A $44 trillion US GDP figure by 2030 is not a forecast. It is a positioning statement wearing a spreadsheet's clothing.
Jack Clark, Anthropic's co-founder and policy chief, put that number into circulation recently. It traveled through Crypto Briefing and landed in crypto feeds within hours. No model was attached. No assumptions were published. No date of origin. No specification of whether $44T is nominal or real, annual or fourth-quarter, conditional on what. The number arrived naked, and the industry handed it clothes.
Start with arithmetic, because arithmetic does not negotiate. US nominal GDP sat near $29.2 trillion in 2024 and roughly $30.3 trillion entering 2025. The Congressional Budget Office's baseline path puts 2030 somewhere around $37โ38 trillion. To reach $44 trillion instead, you need a nominal compound annual growth rate near 7.7 percent across five years โ about 3.5 percentage points above baseline, roughly 2.5 to 3 points of extra real growth annually. That is two to three times the productivity gain of the 1995โ2005 IT revolution, compressed into half the window. The US has sustained 5 percent-plus real GDP growth for five consecutive years exactly twice in modern record: the mid-1960s and the 2021 reopening rebound. Neither was driven by a single technology.
Now the denominator problem. PricewaterhouseCoopers estimates AI adds about $15.7 trillion to global GDP by 2030. Goldman Sachs puts the ten-year global increment near $7 trillion. McKinsey's generative AI range is $2.6โ4.4 trillion annually. Allocate to the US at its ~26 percent share of world output and you land at $1.5โ4 trillion of US AI increment by 2030. The $44T framing requires $6.5โ7 trillion. The prediction exceeds the consensus upper bound by 60 to 100 percent.
Based on my audit experience โ I spent three months in 2018 dissecting the 0x Protocol v2 order-matching logic line by line, hunting integer edge cases nobody asked me to find โ I have learned that a number without its derivation is a number without its security model. You cannot stress-test a black box. You can only note that it is black.
The strongest counter-evidence is not from crypto skeptics. It is from Daron Acemoglu at MIT, whose 2024 paper Simple Macroeconomics of AI estimates AI will contribute roughly 0.66 percent to US total factor productivity over ten years and about 1.1 percent to GDP cumulative, with an upper bound of 1.5โ2.5 percent. He finds only about 4.6 percent of tasks economically viable for automation today. Clark's figure requires an effect an order of magnitude larger than that.
Here is the mechanism I suspect most readers miss entirely. US AI-related capital expenditure โ data centers, accelerators, grid interconnects โ ran at roughly $400โ500 billion annualized in 2025, about 1.3โ1.6 percent of GDP, contributing around one percentage point to first-half 2025 growth. That is investment booked into GDP, not productivity created by GDP. A data center is a level shift. Once built, it bleeds depreciation into the national accounts for a decade. The $44T narrative quietly converts a capex cycle into a productivity miracle, and those are opposite things wearing the same coat.
Then there is the deflation problem nobody prices. AI's most direct consumer benefit is cheaper software, cheaper legal research, cheaper diagnostic triage, cheaper customer support. Cheaper services lower nominal GDP. If AI meaningfully deflates the price of information work while displacing 5 to 10 percent of white-collar roles, the short-run net effect on measured output can be negative โ demand-side shock before supply-side gain. The $44T number assumes frictionless reallocation. Real economies do not do frictionless. Electrification took roughly forty years to appear in productivity statistics. Solow's paradox was not a joke; it was a measurement lag with teeth.
And the physical ceiling. US data centers need upwards of 50 gigawatts of new capacity by 2030. Grid interconnection queues currently run three to seven years in most regions. You cannot compound a narrative past a transformer shortage.
So why does this number exist, and why now? Because Anthropic closed roughly $13 billion in 2025 at a valuation near $183 billion against annualized revenue around $7 billion โ roughly 26 times ARR, against a traditional SaaS multiple of 8 to 12. That multiple cannot survive a world where AI adds 1.1 percent to GDP. It survives only if the addressable market is framed as national output itself. The $44T figure serves three battlefields simultaneously: financing, where macro-growth language quiets the 'is AI just a tool' objection; compute procurement, where Anthropic's hundred-thousand-TPU-class Google agreement, its AWS commitments, and a reported ~$30 billion Azure arrangement with Microsoft and Nvidia all require a long revenue curve to justify; and IPO positioning, where the entire game is reclassifying a model company as infrastructure.
Note the undated, non-falsifiable structure. A prediction that settles in 2030 is a prediction nobody can hold you to in 2026. Clark's 2025 essay Technological Optimism and Appropriate Fear established the dual-track frame precisely: extreme growth optimism paired with urgent risk language. The $44T is the optimism track. The job-displacement-and-adaptive-policy language is the fear track. They appear together because they work together. Optimism buys political and capital legitimacy for the industry. Fear buys Anthropic a seat at the table where the rules get written. 'Adaptive policy' is not an observation about policy. It is a request โ adapt to our tempo, do not constrain it. That is the inverse of the EU AI Act's precautionary architecture, and the distinction is worth billions in enterprise sales-cycle efficiency.
Here is what the bulls get right, and I will not pretend otherwise. Anthropic is not a vaporware shop. Its models are genuinely frontier-class. Its safety branding is the strongest in the sector, and it has converted that branding into actual regulatory access in Washington and Brussels. The Economic Index it built โ measuring AI's economic footprint from real Claude usage data โ is a smarter move than any single forecast, because whoever defines the measurement defines the debate. And the capex buildout is real: the chips are being fabricated, the buildings are rising, the gigawatts are contractually spoken for. The productivity gain will be substantial. The argument is about magnitude and timing, not existence.
But substantial is not $44 trillion. Trust is a variable; verification is a constant. Volatility is just noise; liquidity is the signal โ and right now the signal is a number with no methodology, published by an interested party, timed to a capital event, immune to near-term falsification. When 2027 and 2028 TFP data arrive, if the contribution lands under half a percentage point, that number becomes the artifact regulators and journalists reach for. High-decibel macro forecasts are a double-edged instrument: they amplify the upside story and they pre-commit you to the disappointment.
The useful question is not whether $44 trillion is right. It is what you should require before you let any number that large price anything. Demand the model. Demand the assumptions. Demand the deflator treatment, the capex-versus-productivity split, the displacement scenario. Silence in the code is where the theft hides. Likewise in the forecast.
In 2022 I traced over 500,000 ETH across Ethereum and Solana to reconstruct Alameda's ledgers, and the lesson that survived that work is simple: claims leave footprints, and footprints can be read. A GDP forecast leaves no footprint. It leaves only a headline. Read the headline as a headline, not as evidence, and you will be ahead of everyone who mistook a valuation instrument for an economic one.