100 Gigawatts for What?

This post was written by Claude. It is the output of a conversation prompted by a post from Rhys Lindmark on Substack. As we race headlong into what is clearly hype-induced over-capitalization, I wanted to know how the energy draw of useful work that AI is capable of compared to the buildout proposed by NVIDIA, the hyperscalers, and the investors that finance them.

Handing every white-collar administrative task in the US to AI, plus a generous R&D cushion, works out to roughly 15–25 GW. The buildout being talked about is 100 GW. That gap is the whole story.

What 100 GW actually is

Run continuously, as data centers do, 100 GW is about 876 TWh a year. In rough terms (approximate figures, EIA and LBNL baselines):

  • Households: about 80 million, or roughly 60% of US homes.
  • US electricity: about 21% of current annual consumption.
  • US total primary energy: about 3%.
  • US generating capacity: about 8% of nameplate, and about 13% of summer peak demand.

All US data centers used about 176 TWh in 2023, or 4.4% of electricity. LBNL’s high-end projection for 2028 is 580 TWh. A full 100 GW would be roughly five times today’s total data center load, and well above even that high-end projection.

It is also around-the-clock load. That means firm generation: gas, nuclear, or storage-backed renewables. Covering it with solar alone would take on the order of 400 GW of panels, plus storage.

The stated rationale

The case for building at this scale rests on six arguments.

  1. Scaling laws. Bigger models trained on more data perform better, so better AI means more compute. The bet is that this curve keeps paying off.
  2. Frontier training runs. The International AI Safety Report 2026 projects the largest single training runs in 2030 at 4–16 GW each. A few labs wanting several of those gets you to tens of GW.
  3. Inference demand. Agents, coding tools and enterprise workloads are expected to dominate long-term load, growing with adoption.
  4. Economics. Chips cost far more than the power that runs them, so IFP argues AI data centers must run 24/7 on firm power. Idle hardware is the expensive failure, so builders oversize.
  5. Geopolitics. The same IFP proposal calls for federal “Special Compute Zones” of 5 GW clusters, eventually totaling tens to hundreds of GW. China’s state-coordinated grid buildout is the usual foil.
  6. Market positioning. Chip-for-capacity deals pre-sell compute and lock customers into ecosystems. Whoever builds first holds the workloads.

One scoping note: Epoch AI’s estimate, cited by Carnegie, is over 100 GW of AI data center power worldwide by 2030. Not the US alone.

A sanity check: automate the bureaucracy

So take the most aggressive near-term use case seriously. Give every white-collar administrative task in the US to AI inference, and see what it costs in power.

The assumptions:

  • Workers: 20–80 million. The low end is BLS office and administrative support (about 19 million). The high end adds management, business, financial, legal and government administrative work. Midpoint: 50 million.
  • Tokens per worker-equivalent per day: 1–10 million. A person handles maybe 50,000–100,000 tokens of text a day. Agentic AI is far more wasteful: reasoning, retries, re-reading context. Midpoint: 5 million.
  • Energy per token: 0.1–2 joules, blended across input and output. Google’s reported figure of about 0.24 Wh per median Gemini prompt works out to roughly 0.3–1 J per token. Midpoint: 0.5 J.
  • Overhead: 1.2× for cooling and facility load, and 3× for peaking, since office work happens in business hours.
CaseInputsAverage loadCapacity at 3× peak
Low20M workers × 1M tokens × 0.1 J~0.03 GW~0.1 GW
Mid50M × 5M × 0.5 J~1.7 GW~5 GW
High80M × 10M × 2 J~22 GW~67 GW

Add 10–20 GW for frontier training and R&D across several labs. The total is about 15–25 GW in the middle case and 80–85 GW at the extreme.

The extreme case stacks every pessimistic assumption at once: every administrative job fully automated, highly wasteful agents, inefficient hardware, and no load shifting. Even then it falls short of 100 GW.

So what is the other 75 GW for?

Not for automating American bureaucracy. 100 GW only pencils out if you add global demand, heavy video and multimodal generation, very large training ambitions, or demand that grows far beyond replacing work people already do.

That makes the buildout a bet on demand that doesn’t exist yet. It is a speculative case, not an engineering one. And the people making it are mostly the ones selling the chips, the power and the compute.

The costs are not speculative. Firm 24/7 load at this scale means new gas plants, new transmission and rate-base increases paid by ratepayers who never asked for it, plus water, land and emissions. Railroads in the 1840s and fiber in the late 1990s were real technologies too. Both overbuilt on projected demand, and investors ate the glut.

The caveats cut both ways:

  • Tokens per task is the least-known input. A 10× swing there moves the answer 10×.
  • Cheap inference creates new demand rather than only replacing old work.
  • Hardware efficiency improves roughly 1.4–2× a year, shrinking every number above over a multi-year build.

The last point matters most. The buildout is sized to today’s hardware for demand that may never arrive. By the time it does, the same work may need a fraction of the power.

Sources