AI's Trillion-Dollar Question: IMF Sees Growth Support, BIS Sees a Reckoning

Framing the disagreement precisely: the IMF's growth-supportive read is a near-term, demand-side observation - capex spending shows up in GDP whether or not it's ultimately 'worth it.' The BIS's caution is a medium-term, balance-sheet observation about whether the assets being built will generate cash flows that justify their financing cost.

What would resolve the disagreement in AI's favor: broad-based acceleration in economy-wide productivity statistics over the next two to four quarters, continued strength in hyperscaler revenue growth relative to capex guidance, and stable financing costs for data-centre debt.

What would validate the BIS's reckoning framing instead: hyperscaler capex guidance decelerating faster than revenue growth, rising credit spreads on data-centre-linked debt, or productivity statistics continuing to show no aggregate improvement despite the scale of investment already deployed.

The 19th-century railroad boom and the late-1990s telecoms buildout are the historical comparisons most invoked on both sides - both ended in overbuilt capacity and losses for many individual investors, even as the underlying infrastructure went on to support decades of subsequent growth. That pattern is consistent with the IMF and BIS both being correct at once, just about different investors and timeframes.

For a global economy already navigating a Middle East-driven supply shock, the AI capex cycle is one of the few unambiguous sources of demand-side momentum in 2026. Whether that's a genuine offset or a second bubble stacked on an already uneven recovery is likely to be one of the defining economic questions of the next two years.

Methodologically, when analysts describe AI investment 'adding percentage points to capex growth,' they are generally referring to gross fixed capital formation as measured in national accounts - the same GDP-expenditure component that includes traditional categories like factory construction and machinery. Because AI-related investment is booked in these existing categories rather than a distinct line item, isolating its specific contribution requires modelling assumptions about which capex is 'AI-attributable,' itself a source of legitimate disagreement between analysts using different classification methods.

On financing structure specifically: a meaningful share of the hyperscaler buildout is reportedly financed through a mix of retained corporate cash flow, conventional corporate bond issuance, and off-balance-sheet vehicles - special-purpose financing arrangements and vendor-financing deals between chipmakers and data-centre operators. These structures matter because they can obscure true leverage from a simple balance-sheet read, and because vendor financing in particular has historically been an early-warning sign in prior technology cycles, including parts of the late-1990s telecoms buildout.

The productivity-lag question has academic precedent: economist Paul David's analysis of electrification found measured productivity gains from electric motors took roughly three decades to show up at the aggregate level, because it required complementary changes like factory redesign and workforce retraining. A similar 'productivity paradox' was observed with early enterprise IT investment in the 1970s-80s, prompting Robert Solow's line that computers were visible everywhere except in the productivity statistics. Both episodes eventually resolved into genuine gains - but only after a multi-year lag that, if AI follows the same pattern, would place the payoff well beyond the current cycle's 2028 horizon.

Breaking the investment down by category is useful for judging durability: spending on power infrastructure is the least reversible and longest-lived, effectively a bet on sustained multi-decade demand regardless of any single AI model cycle. Spending on AI accelerator chips sits at the other extreme - a rapidly depreciating asset class given the pace of generational improvement, meaning capex tied to current-generation hardware carries meaningfully higher obsolescence risk than the power and building-shell components of the same data centre.

For a downside-scenario read: if hyperscaler revenue growth decelerated meaningfully below capex guidance for two or more consecutive quarters, the standard historical pattern in prior technology capital cycles is widening credit spreads specifically on debt tranches tied to that capex, followed by capex guidance cuts roughly one to two quarters later. That sequencing - credit spreads first, capex cuts second - is one reason the BIS, as a financial-stability body rather than a growth-accounting one, is watching financing markets as closely as the spending totals themselves.

It's also worth separating the equity-market question from the macro one, since they're often conflated in headlines. A correction in AI-related equity valuations - which trade partly on expected future earnings well beyond 2026 - is a distinct risk from a macro-level capex retrenchment that would actually show up in GDP and employment data. Historically, the dot-com crash saw a sharp equity correction, roughly 78% peak-to-trough in the Nasdaq, without a correspondingly deep recession in the broader US economy, precisely because much of the lost equity value reflected speculative valuation multiples rather than realised economic activity. Whether a future AI repricing follows that pattern - painful for investors, survivable for the broader economy - or instead triggers real capex and employment cuts at scale, likely depends on how concentrated the financing risk turns out to be within a small number of large, interconnected firms versus spread broadly across the economy.