AI's Labor-Substitution Bet Is Out of Margin

On Wednesday, the UK energy regulator Ofgem proposed something that would have been unthinkable a year ago: a fee of up to £712,500 per megawatt for AI data center developers who reserve grid capacity. The proposal landed after UK grid connection requests surged from 41 gigawatts to 125 gigawatts — nearly triple the country's peak demand of 46 GW. Regulators are no longer just watching the AI buildout strain the grid. They are pricing the strain and sending the bill to the companies that cause it. ^[1] This is the third margin to close on the AI industry's central bet. The bet was straightforward: replace workers with AI, let the wage savings fund the infrastructure buildout. It ran on three columns. Column one: layoffs attributed to AI would produce real savings on the P&L. Column two: the cost of running the AI would stay below the cost of the people it replaced. Column three: the expense of upgrading the grid to serve data centers was being passed through to households in their utility bills, and no regulator had yet moved to stop it. ^[2][3] All three columns have now turned red, and they did it at roughly the same time. The savings column was largely fictitious. MIT researchers found that 95% of corporate generative AI pilots produced no measurable impact on profit and loss statements. ^[4] Oxford Economics examined the first eleven months of 2025 and found that AI-related job losses accounted for only 4.5% of total US layoffs — cuts attributed to market conditions were four times larger. The firm concluded that companies are "trying to dress up layoffs as a good news story" and that "firms don't appear to be replacing workers with AI on a significant scale." ^[5]

We suspect some firms are trying to dress up layoffs as a good news story rather than bad news, such as past over-hiring. — Oxford Economics

The layoffs that were announced as AI-driven efficiency are now reversing. Commonwealth Bank reversed 45 customer service terminations. Taco Bell is reassessing voice AI in its drive-throughs. Up to 55% of companies that replaced employees with AI acknowledge they moved too quickly — a phenomenon analysts have labeled the "doorman fallacy": the discovery, after the fact, that the technology cannot replicate human judgment. ^[6] The savings were not just overstated. In many cases they are turning into rehiring costs. The cost column inverted even faster. Nvidia VP Bryan Catanzaro stated that compute costs for his team "far exceed" employee costs. ^[7] Microsoft canceled thousands of internal Claude Code licenses after experimentation "significantly increased operational expenses," redirecting engineers to its own cheaper GitHub Copilot. ^[8] Uber exhausted its entire 2026 AI coding tool budget in four months after using leaderboards to incentivize staff to maximize AI usage — the efficiency drive itself became the cost driver. ^[8] A four-person team at one firm ran up a $113,000 monthly AI bill. ^[9] Meta's CTO issued a memo telling staff not to use AI tools "just for the sake of using them," reversing the company's prior encouragement of high token usage. ^[10]

Nobody should be using AI tools just for the sake of using them. — Andrew Bosworth

The arithmetic had flipped: the tool that was supposed to fund the buildout by cutting wages was now costing more than the wages it cut. Then came the third column — the one that had been quietly carrying part of the cost. US utilities are planning $1.4 trillion in capital expenditures through 2030 for grid modernization driven by AI data centers, a 21% increase over the prior five years. PowerLines, an industry research group, warned these plans are a "leading indicator for future utility rate increase requests," following a 40% rise in utility bills since 2021. ^[2] In Illinois, data center growth is driving up residential electricity costs as utility providers pass infrastructure investment costs to households. ^[3] In Ohio, residential bills are projected to average $800 a month this summer, a 17% increase, compounded by multi-billion-dollar tax breaks previously granted to Google, Amazon, and Meta. ^[11] The pass-through was quietly carrying part of the cost: households paying for AI's power without realizing it. And then regulators began to notice. In February, Anthropic made a voluntary pledge that now reads like an early warning. The company promised to pay 100% of grid upgrade costs and compensate ratepayers for price increases caused by its data centers. CEO Dario Amodei made the company's position explicit. ^[12]

We’ve been clear that the U.S. needs to build AI infrastructure at scale to stay competitive, but the costs of powering our models should fall on Anthropic, not everyday Americans. — Dario Amodei

The pledge was voluntary. What followed was not. In July, Oregon's Public Utility Commission raised data center power rates by 29% under the POWER Act, explicitly requiring data centers to pay their share of grid infrastructure costs. Residential rates dropped 1.3% as a result. State Senator Janeen Sollman said other states are considering similar legislation. ^[13] Then came Ofgem's proposal in the UK — fees of up to £712,500 per MW for grid capacity reservations, a direct monetization of the speculative power demands AI developers have been placing on the system. ^[1] A subsidy that had gone largely unnoticed is now being itemized, priced, and billed back to its source. There is one obvious escape hatch: what if inference costs simply collapse? DeepSeek permanently cut V4-Pro API prices by 75% in May, making its models 12 to 19 times cheaper than GPT-5.5 and Claude Opus 4.7 for equivalent tasks. ^[14] If per-token prices keep falling, the cost column might right itself. But Gartner and Goldman Sachs analysts point out that the rise of autonomous agentic systems changes the arithmetic: agents run continuously rather than on-demand, so total consumption rises even as per-token unit prices fall. ^[8] Cheaper tokens do not mean cheaper systems when the systems never stop running. Three columns of the same ledger. The savings column never materialized at scale. The cost column inverted. And the third column — the one that let households absorb the difference — is being closed by regulators who have started reading the meter. No margin remains on any side. The squeeze has no gap left.