The BIS Just Added a Fourth Scenario to the AI Growth Debate, and It's the One Nobody's Pricing

There is a specific kind of paper that changes a debate, and it isn't the one with the biggest number. Korinek and Suh's 100-300% GDP growth scenarios are bigger than anything in this report. What the BIS added in its 2026 Annual Economic Report is a mechanism, not a forecast, and the mechanism is the part of the AI investment debate that's been mostly absent: what happens to demand while supply gets automated.

I spent my last piece working through the Acemoglu-versus-optimists fight on AI's macro impact. That fight is almost entirely about supply. Acemoglu asks how much of GDP is exposed to automation and how much cost it actually saves. Aghion, Jones, and Jones ask whether bottleneck tasks in the production chain cap the gains even if everything else gets automated. Both are accounting exercises dressed up as growth theory, and they're more rigorous for it. Neither asks who's left to buy the output once the automating is done.

A fourth scenario

Box C of the report ("Transformative AI, long-term growth and r-star") builds a task-based growth model with AI capital automating a rising share of tasks, two classes of households (capital owners and workers), and a forward-looking free-entry condition governing whether firms keep investing in the next round of automation. Varying the scope of automation and how dependent firms' returns are on consumer demand generates four outcomes: business as usual, a bounded productivity boost, explosive transformative growth, and the one that doesn't show up in most investor decks — the demand bottleneck.

"As AI advances, automation increasingly diverts income from labour, which is spent on goods and services, into further AI investment. The consumer base could erode as productive capacity expands. Forward-looking firms, recognising the shrinking future market for AI-produced goods and services, may find it unprofitable to invest in innovating and automating the next task. Productivity stalls not because of technological limitation, but because the demand to justify further capacity expansion is missing. The demand bottleneck becomes the binding constraint."

In the model, output growth under this scenario actually rises first while supply-side forces dominate, then falls below the historical 2% trend as automation stalls out. Every displaced worker is also a lost consumer, so the spending that rewards innovation eventually shrinks. The natural rate of interest, r-star, follows the same arc: it climbs initially, then falls below its pre-AI baseline once the bottleneck takes hold, which the report notes would turn the medium-term inflation picture disinflationary rather than inflationary, the opposite of what most AI-and-rates conversations assume right now.

What I find more interesting than the model itself is that BIS didn't build this in isolation. The report cites Fornaro and Wolf (2026), who show that redistribution toward capital owners with a lower propensity to consume can open a demand shortfall and a liquidity trap in a New Keynesian setting, and Falk and Tsoukalas (2026), who show a similar demand externality can push firms to over-automate relative to the cooperative, economy-wide optimum. Three separate papers converging on the same demand-side failure mode in the same year is a sign this has moved from a footnote to a recognized strand of the literature.

Why this is a different bottleneck

It's worth being precise about what's new here, because "AI hits a bottleneck" isn't itself a new idea. Aghion, Jones, and Jones already gave us a bottleneck: their O-ring effect says that if production needs many tasks done in sequence, automating 80% of them while 20% remain manual doesn't get you anywhere near 80% of the cost savings, because the slowest task still sets the pace. That's a supply-side bottleneck. It's about what AI can technically do.

The demand bottleneck is a completely different failure mode. It doesn't require AI to hit a capability wall anywhere. It requires AI to work exactly as advertised on the supply side while the demand side can't keep up, because the income that used to flow to displaced workers doesn't fully reappear as new spending. It's closer to an old, almost unfashionable worry in economics: an investment boom that outruns the income needed to sustain it. The BIS report draws the parallel explicitly to the canal mania of the 1830s, the British railway mania of the 1840s, the electrification exuberance of the late 1920s, and the dotcom boom, noting all four shared "a genuine technological breakthrough that attracted capital in excess of what commercial returns could ultimately justify." In each case the technology was real. The bust still happened.

What the data already shows

The report doesn't leave the demand bottleneck as a pure thought experiment. It pairs the model with current evidence that the financing side of the AI boom is getting stretched, regardless of which long-run scenario plays out:

• The five largest hyperscalers are on pace to spend over a trillion dollars combined on AI-related capex from 2025 through 2026, a commitment now outpacing their earnings and free cash flow, pushing some toward debt issuance to cover the gap. Nvidia's own CEO has projected cumulative AI capex committed through 2030 in the $3-4 trillion range, by the report's own citation.
• The BIS models this with a contest framework: as competitive pressure pushes capex higher because firms believe only a small number of winners will dominate the market, the net economic surplus for the sector overall — revenue minus capex minus debt service — declines, and turns negative in the report's "AI disappoints" scenario, built on a 50% productivity shortfall with a partial financing unwind.
• Underneath that is a financing structure the report flags as opaque: chipmakers and hyperscalers taking equity stakes in AI labs and neocloud providers, who in turn commit to multi-year purchases of chips and compute, channeling capital back to the investor's own revenue line. Data center construction is increasingly outsourced to third parties leasing facilities back to hyperscalers on long, exit-claused contracts.
• Credit markets are starting to price some of this. CDS spreads on AI-related firms rated BBB and above have begun widening even as equity markets keep pricing in significant upside. Separately, direct lending funds have quadrupled their exposure to the AI and IT sector over the past five years to roughly 15% of their portfolios, with loan terms that haven't adjusted much for the added concentration risk.

None of this proves the demand bottleneck scenario is the one that plays out. It's evidence that the financing side of the boom already has the fragility the model assumes matters, regardless of which long-run growth path wins.

Why a revenue modeler reads this differently

I spend a meaningful share of my working life building unit-economics and LTV/CAC models for an AI product, and the discipline that work has taught me is uncomfortably relevant here. Every model like that has a load-bearing assumption buried in it: that the customer base keeps growing at roughly the rate that justified the original investment. The model doesn't usually advertise that assumption. It just sits underneath the LTV calculation, doing all the work, until something forces you to test it directly.

The BIS's demand bottleneck is that same assumption, tested at the scale of the entire AI capex cycle instead of one product. Acemoglu's framework asks whether AI can technically do enough tasks to matter. The demand bottleneck asks a question every revenue model eventually has to answer: who, exactly, is still earning the income that's supposed to buy what gets built. A model that assumes the buyer base keeps pace with the build-out isn't a forecast. It's an assumption wearing a forecast's clothes.

Signals to watch

Building on the framework from my last piece, here's what I'm adding to the watch list specifically because of this report:

• Hyperscaler capex-to-revenue ratios, and whether debt issuance keeps climbing faster than the revenue it's meant to fund.
• CDS spreads on BBB-and-above AI-related credits versus the broader investment-grade index — a widening gap while equity keeps rallying is an early warning, not a contradiction to dismiss.
• The growth rate of circular financing as a share of total AI-sector financing and forward revenue.
• The gap between earnings-call rhetoric on automation intent and actual sector-level employment and productivity data.
• Private credit concentration in AI and IT lending, and whether underwriting terms start reflecting the added concentration risk.

What would change my mind

Where this leaves the investment framework

I don't think this changes the tiered framework from my last piece so much as it sharpens the caveats inside it. Infrastructure still has the longest runway because the capex is largely already committed, but the report's own numbers are a reminder that "committed" and "self-financing" aren't the same thing, and debt-funded capex is a different risk profile than cash-funded capex. The application layer was already the highest-uncertainty tier; a demand bottleneck would hit it hardest and fastest, since application revenue depends directly on end-customer spending power in a way infrastructure spending, for now, does not.

The thing I keep returning to is that this scenario doesn't require a verdict on whether AI works. The canal builders, the railway financiers, and the dotcom investors weren't wrong about the technology either. They were wrong about the pace at which the economy could absorb what they built, and about who would still have the income to pay for it once the building was done. The hardest question in this entire AI investment cycle was never going to be whether the technology gets better. It's whether enough people are still earning enough to buy what it produces, and on what timeline that gets tested.

This piece reflects my own reading of the BIS's June 2026 Annual Economic Report and the underlying academic literature it cites. It is not investment advice. The demand bottleneck is one scenario among several the BIS itself models, not a prediction, and I'll be wrong about plenty of this in ways I can't anticipate today. Do your own work.