In the first half of 2025, roughly 92% of U.S. real GDP growth came from a single category that made up just about 4% of GDP: investment in information-processing equipment and software. That breakdown comes from Harvard economist Jason Furman, using Bureau of Economic Analysis (BEA) data. Strip those categories out, and GDP growth for the period runs at an annualized rate of about 0.1%.
Pick up the first lens — the nominal one, which puts dollars spent and dollars recovered on the same ledger. The verdict is clear: the visible return on hyperscale AI buildout lags far behind the investment. Through this lens, that 92% is a column propped up by IOUs. Growth is borrowed.
Now switch lenses. Take the same compute spending and run it through the quality-adjusted framework the Peterson Institute for International Economics (PIIE) published this spring. Deflate by the collapse in inference prices, and the "quality-adjusted output" of that same compute grows by over 2,000% per year in both 2024 and 2025 — more than twentyfold. Through this lens, the 92% isn't borrowed growth at all. It's a boom that conventional accounting systematically undercounts, because the national accounts simply don't have a ruler long enough to measure it.
Same numbers. Two lenses. Two opposite conclusions. Some read the echo of the 1996 fiber-optic bubble; others read the Solow paradox dissolving on the eve of an electrical revolution. Whichever lens you stand behind, you'll likely read the same verdict — and the problem isn't the reader. It's the lens.
This piece doesn't take sides. It dismantles one thing: the conclusion is a function of the lens, not of the data itself. And the dangerous part is that the question disguises itself as an objective report requiring no choice at all.
Two rulers, measuring two different things
The BEA ruler has no dedicated AI deflator. AI hardware, software, and cloud services sit folded into existing product categories, deflated by indices built for technologies that moved far more slowly. BEA's own research papers concede that AI activity is hard to isolate in GDP statistics. So when you see a report showing a negative internal rate of return (IRR) on AI investment — the person reading that minus sign did nothing wrong. They have only this one ruler, and its markings were cut to the rhythm of a previous generation of technology.
The PIIE ruler asks a different question. Not "how many dollars in, how many out," but "how much more can one dollar of compute do this year than last." Per-token prices fall roughly 97% a year — about a 35× efficiency gain — so a dollar buys more than ten times the inference it did a year ago. Strip out that "price decline" component, and the remaining "quality-adjusted output" explodes exponentially.
Each ruler is internally coherent. The reader who sees a negative IRR through the BEA lens and the reader who sees twentyfold growth through the PIIE lens are looking at the same negative. Only the developing chemicals differ — and the photograph comes out reversed.
The third lens: turn the gun around
The first two lenses each have their blind spot. BEA measures who made money; its blind spot is who saved money — when AI passes its gains downstream as steep price cuts, that enormous consumer surplus never appears on its scale. PIIE measures how much capability grew; its blind spot is how much of that capability gets captured and paid for. Between capability and revenue, a bridge is missing.
But the third lens is the sharp one. It takes PIIE's own preferred lens and turns it back on the side PIIE would rather not illuminate.
PIIE quality-adjusts only the output side. Inference prices crash → quality-adjusted output soars → output is undercounted. But what if you symmetrically put the same lens on the input side — marking up the book value of capital stock, since a dollar of GPU compute this year is also several times stronger than last year's?
Total factor productivity (TFP) is a residual: total output growth, minus capital's contribution, minus labor's, with technological progress as what's left over. Quality-adjust capital as aggressively as you quality-adjusted output, and the marked-up capital stock balloons, swallowing the entire residual. The measured TFP then falls — possibly turning negative.
This isn't hypothetical. Studies of early-20th-century electrification dissected it precisely: aggressively quality-adjust capital, and the residual method reclassifies all "progress" as capital deepening, making TFP vanish in the data. "Technology racing, total factor productivity falling" isn't a paradox. It's an artifact of adjusting one side and not the other.
PIIE's authors know this. They label their estimates an upper bound, not a point estimate — a first approximation. That twentyfold figure holds only when every assumption lines up and you quality-adjust output alone.
So the third lens exposes not PIIE's math — the math is fine — but its framing choice: which side gets the lens, and which doesn't, is itself a silent inference.
History's reassurance comes with two conditions
You might be thinking: history has shown that even when micro-level investors are wiped out, the macro economy thrives anyway. Half of that is true. The other half rests on premises that don't hold for AI.
Railways. Britain's 1840s Railway Mania saw share prices spike then crash by some 60%, with hundreds of companies going bankrupt. Yet by Crafts's restatement of Hawke's estimates, the social rate of return on British railway investment was about 15%, against a private return of roughly 5%. In the very years micro investors bled out, society at large was making a fortune.
Electricity. Electrical equipment shipments surged, yet total factor productivity fell for decades — Paul David's "productivity paradox." Only after factories were rebuilt and organizational capital reorganized did TFP finally erupt in the 1920s. Decades separated the technology's diffusion from its appearance in the output statistics.
Canals. Most projects of the 1790s canal mania failed, ruining their investors. But the liquidated, cheaply reorganized waterways carried decades of cheap fuel to the first factories — becoming the physical foundation of Britain's Industrial Revolution.
All three cases prove the same thing: in the buildout phase of a general-purpose technology, negative micro IRR and macro prosperity are two faces of the same coin, separated only by time. But the "micro dies, macro lives" reassurance stands on two premises.
First, durable assets. Railway tracks and rights-of-way depreciate over decades or centuries; when an operator goes bankrupt, the right-of-way sells cheap and the trains keep running. AI? Under standard accounting, GPUs carry accelerated depreciation schedules of just three to five years, and each new generation crushes the value of the last; model weights have minimal salvage value against open-source substitutes. A bankrupt railway still runs trains. A bankrupt AI operation leaves behind chips that probably aren't worth powering on.
Second, natural monopoly. Once a canal was dug, no equivalent route existed before the railway; once rails were laid, geographic monopoly formed. AI runs in virtual space, unbound by geography — a frontier model faces every competitor on earth, including open-source substitutes at a fraction of the price. The money micro investors lose settles onto a track with no moat.
Durability and monopoly — railways had both, canals had both, electricity had them partly. AI? Under the currently observable industry structure, the evidence for both premises is thin. Not necessarily absent forever — organizational reorganization, industry standards, state-level compute consolidation could each partly substitute for the old premises in the future. But that's a possibility with open variables, not a realized fact. Cover a structural gap with a variable, and the moment that variable shifts, the whole chain of reassurance has to be recomputed.
The lens's verdict, or yours
Stack the three lenses. The first says growth is borrowed — blind to who saved money. The second says the boom is undercounted — blind to the unbuilt bridge between capability and revenue. The third says even the framework claiming to fix those blind spots is making a silent framing choice of its own.
History's reassurance is withdrawn too. "Micro loses, macro wins anyway" is well documented — but it depends on durable assets and natural monopoly, and AI holds neither.
So if you ever believed "this boom is real," or "this growth is borrowed" — the question worth asking may no longer be which number is right. Across the lenses we've swapped layer by layer, the same numbers never once delivered a verdict.
And the version you believed — did the data tell you that, or did the lens you chose?
No comments:
Post a Comment