The Compute Behind Agentic AI Is Replacing Junior Workers and Reviving Coal

In February, Goldman Sachs became the first major bank to deploy agentic AI as "digital co-workers" for back-office functions — trade accounting, compliance, reconciliation — as part of CEO David Solomon's plan to constrain headcount growth.

Think of it as a digital co-worker for many of the professions within the firm that are scaled, are complex and very process intensive. — Marco Argenti

It was the most visible instance of a technology now moving through the white-collar economy. The same month, the same class of models was straining compute infrastructure so severely that Anthropic began cutting session limits during peak hours, with developers reporting quotas depleting in minutes.

During weekdays between 5am–11am PT / 1pm–7pm GMT, you'll move through your 5-hour session limits faster than before. — Thariq Shihipar

One technology, two effects that read as contradictions: the most advanced software ever built is simultaneously eliminating junior professional roles and driving a data center buildout so power-hungry it is reviving the coal economy. The labor side is no longer a forecast. The Bank of England reported this week that vacancies in roles with high AI exposure decreased by 15% over three years, with administrative and customer service roles declining more than 20%. ^[1]

If AI is genuinely behaving like a general-purpose technology, these productivity and labour-market signals should ultimately be interpreted in coordination rather than in isolation. — Haley Schlicht

In February, Federal Reserve Governor Lisa Cook warned that AI is triggering "the most significant reorganization of work in generations," with displacement that may precede job creation — specifically hitting coding roles and entry-level positions for recent graduates. ^[2]

we appear to be approaching the most significant reorganization of work in generations — Lisa D. Cook

Goldman's deployment of Claude for back-office functions is not an outlier; it is the pattern arriving at its logical first stop — the bank where junior analysts have historically done the reconciliation and compliance work now handled by software. The seam where the labor tool becomes an energy problem is visible in Anthropic's own operations. In March, the company quietly reduced Claude Code session limits during peak hours, with about 7% of Pro users affected and the company acknowledging it is investing in scaling infrastructure. ^[3] The session caps are a small operational adjustment, but they reveal the physics underneath: agentic AI models — the kind that can function as digital co-workers — consume compute at a scale that strains even the labs that built them. That compute demand is now reshaping the American landscape. Over $130 billion in AI data center projects were blocked or delayed in early 2026, including a $1 billion Google proposal in Indianapolis and a $3.6 billion Amazon campus in Tucson. ^[4] New York became the first state to enact a statewide moratorium on certain AI data center developments. In West Virginia, the legislature passed House Bill 2014, stripping local governments of zoning and noise ordinance authority over data centers to clear the path for projects like the 1.3-gigawatt Monarch Compute Campus in Mason County, which caused flooding and foundation damage that led 60% of homeowners in one neighborhood to sell. ^[5]

If we’re going to do this, we should do it like other states do it and do it right, and allow for the people to have a say so in who’s putting the data center in their backyard. — Mike Pushkin

The political cost is high, but the power has to come from somewhere — and what is available now is coal and gas. Wood Mackenzie revised its forecast for the decline of U.S. coal-fired generation from 60% down to 39%, with coal production increasing due to higher demand and delayed plant retirements. ^[6]

We’ve got to stop existing firm capacity from retirement. — Alfie Moon

Energy Secretary Chris Wright has advocated to "stop existing firm capacity from retirement" to avoid blackouts. ^[6]

Energy sobriety has returned to Washington, D.C. Our focus is on Americans and the price of utility and avoiding blackouts. — Alfie Moon

The Trump administration is actively encouraging states to use coal to meet data center demand. ^[7]

Building more gas plants seems like going in the exact opposite direction of what we need to do as a state. — Howard Watts III

NV Energy warned that proposed AI data centers could require electricity equivalent to three times the current demand of Las Vegas, threatening Nevada's 50% renewable goal by 2030. ^[7] Meta is building 10 gas plants for its Hyperion data center in Louisiana; GE turbines are sold out through 2029. ^[8] The clean alternatives are real, approved, and under contract. TotalEnergies signed 15-year power purchase agreements to supply 1 gigawatt of solar power to Google data centers in Texas — the largest renewable PPA the company has signed in the U.S. ^[9] NuScale Power, the first company to receive NRC approval for small modular reactors, is targeting AI data centers as a market. But its first system will not come online until 2030 or later. ^[10]

The race to scale AI has triggered one of the largest infrastructure build-outs in modern history. — McKinsey & Company

That is the structural fact that makes the contradiction durable: the clean energy exists on paper and in contracts, but it arrives in 2030 while the data centers are being built and the coal plants are being un-retired now. The same deployment that is eliminating entry-level white-collar work — at Goldman, across British firms, in the roles Lisa Cook warned about — will run, for the next four years, on coal and gas rather than the reactors and solar farms that are not yet here.