How the 2.7 figure is produced matters more than how large it looks. In "What Investment Data Implies About the AI Transition," Wharton's Jessica A. Wachter and Point72's Jonathan Wachter treat the capex commitments as the observable and solve backwards for the productivity gain that would make an optimizing firm spend that much, calibrating investment through end-2027 to a model of rare booms [2][7]. The paper's causality runs the same direction: an unanticipated jump in productivity, to which firms respond with a surge in investment [17]. Jessica Wachter's own summary is that nobody foresaw the productivity boom and everybody is now playing catch-up [5]. The multiple, then, is a reading of cash flows rather than an independent estimate of what the models will do, which means anyone defending the capex as rational has adopted the forecast whether they ever published it or not.
The scale is clearer in ratios than in headline totals. Hyperscaler AI infrastructure spending goes from $155 billion in 2022 to a forecast $755 billion in 2026, about 4.9 times in four years [3][1]. Five smaller builders, xAI, CoreWeave, Crusoe, IREN and Lambda, add a forecast $95 billion in 2026, roughly an eighth of the hyperscaler number and about $850 billion combined [4][2]. The majority of the total sits on five public balance sheets, which is what turns a sector question into a credit question [3].
Wachter's comparison set is the most useful part of the work for a doubter. The closest analogue she names to a capex-led buildout is late-1990s fiber optics, which she estimates implied a productivity gain of roughly 1.3 to 1.5 times [13]; against the 1.4 midpoint, the AI calibration asks for about 1.9 times as much per boom [3]. The railroad era did reach 2.8 times, but over 60 years, which compounds to roughly 1.7 percent a year [11][7]. The East Asian growth miracles reached 8 to 13 times, each over 25 to 30 years [12]. The model wants a multiple near the top of the historical range on a schedule nothing in that range ran.
The scenario weights deserve the same attention. Each of two years carries a 50 percent chance of a further boom, with any additional booms landing in 2029 and 2030 [6]; read as independent draws, that places the moderate case at 25 percent and at least one further boom at 75 percent [5]. These firms largely build out of operating cash flow, which is why the bankruptcy language demands scrutiny rather than a nod. But the material available to us breaks off at the section asking whether the multiple is achievable, and it never sets out the route from a missed boom to insolvency [16]. So carry the solvency conclusion as Wachter and Wachter's, attributed [1], and the 2.7 as arithmetic anyone can check.
What the model does supply is the width of the fan. The initial boom alone implies about 5 percentage points of extra cumulative GDP growth by 2030, while the scenarios run as high as 58 [8], a spread of nearly twelve to one [6]. Further out, the singularity path expects AI-sector productivity multipliers of 7.1 over 30 years and 188 over 80 years through 2110 [15]. Those long numbers are not a planning input for this quarter; the near ones are. A budget built on 5 points and a budget built on 58 differ mainly in how much fixed capacity they justify buying now, and that is the question in front of most buyers this quarter.
