The Math Behind ‘AI Will Take Your Job’ Is Missing One Crucial Number.

Summary

The video revisits the claim that AI could replace office workers and asks how much electricity that would require. The narrator says his earlier estimate of about 100 GW was flawed: it assumed 100 million workers, a dedicated 700 W chip for each worker, and counted cooling twice.

Reworked Estimate

The narrator uses U.S. labor statistics to count about 52 million office jobs, or roughly 87 million under a broader definition. He estimates that a full-time worker works about 1,800 hours per year and uses a programming-agent benchmark averaging 4.17 million tokens per task. He treats one benchmark task as equivalent to one hour of human work, while acknowledging that this is an uncertain assumption.

To estimate inference energy, he draws on Microsoft research, adjusts for different job types, and applies his own estimate of 80% input caching. His best-case average estimate is about 27 W per replaced worker, or 12 TWh per year / 1.4 GW average power for 52 million workers. He gives a broader plausible range of 30–150 W. The estimate excludes model training.

The Main Variable: Speed and Chip Sharing

The central argument is that average energy use does not determine the infrastructure required. The crucial factors are how quickly each agent responds and how many agents share the same chips.

Using a benchmark of a large open model running on an Nvidia GB200 rack, the narrator describes this trade-off:

• At 90 tokens per second, a chip serves many agents, at about 24 W per agent.
• At 165 tokens per second, the estimate rises to about 300 W per agent.
• At 220 tokens per second, it reaches about 1,619 W per agent.
• At 286 tokens per second, one agent needs nearly five chips, or about 9,746 W.

In this example, tripling token speed raises energy per agent by roughly 400 times. The rack’s overall power budget stays nearly constant; faster service means fewer agents can share it. The narrator says the earlier 700 W figure roughly corresponds to one agent receiving dedicated hardware at around 190 tokens per second, rather than being a typical average.

Power Scenarios and the Grid

Depending on the chosen speed, the narrator estimates that replacing all office workers could require:

Speed Estimated power 90 tokens/s 600 MW 178 tokens/s 12 GW 220 tokens/s 41 GW 286 tokens/s 244 GW

The answer can therefore range from less than current U.S. data-center consumption to far more than projected AI data-center capacity. The narrator’s point is that job-replacement claims—and rebuttals—are incomplete unless they specify response speed and hardware-sharing assumptions.

He also distinguishes average power from the capacity infrastructure must support at peak. Data centers need cooling and reserve capacity; he cites a power-usage overhead of roughly 1.15–1.5, depending on the facility. He says faster service may require more chips even if it does not make each chip draw more power.

Reliability, Oversight, and Job Replacement

The video argues that AI agents may use many tokens because they make errors and retry. The narrator cites a benchmark where the best model is wrong on 1.8% of simple document-summary cases, then illustrates how small per-step error rates compound across long sequences of agent actions.

He also discusses a METR study of experienced developers: participants using AI tools took 19% longer on the tested tasks, despite expecting to be faster. He notes that the study used tools from early 2025.

These points support the narrator’s view that human review may remain necessary. If a person must wait for and check an agent’s output, employers may favor faster—and more energy-intensive—service.

Efficiency and Demand Growth

The narrator explains the Jevons paradox: greater efficiency can make a technology cheaper to use, increasing total demand. He applies this to AI, arguing that efficiency gains may be overtaken by growth in the length and complexity of tasks agents can perform.

He calculates that projected efficiency improvements could be consumed within roughly 1.75–2.5 years, depending on the assumptions. He says even a claimed 10× improvement in a future Nvidia generation could be overtaken on a shorter timescale.

Infrastructure and Grid Stability

The video says that new generation capacity does not arrive immediately. It cites a record year for U.S. additions but argues that nominal capacity overstates dependable output, particularly for solar and batteries. It also cites a large queue of proposed grid connections and long median wait times.

Separately, the narrator raises a grid-stability concern: he says NERC issued its highest alert level after AI data centers rapidly dropped more than 8,000 MW of load. He argues that building more power plants does not by itself solve sudden load changes.

Guides and Resources

• The narrator provides a spreadsheet template for changing assumptions—such as worker counts, task intensity, caching, and agent speed—and recalculating the estimates.
• The video functions as a step-by-step calculation guide, explaining tokens, tokens per second, average versus peak power, and chip sharing.
• A separate video on data-center cooling is mentioned as a future topic.

Main Speaker and Sources

The main speaker is the video’s narrator, who identifies himself as a chemical engineer with a PhD in energy management.

Sources discussed in the video include the U.S. Bureau of Labor Statistics; a programming-agent token-use benchmark; Microsoft research on AI inference energy published in Joule; an Nvidia GB200 rack benchmark; Berkeley Lab data-center energy projections; METR’s developer study; an Anthropic study on task coverage and employment; and NERC grid alerts.