Environment, Energy & Compute · Capstone brief

ENV-03 · Grid intensity of a query

Estimate energy/carbon for representative LLM/API workloads; show sensitivity to hardware and location assumptions.

The question

Estimate energy/carbon for representative LLM/API workloads; show sensitivity to hardware and location assumptions.

Sources / materials

Teacher-approved assumptions table (tokens or request counts, hardware class, grid intensity). Public grid or generation mix sources such as EIA explainers/data (https://www.eia.gov/). Spreadsheet with sensitivity (± range on at least two assumptions). Student must log any compute used for the brief itself.

Expected failure modes

A single magic gram-CO₂ number. Copying a blog’s factor without citation. Ignoring location. Pretending the student’s own inference is free.

Done looks like

Assumptions table with sources; central estimate; sensitivity table; one-page interpretation; explicit list of what the estimate cannot support.

Five C's

CT: estimate vs. measurement. CR: transparent assumptions. CO: peer attacks one assumption. CM: one-page intensity brief. CZ: communities on the grid that serves the compute.

Mentor role

Env-science or physics teacher reviews assumptions before final numbers. Standing instruction: reject uncited intensity factors and reject false precision. School-supervised.

Rubric calibration

R1: one workload class. R2: arithmetic shown. R3: location/hardware variants. R4: sensitivity present. R5: page readable. R6: non-claim list on precision.

Two ways this goes wrong

(a) One magic number from a tweet. (b) No sensitivity, maximum confidence.

Checkpoint suggestions

  • Week 1–2: Workload definition and sources approved; assumptions table locked.
  • Week 4–5: Central estimate + sensitivity; peer attack on one assumption.
  • Week 7–8: One-page brief; student’s own compute log; mentor review.

Credit lane fit

Lane A immediately. Natural precursor to ENV-04. No verified credit claim.