When a number lands without methodology, it is not a forecast. It is a signal.
Sometime in late 2025, a headline circulated through crypto and tech aggregators claiming that Anthropic โ the AI lab behind Claude โ had projected artificial intelligence would add $44 trillion to global GDP. The story carried no working paper, no baseline year, no production function, no task-exposure weighting, no link to original research. Just a figure, attributed to a lab, laundered through a publication that has no editorial relationship to the AI economy.
I have audited enough whitepapers to recognize this pattern. In 2017, at twenty-three, I spent forty hours manually tracing the ERC-20 implementation of Golem Network against its stated economic model, and what I found was not a fraud but a gap โ a space between what the marketing promised and what the contract could actually execute. The integer overflow in the distribution algorithm was real. So was the mismatch between a computational marketplace narrative and the code that was supposed to deliver it. That experience taught me a permanent habit: when someone gives you a number, ask for the function signature behind it.
Anthropic's $44 trillion has no function signature. And that absence is itself the most interesting data point in the story.
The figure, if taken seriously, describes something historically unprecedented โ and it does not survive contact with the arithmetic.
Let us reconstruct what $44 trillion would actually require. US nominal GDP in 2025 sits near $30.5 trillion. If the trend growth rate from the pre-AI decade โ roughly 4 percent nominal annually โ had continued untouched, 2030 US GDP would land around $37.1 trillion. Reaching $44 trillion implies an annual nominal growth rate of about 7.6 percent sustained for five consecutive years without interruption. That is an additional 1.2 to 1.4 percentage points per year layered on top of baseline, compounded, unbroken.
This is not cautious optimism. This is a projection that compresses six to seven years of normal American productivity growth into a five-year window. Compare it against the institutions that actually publish their models. Goldman Sachs, in its 2023 assessment, put generative AI's contribution at roughly 1.5 percentage points of productivity growth over a decade, translating to a 15 percent GDP level increase by around 2033. Anthropic's implied figure arrives twice as fast. Daron Acemoglu's task-exposure methodology, published in 2024, produced a cumulative total factor productivity gain of roughly 0.66 percent over ten years โ a GDP impact of about 1.1 percent. That is a sixteen-fold divergence. McKinsey's 2023 global estimate of $2.6 to $4.4 trillion in annual incremental output, apportioned to the US, implies something in the range of $0.6 to $1.0 trillion per year. The gap between that and the Anthropic headline is six to ten times.
When credible estimates for the same phenomenon span one to two orders of magnitude, the honest conclusion is not that one camp is right. The honest conclusion is that the field has no reliable measurement basis at all, and any single figure presented without its model is a rhetorical instrument rather than an economic one.
There is a second problem, more structural, and it is the one that most coverage missed entirely: the GDP contribution of AI today is overwhelmingly capital expenditure, not productivity.
In 2025, data center and information processing equipment investment contributed somewhere between half a percentage point and 1.1 percentage points to US GDP growth โ a substantial share of total growth. In expenditure-accounting terms, that is the I in C plus I plus G plus net exports. It is spending, not efficiency. And spending is reversible. The telecom equipment buildout around 2000 contributed meaningfully to growth in the late nineties, then flipped to a drag after 2001. If the AI capex cycle peaks, the GDP contribution does not gracefully decelerate. It goes negative.
The accounting gets worse. A large share of GPUs are imported, and imports subtract from GDP, which means the net domestic contribution of building data centers is systematically overstated by headline figures. Simultaneously, consumption of fixed capital rises as those assets depreciate, which suppresses net national income even as gross domestic product climbs. GDP can rise while the income actually available to households rises far less. A forecast that cites the gross figure while ignoring the depreciation offset is not measuring welfare. It is measuring activity.
Then comes the paradox that sits at the heart of every AI productivity projection, and which the $44 trillion number walks directly into. The projection is denominated in nominal GDP. But the primary economic effect of AI is to compress the price of the services it substitutes. Code, copywriting, customer support, translation โ costs in these categories have already fallen by fifty to eighty percent. If AI genuinely raises real output while simultaneously depressing prices, nominal GDP growth will run below real GDP growth. The more authentically AI transforms the economy, the harder it becomes for a nominal dollar figure to be reached. The forecast is internally self-defeating. This is what happens when a technology's deflationary nature collides with a metric that rewards inflation.
And the diffusion timeline does not cooperate either. General-purpose technologies historically take fifteen to forty years to show up in aggregate productivity statistics โ the Solow paradox, when computing was everywhere except in the productivity numbers, lasted roughly a decade. Anthropic's own Economic Index data shows that as of 2025, only about 4 percent of occupations use AI across more than half of their tasks. The METR randomized controlled trial from July 2025 found experienced open-source developers working with AI assistance completed tasks 19 percent slower โ while believing they were 20 percent faster. That perception-reality gap is the hardest evidence we have against the assumption of a five-year economy-wide productivity leap. If the people closest to the technology, best positioned to use it, systematically misjudge their own gains, the aggregate forecast built on extrapolated gains deserves severe discounting.
Which brings us to the question the number was never designed to answer, and the one that matters most: who benefits from publishing it?
Anthropic's revenue is almost entirely enterprise API and Claude subscriptions. Enterprise AI budget approval runs through executive conviction โ the belief that AI is a structural transformation rather than a departmental experiment. A $44 trillion anchoring figure, repeated often enough, lowers the psychological threshold for that approval. The lab's annualized revenue moved from roughly one billion dollars at the start of 2025 to something in the seven billion range by year's end, one of the fastest growth curves in the industry, with a heavy concentration in coding agents and enterprise API traffic. Menlo Ventures' 2025 survey placed Anthropic at roughly 32 percent of the enterprise API market against OpenAI's 25 percent โ a genuinely strong position.
But set seven billion against a $44 trillion forecast and the arithmetic is uncomfortable. If AI were to add six to seven trillion in annual output by 2030, and the leading enterprise API vendor captured even a generous fraction of that, the implied capture rate today is around one-tenth of one percent. The gap between forecast and reality spans three orders of magnitude. Meanwhile the lab has committed to hundreds of billions in multi-year compute contracts with cloud providers, meaning it must drive revenue into the tens of billions within a few years. That commitment is a financial motivation to publish optimistic narratives, and it is absent from every retelling of the headline.
The falsifiability structure is instructive. The prediction resolves in 2030. It carries no published model, no confidence interval, no decomposition between gross and net output. The publication cost is near zero. The citation benefit is enormous. The cost of being wrong is negligible, because the claim was framed as what AI "could" do. This is low-cost, high-leverage narrative engineering, and I have watched the same pattern in crypto for a decade โ token whitepapers that promised planetary-scale computation and delivered a Discord server. Hype creates noise; protocols create history.
There is a further distinction the forecast elides entirely: AI-created GDP versus AI-redistributed GDP. A large share of AI's economic value is share transfer, not net creation. When customer support migrates from a business process outsourcing firm in Manila to a software vendor in San Francisco, output moves between sectors. In aggregate GDP it may register as a price decline rather than a volume increase. We have seen the crypto equivalent many times: liquidity mining APY is not yield, it is the project subsidizing its own TVL figure. Stop the incentives and the users evaporate. Stop counting redistribution as creation and a large portion of the AI GDP story does the same.
And if the optimistic case is right โ if six to seven trillion in output really does materialize within five years โ the distributional consequence arrives before the dividend. Labor's income share compresses faster than policy can respond, and the political backlash may well dismantle the diffusion machinery before the productivity gains land. That is the risk nobody publishes, because it does not fit in a headline.
The competitive logic behind the release is legible enough. Anthropic's strategic position is narrow and deep: first-tier capability in coding, agents, and long-context reasoning, deliberately absent from multimodal generation, deliberately absent from consumer distribution, dependent on AWS, Google Cloud, and Azure for compute and channel. It has no proprietary cloud, no silicon, no consumer entry point. Publishing macroeconomic forecasts is a way to claim the "economic impact definer" narrative adjacent to the capability narrative โ differentiating from a competitor whose advantage runs through consumer mindshare. But there is a brand cost. A lab whose core asset is honesty and safety cannot casually traffic in low-verifiability optimism without paying for it in credibility. When I audited the custody architectures behind the 2024 spot Bitcoin ETF applications โ multi-signature setups and threshold signature schemes compared against open standards โ the finding that mattered was not that the cryptography was weak. It was that compliance-driven centralization had quietly redefined what the product was, and the marketing language had not caught up. The same drift applies here. The narrative and the code have diverged, and the narrative is doing the trading.
I have been through this particular disillusionment before. In 2022 I had already documented, in private notes, the brittleness of the Terra mechanism before the collapse โ and I watched the market refuse to read it. The retreat to Sรฃo Paulo afterward was not drama. It was recalibration. Reverse-engineering the UST burn logic taught me that confidence does not fail gradually; it fails at a precisely identifiable tipping point that everyone can see in hindsight and nobody prices in advance. The $44 trillion number is the same species of object. It is not a datum. It is a confidence instrument, and confidence instruments are always priced above their intrinsic value until they are not.
So what should a serious reader take from a headline that arrived without a model? Not the number. Take the fact that the number was published at all, and by whom, and at what point in the capital cycle. That is the actual information content. The forecast tells you where the lab's balance sheet needs the world to be in 2030; it does not tell you where the world will be.
If you want a working discipline from this, here it is: when any institution publishes a macroeconomic projection, ask three questions before you internalize the figure. What is the baseline and the time horizon? Is it nominal or real, and what deflator assumption does it carry? And is the reported figure gross or net of displaced output? Those three questions alone can move a projection by trillions in either direction, and the failure to answer them is not an oversight โ it is a feature of narrative construction.
Fragility is the price of infinite composability. The same holds for forecasts. A number that composes with every narrative, that requires no methodology to be cited, that carries no cost when wrong, is not robust. It is fragile exactly where it appears strongest.
The real question is not whether AI will add $44 trillion to global GDP. The real question is what happens to an industry's credibility when the gap between its models and its marketing becomes visible to the people writing the checks. That gap is closing. The data center capex is real, the depreciation schedules are real, and the import subtraction is real. The productivity numbers are not yet. Somewhere between the two lies the actual story, and it will not be told in a headline.
Hype creates noise. Protocols create history. The forecast will be forgotten; the architecture it justified will still be standing in 2030, and it will be judged on what it produced rather than what it promised.