The number arrived without a footnote. Forty-four trillion dollars — the size Anthropic's Jack Clark says the American economy could reach by 2030, if artificial intelligence keeps its promises. I read it on a Tuesday, in the flat grey light of a Singapore afternoon, and I did what I always do when a stranger hands me a number: I opened a blank sheet and tried to make the arithmetic confess.
It took four minutes. Four minutes to discover that $44 trillion by 2030 is not a forecast in any sense I recognize. It is a covenant — a promise made in public, sealed with a signature, and payable only in a year distant enough that no one will remember who signed it. My code was the covenant, not just the contract; I learned that in 2020, auditing immutability and mistaking permanence for honesty. The $44 trillion figure has the same texture. Immutable. Unfalsifiable. Beautiful, in the way that unbuilt things are beautiful.
Jack Clark is not a careless man. He is a co-founder of Anthropic and its policy lead, and in 2025 he published a long essay, Technological Optimism and Appropriate Fear, that set out a two-track posture: extreme optimism about growth, extreme urgency about risk. The $44 trillion belongs to the first track. The layoffs-and-adaptive-policy language that travels with it belongs to the second. That they arrive in the same breath is not an accident of journalism; it is a designed pairing.
Anthropic enters this year with roughly $7 billion in annualized revenue, a valuation near $183 billion — about 26 times ARR, against a traditional SaaS multiple of 8 to 12 — some $13 billion of fresh capital raised, and compute commitments that stagger the imagination: a million-chip scale agreement with Google's TPUs, an AWS footprint, and a reported $30 billion Azure arrangement with Microsoft and Nvidia. Read those commitments as a balance sheet, and you will see why the macro story matters. Compute contracts are liabilities with dates. The revenue that services them is a curve. And a curve needs a story to justify its slope.
I have spent thirteen years watching this machinery from the inside — first as a student annotating ICO whitepapers, later as a developer auditing contracts for philosophy rather than exploits. The machinery has not changed. Only the asset class has. In crypto we call it market-making; in AI policy they call it macroeconomics. Same engine, different ticker.
Here is the arithmetic, which took four minutes and which the coverage did not perform.
American nominal GDP was about $29.2 trillion in 2024 and roughly $30.3 trillion in 2025. The Congressional Budget Office baseline puts 2030 near $37–38 trillion. To reach $44 trillion, the economy must compound at roughly 7.7 percent nominal annually from 2025 — about 5 percent real if inflation holds near 2.5. That is 3.5 percentage points above baseline, year after year, for five consecutive years.
The United States has produced five straight years of 5 percent-plus real growth twice in modern memory: the mid-1960s and the post-pandemic rebound of 2021. Neither was driven by a single technology. Neither was compressed into a five-year window. The IT revolution — the closest analogue anyone can offer — delivered perhaps 2 to 3 percentage points of productivity gain spread across a decade. The $44 trillion claim asks for two to three times that, in half the time.
Against consensus, the number does not stretch; it snaps. PwC puts AI's global GDP contribution by 2030 at $15.7 trillion. Goldman Sachs estimates roughly $7 trillion over ten years, about 7 percent of global output. McKinsey's generative-AI range is $2.6–4.4 trillion annually. If America is roughly 26 percent of world GDP, its share of those increments lands between $1.5 and $4 trillion — and $44 trillion requires $6.5 to $7 trillion above baseline. The prediction exceeds the upper bound of the consensus range by 60 to 100 percent.
The most rigorous counter-estimate is also the least quoted. Daron Acemoglu's Simple Macroeconomics of AI, published in 2024, puts AI's cumulative TFP contribution to the US economy at about 0.66 percent over ten years, and total GDP impact near 1.1 percent, with an upper bound of 2.5. He finds only about 4.6 percent of tasks economically automatable today. That is an order of magnitude below the implied claim — and it arrives with a model, assumptions, and a replication path. The $44 trillion arrives with neither.
There is a subtler problem, and it is the one I keep returning to. In 2025, American AI-related capital expenditure — data centers, accelerators, power — ran at $400–500 billion annualized, about 1.3 to 1.6 percent of GDP, and contributed roughly a full percentage point to first-half growth. That is investment entering the GDP identity, not productivity leaving it. A data center is counted when it is built; afterwards it becomes depreciation, a drag. A narrative that reads capex as productivity is reading a subsidy as yield — and I have watched that mistake hollow out a hundred treasuries.
Which brings me to the diffusion problem. General-purpose technologies do not deploy at the speed of their own benchmarks. Electrification took roughly forty years to appear in productivity statistics. Brynjolfsson's productivity J-curve describes the same lag: organizations reorganize slowly, skills re-train slowly, regulation adapts slowly. Even if model capability arrives in 2027, the complementary capital — process redesign, workforce retraining, legal adaptation — will not be fully installed by 2030. The Solow paradox was not ignorance. It was a timetable.
And the physical constraint is not rhetorical. Meeting 2030 demand requires on the order of 50 gigawatts of new American data-center power. Grid interconnection queues in most regions run three to seven years. You cannot compound your way past a transformer.
This is where the crypto reader should stop feeling superior and start feeling implicated.
In 2021 I spent a season watching liquidity mining programs print four-digit APYs. The mechanism was elegant and honest in its dishonesty: the protocol paid for its own TVL, and when the emissions stopped, the TVL left within a week. Everyone in the room knew. The number was real, the yield was real, and the users were a rumor. Liquidity mining did not measure demand; it measured the willingness of a treasury to buy a chart. Every broken token taught me how to hold value — and the first lesson was that a subsidized number is not a holding.
The $44 trillion figure has the same shape. It measures the willingness of a well-capitalized company to buy a narrative at the exact moment it needs one — before an IPO, during compute-contract negotiations, in front of legislators deciding whether to write rules that bind. And like the APY, it is unfalsifiable until the emissions stop. In 2030 the subsidy ends and we will discover who was actually there.
We have a second, quieter analogue. I have argued for two years that the data-availability layer is overbuilt — that 99 percent of rollups do not generate enough data to justify dedicated DA, and that the market priced a category before it priced a need. The AI macro story is the same error at national scale: a category, 'growth engine,' priced before the mechanism is specified. Neither claim is false. Both are premature, and prematurity is the most expensive thing you can sell.
What makes this more than commentary is that Anthropic is not only telling the story; it is building the measuring instrument. Its Economic Index, drawn from real usage data, is statistical infrastructure — an attempt to define how AI's economic impact will be counted. Whoever defines the metric defines the market. That is the insight worth carrying: the $44 trillion is not the asset. The index is. The number is marketing; the measurement is the moat. I have seen this before, in protocols that shipped a dashboard before they shipped a product and discovered the dashboard was the product.
Now the turn I owe you, because a good contrarian reading must cut against the writer as well.
Suppose the number is directionally right for entirely wrong reasons. Inflation alone, at 2.5 percent, mechanically lifts $30.3 trillion to about $34.3 trillion by 2030 — roughly a third of the journey, delivered by the unit of account rather than by any model. Add sustained AI capex and you can approach the figure while productivity barely moves. Clark would be vindicated and Acemoglu would be correct, simultaneously, which is the most dangerous outcome of all: a true headline attached to a false mechanism.
There is a reverse effect almost nobody models. If AI collapses the price of software, education, and information services, nominal GDP can be pushed down, not up. Consumer surplus is enormous and statistically invisible. A world of abundance can look, in the national accounts, like a world of stagnation — and that is before we count the demand shock of 5 to 10 percent white-collar displacement, which is negative in the short run unless you assume frictionless reallocation. The prophecy may be wrong in the direction its authors did not anticipate.
And the double edge: a prediction that cannot be settled until 2030 is a prediction that cannot be defended until 2030. If TFP contribution lands below half a percentage point in 2027–2028, the number becomes the exhibit in someone else's hearing. I learned this in the winter of 2022, when the market took 40 percent of my colleagues and every confident forecast I had written aged into evidence. In the silence of the bear, we heard the truth. The truth was that we had confused a cycle for a covenant.
One more honesty, since crypto likes to sneer: we invented this genre. We sold GDP-adjacent futures on a whitepaper and called it a network. The AI labs learned the play from us.
So here is what I am watching, and what I would ask Clark if the room were quiet.
Show me the transmission mechanism. Is the gain total factor productivity, capital deepening, or labor-force participation? The three imply different policies, different valuations, and different losers. Show me whether consumer surplus is inside the model or outside it. Show me the power contracts, because 50 gigawatts is a schedule, not an aspiration.
And then show me the index — because that is the artifact that will outlive the forecast. My code was the covenant, not just the contract. If we are going to write a covenant about the next decade of American output, I would like to see the ledger it is written in. Wouldn't you?