The AI boom under the BIS lens: real revolution, opaque financing, possible overcapacity

The Bank for International Settlements does not deal in sensation. When it files investment in artificial intelligence among the pressure points that “warrant attention”, on a par with inflation and fiscal stress, you should read the source document rather than the headlines. In a thread published on 29 June 2026, Finneko (@finneko_prgrm) relays and comments on chart 13 of the BIS annual report, whose caption says the essential: corporate credit is vulnerable to a repricing should AI disappoint. Here we go back to the primary source, and test that outlook against the state of economic research on technological revolutions.

Finneko’s thread draws five ideas from the BIS that deserve to be taken seriously: a productivity gain on one task does not make a durable rise for the whole economy, the investment race is partly defensive and risks overcapacity, a bottleneck may appear on the demand side, the ecosystem’s financing is opaque, and the market’s error would be to turn a true story into financial certainty. Let us go back to the record.

What the BIS actually says

The diagnosis rests on data. BIS bulletin no. 120, by Aldasoro, Doerr and Rees, sizes the phenomenon: by mid-2025, spending on data centers and semiconductor plants equalled 1% of US GDP, and total computing-related investment reached 5% of GDP, above its 2000 dot-com peak. AI contributed 0.4 point to US growth over three years, and computing as a whole nearly half of recent growth. The trajectory is not weakening: the BIS projects annual data-center spending rising to 0.8 to 1.3% of GDP, against 0.5% today.

The hard point is financial. The five big hyperscalers plan more than $1,000 billion of AI-related capex across 2025 and 2026, a pace that exceeds their earnings and free cash flow and pushes them toward borrowing. Chart 13 shows three converging moves: investment increasingly financed by debt, rising credit risk, circular financing become commonplace. Since January 2025, the CDS premiums of AI issuers rated BBB or better have risen while the comparable broad quality index has eased: credit is starting to single these issuers out.

Circular financing, the blind spot

This is the passage Finneko foregrounds most, and he is right to press it. The BIS defines circular financing as an arrangement where hyperscalers take stakes in AI labs in exchange for purchase commitments, which sends capital back to investors as revenue. Added to this are leases on data centers built by third parties, take-or-pay capacity clauses, and off-balance-sheet commitments, where leverage does not vanish for all that. Our piece on AI circular financing details these loops.

Private credit is at the heart of the shift. According to the BIS, direct lending by private-credit funds to AI-linked firms has gone from close to zero to more than $200 billion, its share rising from under 1% to nearly 8% of the stock, with a projection of $300 to $600 billion by 2030. Yet the spreads demanded on these loans, 6.2 points against 6.1 for other sectors, are almost identical. Lenders are therefore treating AI risk as average risk, while equity valuations assume off-the-charts returns. One of two things: either credit underestimates the risk, or equities overestimate future profits. This dissonance between debt and equity is the report’s real signal. The backdrop is a private-credit market grown colossal: assets managed by these funds have risen from about $100 billion in 2010 to more than $2,200 billion today. The BIS March 2026 Quarterly Review speaks of “shadow leverage” for commitments economically akin to debt but largely held off balance sheet, and notes gross hyperscaler bond issuance topping $100 billion in 2025, at long maturities matched to the assets’ lifespan. For the mechanics of private credit, see our guide on reading private credit.

Productivity cannot be decreed

Finneko’s first point, the gap between a gain on one task and a gain for the whole economy, is one of the best-established results in economic history. In 1990, Paul David showed, with the example of the dynamo, why electrification only lifted productivity decades later: factories had to be reorganised, the central drive shaft abandoned for distributed motors. The technology was there long before its gains showed up in the statistics.

Contemporary research confirms this lag. Brynjolfsson, Rock and Syverson formalised in 2021 a productivity J-curve: general-purpose technologies demand complementary intangible investment, in reorganisation and skills, so that measured productivity first stagnates before accelerating. And Daron Acemoglu’s cautious 2024 estimate puts the total-factor-productivity gain attributable to AI at no more than 0.66% cumulatively over ten years, below the 1% threshold, far from the promises. His reasoning starts from a task-based model: as long as AI’s effect passes through cost savings at the level of each task, its macroeconomic effect follows from the share of tasks actually touched and the average saving per task. Nothing rules out a deep transformation, but it will be measured in years, not quarters.

Defensive over-investment and the lessons of bubbles

The second point, the defensive race for capital, echoes a known dynamic. Carlota Perez described in 2002 how each technological revolution passes through an installation phase where financial capital runs hot, over-invests and inflates a bubble, before a break and then a soberer deployment. Railways, electricity and the Internet bubble all combined a genuine breakthrough and an excess of capital invested too fast. Recent work by Rahil Solanki, in 2026, likens the AI circular economy to three precedents: telecom vendor financing in 2000, the off-balance-sheet opacity of 2008, and shale-oil over-investment.

The BIS itself calibrates what comes next. The end of past investment booms came with a growth slowdown of more than 1 point on average, and nothing indicates that a boom, even carried by a genuine technological advance like the Internet, leads to durably stronger growth. The sharpest setback followed the Internet bubble, modest though it was relative to GDP. The size of the shock does not depend only on the size of the boom.

The demand bottleneck

The third point, more forward-looking, stays open. If AI shifts income from labour to capital, whose propensity to consume is lower, an economy more productive in theory can run into insufficient demand. The BIS raises this risk without quantifying it; it echoes the literature on automation and the labour share, in Acemoglu and Restrepo, who documented how automation weighed on the wage share of US national income since the 1980s. A conditional, medium-term risk, that rests as much on the sharing of value as on the technology.

What it is worth

The synthesis holds in a balance. The BIS does not say AI is a bubble, and its bulletin calls the macro-financial risks moderate at this stage. But the annual report goes up a notch: a repricing, through higher rates or AI disappointment, could be as disruptive to credit as the 2008 crisis, all the more so as US households are heavily exposed to equities and non-bank institutions have become the largest holders of advanced-economy sovereign debt.

The real frictions are not abstract. McKinsey puts at close to $6,700 billion the global data-center capex needed by 2030 to keep up with compute demand, and the IEA reminds us that electricity, the grid and cooling are becoming a physical constraint as serious as financing. Add semiconductor shortages and competition that can compress margins, so many obstacles between the promise of use and actual profit.

Finneko’s contribution is to name the cognitive error: taking a true story and treating it as financial certainty. The revolution is probably real, but that compels acceptance of neither any level of capex nor any structure. For anyone who wants to judge on the record, three dials beat a thousand narratives: the gap between capital spending and free cash flow, the share of circular financing in announced revenue, and the credit premium of AI issuers. Credit often sees before equities do. This data is public, and the BIS has just published its map.

---

Primary sources:

• BIS, Annual Economic Report 2026, 28 June 2026: chart 13, debt financing, rising CDS of AI issuers, circular financing, repricing compared with 2008.
• BIS, press release on the 2026 annual report, 28 June 2026.
• BIS, bulletin no. 120, “Financing the AI boom: from cash flows to debt” (Aldasoro, Doerr, Rees), 7 January 2026: investment figures (1% and 5% of GDP, 0.4 point of growth), shift to debt, private credit, historical perspective on booms.
• BIS, Quarterly Review, box on “shadow leverage”, March 2026.
• IMF, Global Financial Stability Report, chapter 2 “The Rise and Risks of Private Credit”, April 2024.
• IEA, Energy and AI, April 2025.
• McKinsey, “The cost of compute: A $7 trillion race to scale data centers”, 28 April 2025.
• Daron Acemoglu, “The Simple Macroeconomics of AI”, NBER Working Paper 32487, 2024.
• Erik Brynjolfsson, Daniel Rock, Chad Syverson, “The Productivity J-Curve”, NBER Working Paper 25148, 2021.
• Rahil Solanki, “The AI Circular Economy: Systemic Risk, Vendor Financing, and the Keystone Problem”, SSRN, April 2026.
• Paul A. David, “The Dynamo and the Computer”, American Economic Review, 1990.
• Carlota Perez, Technological Revolutions and Financial Capital, 2002.
• Starting point: Finneko thread (@finneko_prgrm), 29 June 2026.