A visual essay · power → tokens

Electrons to tokens.

A single electron leaves a transmission line at 230,000 volts and ends its working life inside a tensor core at 0.8 volts, having helped emit one token. Scroll through the eight transformations that make AI possible, and expensive.

Scroll
Stage 1 · Transmission01 / 08

230,000 volts in a wire

Power leaves a generator at hundreds of kilovolts. The reason is unromantic: losses scale with current squared, so the cheapest way to ship watts long distances is to push them at brutally high voltage.

Voltage

230–765 kV AC

Frequency

60 Hz

Line loss

~3–7% per 100 mi

Eight transformations. Roughly twenty-five percent loss.

From the generator to the tensor core, somewhere between 20 and 30 percent of the original electrical energy never reaches the matmul. Transmission, transformers, cooling, conversion losses inside the server, the VRM stepping down to 0.8 volts: each stage takes a tax.

The frontier model you talk to runs on top of that whole stack. The bottleneck that gets discussed first, chips, is rarely the one that actually bites. Power, transmission, cooling, and packaging show up earlier, take longer to build, and answer to no Moore's law.

The reason a frontier lab cares about a county zoning hearing in Loudoun is the same reason it cares about ASML's production schedule. Both are step-down moments, places where the supply upstream has to be re-shaped to fit the demand downstream, and neither one moves on the timescale of a model release.

Sources

  • Lawrence Berkeley National Lab. 2024 United States Data Center Energy Usage Report. DOE LBNL-2024 ID number 2001637. Power-draw ranges, PUE distribution, and projected demand.
  • Open Compute Project. OCP 48 V Power Architecturespecifications. Server-rack and PSU voltage standards.
  • NVIDIA. H100 Tensor Core GPU architecture whitepaper(TDP, FP8 throughput, HBM bandwidth) and GB200 NVL72 datasheet (~120 kW rack draw).
  • US Energy Information Administration. Transmission and distribution losses in the US grid (3–7% typical, ~5% national average).
  • IEEE Spectrum and Uptime Institute. Data center electrical topology (transformer, switchgear, UPS, busway, PDU stages) and typical generator-start ride-through windows.
  • Semianalysis and ServeTheHome teardown coverage. GPU package construction (HBM placement, BGA pin count, on-board VRMs, core voltage and current at TDP).
  • Energy-per-query estimates: Epoch AI and Stanford HAI AI Index 2026 figures, plus published lab disclosures. Quoted ranges, not point estimates; actual energy varies by 10× between models and prompts.

Numbers are typical industry values, not specs of any single site. Where ranges are given, ranges are real.