes years. Electrification took roughly four decades to appear in productivity statistics, and the delay was not because the dynamo was underpowered. It was because factories had to be rebuilt around the unit drive rather than the line shaft.
Now compress the timeline. Even if model capability clears every threshold by 2027 โ a generous assumption โ enterprise process reengineering, workforce reskilling, and regulatory adaptation cannot plausibly complete at national scale by 2030. The claim requires not just invention but saturation, and saturation is a sociological process measured in business cycles, not product releases.
Then there is the physical constraint, which is the least glamorous and most binding.
Meeting an AI-driven growth path by 2030 would require on the order of 50 gigawatts or more of additional United States data-center power. Grid interconnection queues in most American markets currently run three to seven years, and in the tightest regions longer. You cannot CAGR your way past a queue. A forecast that ignores the interconnection queue is a forecast about software written by people who have never waited for a transformer.
What the number is actually for
Having established that the figure does not survive its own arithmetic, the interesting question becomes: what work does it do?
Three markets, simultaneously.
Financing. A 26-times-ARR multiple requires a story in which the addressable market is not merely large but civilizational. The GDP-increment narrative is the terminal node of that logic chain. If AI's true contribution is Acemoglu's 1.1%, the total addressable market contracts by an order of magnitude and the multiple has no defense. The forecast is not decoration on the fundraising; it is a load-bearing wall of it.
Compute procurement. Commitments measured in billions of dollars per year require counterparties to believe in a long, steep revenue curve. That curve is underwritten by a penetration narrative. When Anthropic negotiates TPU allocations or Azure capacity, the macroeconomic story is not small talk โ it is the credit support.
IPO positioning. The single most valuable reframing available to a model company is the shift from "software vendor" to "infrastructure of national productivity." Infrastructure multiples are different from software multiples because the story is different. The $44 trillion figure is the linguistic mechanism of that reframing. It is doing the work that a rate-case filing does for a utility.
I have watched this pattern before, and I have watched it fail. During DeFi Summer 2020 I ran a research desk through twelve consecutive protocol deep-dives, and the tell was always the same: the teams that talked about total addressable market instead of emissions schedules were the ones whose tokens did not survive the following spring. We pulled five million dollars of subscriber exposure out of Curve-linked farms days before the token broke, not because we were clairvoyant but because we had learned to weight the mechanism over the narrative. The mechanism is the emissions schedule. The narrative is the TAM slide. Every cycle, the same substitution.
There is a fourth function, less discussed and more durable than the other three: the measurement regime.
Anthropic's Economic Index โ a project that measures AI's economic footprint using real Claude usage data โ is not a research curiosity. It is a claim on the definitional layer. Whoever establishes the instrument by which AI's economic impact is counted establishes the frame within which every subsequent argument is conducted. That is worth more than any single forecast, because a forecast expires in 2030 and a measurement standard does not. Reading the code that writes the culture means noticing when a company stops arguing about the number and starts building the ruler.
The regulatory moat hiding inside "adaptive policy"
The most quoted phrase from this episode, after the dollar figure, is "adaptive policy." It reads like a description. It is a request.
The subtext of adaptive is directional: policy should adapt to the pace of technology, which in practice means refrain from hard ex ante constraints, rather than constrain the pace of technology, which in practice means pre-deployment approval. Those are opposite architectures, and the choice between them determines the shape of the industry for a decade.
Anthropic's disclosed positions are internally consistent in a way that deserves more attention than it gets. Support for chip export controls. Support for frontier-model transparency reporting. Opposition to a federal moratorium on state-level AI legislation. Advocacy for scalable oversight. Every one of these positions raises the compliance burden on frontier training while leaving the company's own capability expansion unconstrained. For a firm that is neither the best-capitalized nor the best-compute-endowed but plausibly the strongest safety brand, that is not hypocrisy โ it is optimal strategy.
I have written about this mechanism in the context of exchange proof-of-reserves exercises, and the underlying logic transfers cleanly. Compliance architecture is never costless and never neutral. It is a filter, and filters are worth more to the largest incumbent than to any challenger. When a regime requires continuous attested liabilities, the exchanges that can afford the attestation infrastructure gain a moat and the ones that cannot exit the market; the customers absorb the cost either way. The same structure applies here. A world of pre-deployment licensing favors the entity with the largest legal department, the most mature safety documentation, and the most established relationships with the regulators writing the license. That is not an argument against safety rules. It is an argument for reading who benefits from their specific shape.
This also explains why "adaptive" carries direct commercial value. Anthropic's enterprise customers concentrate in financial services, healthcare, and legal โ the three most heavily regulated verticals in the economy. Under a post-deployment accountability regime, sales cycles stay short and growth curves stay steep. Under a pre-approval regime, every enterprise deal inherits a regulatory review and the curve flattens. A single adjective in a podcast is doing measurable work on the revenue model.
The contrarian read: the number may be too low, and it still would not mean anything
Here is where I part company with both camps.
The skeptics are attacking the $44 trillion figure with better arithmetic. They are right on the arithmetic and wrong on the frame. The optimists are defending it with capability curves. They are right about the curves and wrong about the instrument.
Both sides are assuming GDP is the correct measuring rod. It is not, and the assumption is doing more damage to the discourse than any single wrong number.
Consider what AI does to prices. A large share of AI's value to consumers arrives as consumer surplus โ free or near-free capability substituting for paid services in software, education, legal information, diagnostic triage, translation, and routine analysis. Consumer surplus is, by construction, largely invisible in GDP. Now push further: if AI meaningfully deflates the price of software, information services, and administrative labor, then nominal GDP could be lowered by the very technology that is improving welfare. A successful AI deployment and a disappointing GDP print are not mutually exclusive. They are, in the deflationary scenario, the same event.
This cuts both ways and it cuts deep. It means the $44 trillion figure could be simultaneously too low as a welfare statement and unachievable as an output statement, and neither of those facts would resolve the argument. It means the entire debate โ the CBO baselines, the McKinsey ranges, the Acemoglu task shares, the Clark headline โ is being conducted on an instrument that is being actively destabilized by the thing it is measuring.
Which brings me to the structural property of the claim that I find most telling. It is a prediction with no cost, no risk, and no settlement date until 2030. That is not an accident of timing; it is the defining feature of the genre. A forecast that cannot be marked to market for five years functions as a free option written against the future credibility of anyone who might later want to argue otherwise. It is, in the precise sense, unaccountable โ not because the speaker is dishonest, but because the instrument has no expiry and therefore no discipline.
Compare the accountability structures. Acemoglu published a model with assumptions that can be checked, reproduced, and falsified next quarter. Clark offered a sentence with none of those properties. The asymmetry is not about intelligence or integrity. It is about the fact that one party is producing knowledge and the other is producing positioning, and the market currently pays more for the second.
And watch the downside case, because it is rarely priced. If 2027 and 2028 data show AI's TFP contribution running below half a percentage point, the $44 trillion number becomes the most quotable artifact in the hands of every regulator, journalist, and competitor who wants to argue that the industry oversold. High-decibel macro forecasting is a levered position on your own credibility. The same sentence that helps price an IPO in 2026 can be read back into the record in 2029, and it will be read back by people who do not care that it originated in a podcast.
There is a third scenario that neither camp discusses and that I think is more probable than either: an AI capital expenditure cliff. If the growth attributed to AI in 2025 and 2026 is substantially capex rather than productivity, then the depreciation tail arrives on a schedule. Data centers built in 2026 depreciate regardless of whether the productivity shows up in 2028. A market that has priced a productivity narrative and receives a depreciation narrative instead does not correct gently. Navigating the storm to find the steady current requires knowing which of those two you are actually holding, and the honest answer for most AI-adjacent exposure right now โ including a great deal of the tokenized version โ is that nobody has separated the two.
What a bear market does to a number like this
We are not in the phase of the cycle where this kind of claim gets tested gently.
In a drawdown, growth narratives become scarce and therefore expensive. The marginal bid in a bear market is not for returns; it is for a reason to believe that the drawdown is temporary. Macro-scale AI optimism supplies exactly that, which is why it travels so efficiently through channels that would normally be skeptical โ including, and especially, digital-asset media, where the audience is already primed to accept that a technological step-change will reprice everything they hold.
The practical consequence is a specific failure mode I have watched repeatedly since 2017. Readers do not evaluate a macro forecast on its merits; they evaluate it on its compatibility with their existing position. A holder of AI-adjacent tokens reads "$44 trillion" as confirmation. A holder of nothing reads it as an entry signal. Neither reader asks the only question that matters, which is whether the forecast's mechanism is intact.
So here is the mechanism test I would apply, and I would apply it to any protocol making a similar claim: what is the transmission channel from the asserted capability to the measured outcome, and where in that channel can it break? For a Layer 2, the channel runs from proving cost to throughput to fee revenue, and it breaks at the proving cost line โ a fact that has been true for three years and that most operators still will not price into their runway. For an AI growth forecast, the channel runs from capability to task automation to process redesign to measured output, and it breaks at the third link, in the organizational reconstruction, which is the slowest and least glamorous node in the chain. Every forecast has exactly one link that everyone skips. Find the skipped link and you have found the falsification condition.
The takeaway
The $44 trillion figure will not be settled in 2030, because by 2030 the people who cited it will be citing something else. That is how these instruments work. They are not held to maturity; they are rolled.
What can be settled is the shape of the claim. It requires a 7.7% nominal growth path that the United States has essentially never run, an increment 60% to 100% above the consensus upper bound, a productivity transmission roughly an order of magnitude larger than the best available academic model, and a diffusion speed that contradicts everything we know about how general-purpose technologies enter the national accounts. Any one of those is arguable. All four together are a marketing document.
What I would actually watch is narrower and more diagnostic. First, the capex-to-revenue ratio in AI infrastructure through 2027, because that number tells you whether you are looking at a productivity story or a construction story, and the two have opposite duration profiles. Second, whether "adaptive policy" migrates from interview transcripts into legislative text, because that migration is the actual event being priced. Third, whether Anthropic's Economic Index publishes its methodology in reproducible form, because a ruler you cannot inspect is not a ruler โ it is a brand.
And beneath all of it, the question that the industry keeps declining to answer: if the value of machine intelligence arrives primarily as deflation โ cheaper software, cheaper analysis, cheaper expertise โ then the welfare gain and the GDP number will move in opposite directions, and we will have spent five years arguing about a metric that was never designed to see the thing we care about.
Which raises the only question worth carrying forward. When the productivity data finally prints in 2028, will anyone go back and check the arithmetic โ or will the number simply have been rolled into the next round, at a slightly higher valuation, on a slightly longer horizon?
Navigating the storm to find the steady current has never required a bigger number. It has required a smaller one, defensible under inspection. That is the whole discipline, and it is the one thing no forecast has ever been able to buy.