Forty percent, annualised: the US-Europe AI capex gap is real, the bust risk sits elsewhere

The 40% is cumulative over six years, 2021 through the end of 2027, which compounds to roughly 5.8% a year in real terms ^[1][7]. Europe's figure annualises to about 1.9% ^[4][11]. The cumulative ratio is 3.3 to one ^[8], which is the number that gets quoted; the annual gap is under four percentage points, which is the number that turns up in a procurement plan. Both are accurate. Only one of them describes a boom.

The boom is somewhere else. Google, Meta, Microsoft and Amazon are lined up to spend, in a single year, about 1.25 times the entire recorded global AI funding total for 2025 ^[3][12][9]. The two figures do not measure the same thing: one is infrastructure capex at four companies, the other counts private investment, M&A, public listings and minority stakes across the whole market ^[3][14]. Treat it as a magnitude check, not a like-for-like. It still makes the point that a single-digit corporate growth trend and a small number of balance sheets are carrying very different loads.

That 2025 funding total came off a base of roughly $253 billion a year earlier ^[12][18]. After a 129.9% year, a flat 2026 would represent no decline in spending at all and would still read as a rupture ^[12]. That is the mechanism behind the BIS language: the American investment boom is tied to continued growth in AI spending, so the failure mode is not firms spending less, it is firms stopping the acceleration ^[5].

Karsten Junius, head economist at Bank J Safra Sarasin, is relaxed about the euro area on the grounds that, in his words, "AI investment in the US is not going to continue at this scale indefinitely" ^[10]. He may well be right, and that is the awkward part. The route back to parity he describes runs through American deceleration, which is the same event the BIS is calling a bust ^[5][10]. The other route runs through European capital that is not present: a more fragmented technology market with fewer companies at the scale of the largest US players, and no comparable cash flows to divert into data centres and chips ^[6]. Junius's own downside is that the average European's quality of life keeps sliding relative to Americans ^[17].

The China comparison rewards the same arithmetic. US private AI investment is put at 23 times China's ^[16], and the usual correction is Beijing's state money, estimated at $184 billion over 23 years ^[15]. Spread across those years, that averages about $8 billion annually ^[19], and the four US hyperscalers' 2026 plan is close to four times the cumulative 23-year state figure ^[20]. State support may narrow a capability gap. It does not narrow this spending gap.

So if the question is who is exposed to a reversal in AI capex, the 40-versus-12 split is not where the answer lives. Four capital budgets are.