The $44 Trillion Figure Is a Valuation Narrative, Not an Economic Forecast

Anthropic co-founder Jack Clark recently put a number on the table: AI could push US GDP to roughly $44 trillion by 2030. The figure traveled fast through crypto and tech media. Almost nobody asked what it would take to be true. That is the tell. A prediction with no model, no assumptions, no confidence interval is not an economic forecast. It is positioning. I have spent enough time inside central bank working groups to know what a credible macro projection looks like. It has a baseline, a transmission mechanism, and a stated error band. The $44 trillion claim has none of these. What it does have is a speaker whose employer is preparing for a public listing, and an audience primed to hear good news. Start with the arithmetic, because the arithmetic is not ambiguous. US nominal GDP closed 2024 near $29.2 trillion. The Congressional Budget Office baseline puts 2030 somewhere around $37 to $38 trillion. To reach $44 trillion, the US economy would need roughly 7.7% nominal compound annual growth from 2025 through 2030. Assume 2.5% inflation, and that still implies real growth above 5% every single year for five years. The United States has done that twice in the post-war era: the mid-1960s boom and the 2021 reopening. Neither was driven by a single technology. Both were driven by fiscal and monetary conditions that no longer exist. Put differently, the $44 trillion figure requires roughly $6.5 to $7 trillion in additional output above baseline, a 17 to 18% upward revision to the entire US economy inside half a decade. Now compare that to what serious modelers have published. PwC estimated AI adds about $15.7 trillion to global GDP by 2030. Goldman Sachs put the ten-year global increment near $7 trillion. McKinsey sized generative AI at $2.6 to $4.4 trillion annually. Allocate the US at roughly a quarter of world output, and the consensus band for American AI upside by 2030 lands between $1.5 and $4 trillion. Clark's number sits 60 to 100% above the top of that band. Then there is Daron Acemoglu at MIT, who did the unglamorous work of building the actual task-level model. His 2024 paper estimates AI contributes about 0.66% to US total factor productivity over ten years, translating to roughly 1.1% cumulative GDP growth, with a generous upper bound near 2.5%. He found only around 4.6% of tasks economically viable for automation given current cost structures. The gap between that and $44 trillion is not a difference of opinion. It is an order of magnitude. Here is the part that gets lost in the headline. A meaningful slice of AI's current GDP contribution is capital expenditure, not productivity. Data centers, GPUs, power infrastructure, cooling: that spending is booked as investment and counts in the national accounts. Annualized, US AI-related capex is running $400 to $500 billion, roughly 1.3 to 1.6% of GDP, and it contributed about a full percentage point to growth in the first half of 2025. Code enforces what contracts cannot, but accounting rules enforce what optimism cannot: a dollar of capex adds to GDP once, then becomes a depreciation drag. Building the factory is not the same as running it efficiently. This distinction matters because the $44 trillion narrative quietly converts a capex cycle into a productivity miracle. Those are different phenomena with different durations. Capex is a pulse. Productivity is a ratchet. One of them can reverse in a single earnings season. The historical record on general-purpose technologies is equally inconvenient. Electrification took roughly four decades to show up in productivity statistics. The Solow paradox held for two decades after computers arrived in offices. Brynjolfsson's productivity J-curve describes exactly this pattern: organizational restructuring lags capability by years, not months. Even if frontier models hit every capability milestone by 2027, reengineering enterprise workflows, retraining a workforce, and adapting regulation inside three years is not a schedule any economy has ever kept. My own modeling work on central bank digital currencies showed programmable money could shorten the monetary policy transmission lag by roughly 15%. Useful, but incremental. Nobody in that room proposed that a new settlement rail would add seven trillion dollars to national output in five years, because the people in that room understood the difference between mechanism and outcome. So why this number, and why now? Because Anthropic closed roughly $13 billion in funding at about $183 billion valuation, on annualized revenue near $7 billion. That is a multiple around 26 times ARR, versus the 8 to 12 times that traditional software commands. Sustaining that multiple requires a specific belief: that the addressable market is not software, but the entire economy. National GDP displacement is the terminal node of that logic chain. If AI instead delivers Acemoglu's 1.1% increment, the addressable market contracts by an order of magnitude and the multiple does not survive contact with a spreadsheet. The same narrative does work on two other fronts. Anthropic has committed to compute at staggering scale: TPU capacity with Google measured in millions of chips, alongside AWS and a roughly $30 billion Azure arrangement involving Microsoft and Nvidia. Those commitments are underwritten by long-term revenue curves. Revenue curves are underwritten by market penetration assumptions. Macro forecasts feed the whole stack. And ahead of a listing, the single most valuable sentence an investment banker can say is not 'this is a software company' but 'this is a productivity infrastructure layer.' I have run this playbook from the other side. In 2021, I assessed the NFT boom through a liquidity lens rather than a community lens, and concluded that retail speculation had decoupled from utility. The signal was not in the enthusiasm. It was in the structure of who needed the enthusiasm to be true. The same read applies here. A prediction that requires no cost, accepts no downside, and cannot be tested until 2030 is not a forecast. It is an option written against the credibility of everyone who repeats it. Which brings me to the contrarian point. The consensus critique is that Clark is too optimistic. I think the more interesting problem is that he may be describing the wrong variable. If AI dramatically cheapens software, education, legal research, and diagnostic services, the effect on nominal GDP could be deflationary. We would be living through the largest consumer surplus event in modern history while the headline number underperforms, and commentators would call it failure. GDP measures what we pay for things, not what we get. A technology that makes intelligence abundant may show up in the accounts as slower growth, not faster. An unemployment shock in white-collar roles would compound this: displaced engineers and analysts cut consumption before they retrain, and demand-side drag arrives before supply-side gain. Physical constraints sit alongside. US data centers likely need well over 50 gigawatts of new capacity by 2030. Grid interconnection queues currently run three to seven years in most regions. No narrative compresses that queue. None of this makes AI unimportant. It makes the number a poor instrument. Yields dissolve; infrastructure remains. The infrastructure being built right now is real, the compute commitment is real, the deployment is real. What is not real is the precision of a figure pulled from a conversation and rendered as economic fact. The honest version of Clark's argument would be: AI is a general-purpose technology whose measured contribution will lag visible capability by years, and the policy question is how to shorten that lag. That argument is defensible. The $44 trillion version is a fundraising document wearing a forecast's clothing. Watch what Anthropic's own Economic Index reports over the next eighteen months. Whoever defines how AI's economic impact is measured defines the terms of the debate itself. That is worth more than any single projection. When the data arrives, the question will not be whether the model was right. It will be who gets to write the baseline.