Anthropic plans to spend $518 billion from a 2025 revenue base under $5.2 billion

"More than 100 times" sets a ceiling: Anthropic's 2025 revenue was below $5.18 billion ^[15]. Columbia Business School's Stijn Van Nieuwerburgh uses a 10% return as the hurdle for the US sector ^[9]. Apply it to the $518 billion and the spending would need to earn about $51.8 billion a year once it is all spent ^[16]. Anthropic took in at most a tenth of that during 2025 ^[17]. Reuters did not report how many years the plan covers. If it fell inside Van Nieuwerburgh's 2025-2032 window, one company would account for about 6% of his US estimate ^[21].

JP Morgan's Nvidia test looks modest as an annual rate and steep once compounded. Over ten years the CBO's 1.75% baseline adds about 19% to US productivity ^[18]. JP Morgan's band adds about 34% to 63% ^[19], or between 1.8 and 3.3 times the cumulative gain the budget office expects ^[20]. By JP Morgan's estimate, that is the bar for one chipmaker's valuation ^[5].

Revenue is the other test. Bain's $4.2 trillion over five years averages $840 billion of new revenue a year ^[14]. Van Nieuwerburgh's bar is an annual run rate: about $3.55 trillion of revenue by 2032, for a sector that, Reuters reported, earns a fraction of that now ^[9]. "The question is whether the applications arrive in time to pay for it," Bain's study said ^[7].

Two outcomes would vindicate the spending. In the first, productivity growth reaches JP Morgan's pace and the CBO baseline turns out too low ^[5]. Jasjeet Sekhon, chief strategy officer at Google DeepMind, named the second at a UC Berkeley summit in August. He said recursive self-improvement, models teaching themselves, was a "key part of the investment thesis" ^[12]. JP Morgan described the case where neither happens: "Historical precedent suggests that technology-driven booms often end when infrastructure buildouts cease to deliver sufficient returns," the bank wrote ^[13].

In my view the evidence supports the cautious reading for now. Anthropic's plan asks buyers to pay today for an annual return at least ten times its 2025 revenue ^[17], and the sector-wide revenue needed for a 10% return is still several multiples away ^[9]. The counter-case is that a single company's sales can compound far faster than national productivity, and an IPO prices one company's cash flows. The cautious reading would be proved wrong if Anthropic's sales climbed toward $51.8 billion a year ^[16], or the sector's toward $3.55 trillion by 2032 ^[9].