AI Costs More to Build and Less to Buy. The Difference Is on Your Bill.

Meta's capital spending forecast for this year now runs to $145 billion — nearly its entire projected 2026 EBITDA of $145 billion — and $10 billion of that was added in a single revision, the cost of memory-chip inflation alone ^[1]. In the same stretch, OpenAI cut the price of its Luna model by 80% and Anthropic shipped a high-performance model at half the cost of its top system, both to hold off cheaper Chinese rivals ^[2]. Morgan Stanley has already done the arithmetic on where that leaves the industry: a fully optimized data center running the latest Nvidia chips costs $25 billion a year to rent and produces $23 billion in output ^[3]. The inputs to AI keep getting more expensive; the thing AI sells keeps getting cheaper. The gap has to close somewhere, and it is closing on the people who buy electricity and computers. Start with the power bill. Ohio households are projected to average $800 a month for electricity between June and September, a 17% jump ^[4]. Duke Energy is asking North Carolina regulators to move the typical residential bill from $143 to $168 a month by 2028 to fund grid expansion for data centers ^[5]. Across 13 states, 67 million Americans have already seen their bills rise, with the market monitor attributing 64% of the increase in one capacity market to data center demand ^[6]. It is not only electricity. Memory-chip shortages have pushed up personal computer prices for the first time since the early 1980s ^[7], and Apple raised prices on iPads, MacBooks, and home devices to offset rising chip costs ^[8]. The regulatory track runs parallel. The EPA has proposed ending the mandatory public notice and comment period for air permits at data centers ^[9], and separately proposed letting builders start construction on piping, wiring, and site work before a permit is granted — sunk costs that make it politically difficult to reject the permit later ^[10]. FERC has ordered grid operators to accelerate connections for large AI loads ^[11], and the Energy Secretary directed the commission to speed data center hookups using a federal authority one former FERC chairman called an unprecedented expansion of federal control over the states ^[12]. The consumer-protection language sits inside the same orders that mandate faster buildout ^[13]. The administration's answer to the cost question is a voluntary pledge in which seven tech firms committed to pay their own way — non-binding, with final spending still subject to state regulator approval, even as utilities plan $1.4 trillion in capital spending that analysts call a leading indicator of future rate increases ^[14]. Then the question that would justify the whole bill: productivity. Of the 60 largest US-listed financial firms, one has reported a realized dollar return from AI across three years of earnings calls ^[15]. The Federal Reserve is split, with the chairman arguing AI will be disinflationary like the 1990s and other officials warning of a shock if the productivity never shows up ^[16]. Oregon shows the squeeze is being resolved at consumers' expense by choice, not necessity. The state's POWER Act raised data center power rates by 29% while cutting residential rates by 1.3%, and other states are now considering similar legislation ^[17]. That is the direction of travel that protects households — and it runs against the grain of federal policy, not with it. Goldman Sachs now projects $7.6 trillion in AI infrastructure spending through 2031 ^[18], and the banks are pricing in a commodity upcycle on the inputs that number assumes ^[19]. The math that launched the buildout is already obsolete, and the gap between what AI costs to build and what it earns is widening, not closing.