Productivity or Reckoning

Issue XVII set out the arithmetic facing Washington in stark terms. Federal debt is on course to exceed its 1946 post-war peak within years, net interest costs are compounding faster than the economy that must service them, and only two levers remain: borrow less, or grow faster than the debt grows. We argued that cutting the deficit is the more reliable lever, but also the more politically constrained one, in the United States and in most of the advanced world. That leaves growth carrying more of the load than it comfortably can, and it is why the G20's own communiqué reached for artificial intelligence as the productive capacity that might do the lifting. This issue tests that proposition directly. Is AI-driven productivity growth capable of lifting profits, tax receipts and GDP by enough to bring the debt trajectory back under control, or is it another number too small for the scale of the problem?

Our approach is deliberately narrow and arithmetic rather than aspirational. We look first at what the productivity data is actually showing in the United States today, as the clear global leader in AI adoption. We then set out the range of published projections for how far AI-driven productivity growth could run over the next decade. Finally, we run those projections directly against the government's own debt sustainability arithmetic, and against the historical record of the one occasion the United States actually did grow its way down from a comparable debt burden, to see whether growth alone has ever really done the job it is now being asked to do again.

What the Productivity Data Is Actually Showing

The headline signal is genuinely positive. US labour productivity has grown at an annualised rate of 2.4 per cent since the start of 2024, against an average of 1.6 per cent in the five years before the pandemic, and the Federal Reserve Bank of Dallas finds a positive relationship between AI exposure and productivity growth at the sector level that holds up in international comparisons. Deloitte's analysis of nonfarm business output per hour shows growth of 2.6 per cent per quarter on average since 2023 — more than double the prior decade's 1.2 per cent pace — alongside a surge in tech-related capital investment. Stanford's AI Index puts the estimated US consumer surplus from generative AI tools at US$172 billion annually by early 2026, up from US$112 billion a year earlier, and finds measured gains of 14 to 15 per cent in customer support, 26 per cent in software development and 50 per cent in marketing output at firms that have adopted the technology.

Set against that is a more sobering read from inside the boardroom. A survey of nearly 6,000 senior executives across four advanced economies found 70 per cent of firms report using AI, yet 90 per cent say it has had no measurable impact on productivity or employment so far, with adoption concentrated in larger, already-productive firms whose leaders expect only a further 1.4 per cent productivity gain over the next three years. The pattern is consistent with what manufacturing research calls the AI J-curve: initial adoption slows measured productivity, as firms absorb the cost of reorganising workflows and retraining staff, before any stronger gains show through. On this reading, the encouraging aggregate numbers may say more about a handful of frontier adopters and a hot capital expenditure cycle than about an economy-wide productivity transformation that has actually arrived.

How Fast Adoption Is Actually Moving

Enterprise uptake has clearly accelerated. Around 78 per cent of organisations report using generative AI in at least one business function, up from 55 per cent a year earlier, and the enterprise AI software market is forecast to grow at a compound annual rate above 34 per cent through the mid-2030s. Yet breadth of adoption is not the same as depth of transformation. Separate research finds barely 1 per cent of companies consider themselves mature AI users even as more than 92 per cent plan to increase investment, and only 18 per cent of US firms had adopted AI in any form by the end of 2025 on Federal Reserve estimates. The picture is one of an economy at the very beginning of a diffusion curve that history suggests typically takes fifteen to twenty years to run from initial breakthrough to broad-based productivity payoff — not one already mid-cycle.

A Decade-Out Forecast That Spans a Fifty-Fold Range

This is where the debate becomes genuinely difficult to arbitrate, because the published estimates of AI's cumulative contribution to growth disagree with each other by an order of magnitude, and in some cases by more. At the conservative end, MIT's Daron Acemoglu, using a task-exposure model grounded in Hulten's theorem, estimates AI will lift US total factor productivity by no more than 0.7 to 1.1 per cent cumulatively over ten years. The Penn Wharton Budget Model has AI raising US productivity and GDP by 1.5 per cent by 2035, rising to nearly 3 per cent by 2055, with the boost to annual growth peaking in the early 2030s before fading.

We think the honest reading of this spread is not that one estimate is right and the rest are wrong, but that the underlying methodologies are measuring different things. The conservative estimates model what share of existing tasks AI can profitably perform today, extrapolated cautiously. The larger estimates extrapolate from early productivity experiments, in some cases assuming today's frontier model capabilities scale smoothly into the workforce over a decade. Both are legitimate exercises. Neither is a forecast in the sense a bond investor would recognise, and the gap between them — fifty-fold — is itself the most important fact in this debate: nobody actually knows, within an order of magnitude, how large this dividend will be.

Running the Numbers Against the Debt

The Congressional Budget Office's own scenario analysis lets us test the productivity case directly against the fiscal arithmetic that matters. Under its extended baseline, federal debt held by the public reaches 156 per cent of GDP by 2055 — already close to double the 1946 post-war peak of 106 per cent. CBO has separately modelled what happens if nonfarm business productivity growth runs half a percentage point per year faster, or slower, than that baseline for three decades. The faster case brings 2055 debt down to 113 per cent of GDP. The slower case pushes it up to 203 per cent.

Two things stand out. First, productivity genuinely matters to this arithmetic — a sustained half-point annual difference in productivity growth is worth roughly 90 points of debt-to-GDP by 2055 in either direction. Second, even the favourable case — a full half-point of extra productivity growth sustained for thirty straight years, a considerably larger and more durable effect than Penn Wharton's own central estimate for AI specifically — still leaves debt at 113 per cent of GDP, above the 1946 peak and above where the United States stands today. On CBO's own numbers, AI-scale productivity gains can meaningfully soften the trajectory. They cannot, on their own, resolve it.

The Penn Wharton Budget Model has attempted the more direct version of this exercise, estimating that AI could reduce cumulative federal deficits by around US$400 billion over the 2026 to 2035 budget window. Set against CBO's projected cumulative deficit of roughly US$22 trillion over the same period, that is a rounding error — well under two per cent of the gap. Even granting that this is an early and conservative estimate likely to be revised as adoption data accumulates, the order-of-magnitude gap between what AI is expected to contribute and what the debt trajectory requires is the central finding of this issue.

Growth Alone Has Not Done This Before

There is a natural counter-argument: the United States has grown its way down from a comparable debt burden before, from 106 per cent of GDP in 1946 to 23 per cent by 1974. If it happened once, can it not happen again through an AI-sized productivity boom? The historical record is less encouraging than the headline suggests. Recent research revisiting that episode finds that economic growth on its own, stripped of everything else that was happening at the time, would have reduced the ratio by only 22 percentage points over the following 76 years — from 106 per cent to around 84 per cent. The rest of that dramatic decline was done by primary budget surpluses sustained for decades, a burst of surprise inflation through the 1970s, and financial repression: the Federal Reserve's deliberate capping of Treasury yields below market levels from 1942 to 1951, which taxed bondholders through negative real returns rather than taxing anyone through visible fiscal measures.

None of the three tools that did the heavier lifting after 1946 are readily available, or even desirable, today. Sustained primary surpluses require exactly the spending restraint or tax increases that Issue XVII identified as politically constrained. A repeat of 1970s-style surprise inflation would be a policy failure, not a policy choice, and would be resisted by the same bond market whose patience is already being tested. Financial repression of the kind practised in the 1940s would require capital controls and forms of intervention in the Treasury market considered incompatible with an open financial system and confidence in the dollar as a reserve currency. If growth is left to do the job alone, the evidence suggests it is being asked to do roughly five times more work than it has ever done before.