Course project · Concept

A proposed standard

Greener
Hours.

Every AI API response carries a new number: how clean the power behind it was. Three tools show what that unlocks: a chat indicator, a job scheduler, and a procurement dashboard.

Skip to the live demo

Course
Design for a Warming World
Program
MS Strategic Design & Management · Parsons
Term
Spring 2026
View the full deck ↗

§ 01 · The problem

The user sees the AI.
They don't see any of this.

945TWh

Projected global data-center electricity demand by 2030, more than double 2024's 415 TWh, with AI identified by the IEA as the lead driver.

025050075010002002018240202029020224152024~5802026~76020289452030Global data-center electricity · TWhsolid · actual outlined · projection
VISIBLETHE WHOLE INTERFACEUserQUERIES THE MODEL↑ THE LINE OF VISIBILITY ↑DARK MATTEREVERYTHING THAT MAKES THE INTERFACE WORKAIDatacenterGridSolarWindGasCoalOilRENEWABLEFOSSIL
Supply side
Aggregate, annual, self-reported. Not per request.
Demand side
Zero signal. Nothing per request, period.
Dev tools
Voluntary, research-grade, inconsistent. Not procurement-grade.

IEA, Energy and AI (April 2025); Shehabi et al., 2024 U.S. Data Center Energy Usage Report, LBNL (Dec 2024). Dark-matter framing after Dan Hill, Dark Matter and Trojan Horses (Strelka, 2015); AI energy estimates: IEA, Energy and AI (2025).

§ 02 · The reframe

Visibility doesn't change end-user behavior. But visibility paired with a standard creates supply-side pressure.

The same pattern as HTTPS, energy labels, SBOM, and cloud-carbon dashboards.

End users didn't change.
Providers did.

A backend protocol made visible at the UI layer through a tiny glyph. No one wanted their site flagged “Not Secure” in Chrome. Within a decade the entire web shifted.

0%2550751002018 · CHROMEFLAGS “NOT SECURE”↓ THE INFLECTION'94'10'18'25HTTPS share of web traffic · 1994–2025

The same pattern, across domains

EU energy labels
Regulatory disclosure

The A–G grade reshaped manufacturer behavior more than consumer behavior; appliances were re-engineered to climb the visible scale.

FDA nutrition labels
Mandatory disclosure

Barely moved consumers; drove reformulation (trans fats, sodium, sugar) because makers competed on what was visible.

Cloud carbon dashboards
Already happened

AWS, Azure, GCP shipped per-service carbon in 2021–22 on procurement demand. The pattern works; the AI-specific extension is what's missing.

Google Transparency Report; Let's Encrypt annual reports (2017–24). Curve directional. Inflection: Chrome 68 (July 2018) flagged all non-HTTPS pages.

§ 03 · The product · the standard

One standard. Three surfaces.

The substrate
Open API standard · per-request carbon disclosure as a response header.
EXISTING AI API CALL · UNCHANGEDPOST /v1/messagesAuthorization: Bearer ...Content-Type: application/jsonRESPONSE200 OKContent-Type: application/jsonX-Compute-Carbon-Intensity: 412 gCO2eq/kWhX-Compute-Region: us-east-1X-Compute-Confidence: highSDK PARSESSURFACED ACROSS THREE TIERST1 · CHAT INDICATORInline glyph · “current compute window”T2 · FLEXIBLE SCHEDULERDeadline · route to cleanest windowT3 · ENTERPRISE DASHBOARDPer-team aggregation · procurement view

Open-source spec · reference SDK in Python, TypeScript & Go · piggybacks on existing API plumbing. Adoption is three headers.

§ 04 · The three surfaces

One indicator, one scheduler, one dashboard.

↻ greenerhours.dev/console
End-user viewchat.ai
MODELSonnet 4.7 ▾
High intensity · 462 gCO₂/kWh
Hi, this chat is wired to a live model API through a server-side proxy, and it can see the carbon state in the indicator above. Ask me anything.
COMPUTE WINDOW · CURRENT
462 gCO₂eq/kWh|us-east-1|high confidence
Dev-team viewconsole.ai/scheduler
Submit job
Batch transcription · 412 audio files
Deadlineby today · 22:00
Recommendedtoday 21:00 · 284 g 39% vs now
…or tap a bar on the forecast to pick your own window; the system still checks it against your deadline.
Demo clock14:00 · 462 g
fast-forwards the day so you can watch jobs run
48h grid forecasttap a bar to pick · jobs marked ↓
Job queue · 2cleanest-window routing
Batch transcription · 412 audio files
sub 13:00 · by 21:00
17:00
168 g
51%
Queued
Overnight feedback summary
sub 12:00 · by 20:00
16:00
194 g
46%
Queued
Procurement viewconsole.ai/footprint
AcmeCorpCOMPUTE FOOTPRINT
May 2026 ▾
Operational efficiency
TOTAL REQUESTS
2.40M
+18% vs Apr
AVG INTENSITY
462g/kWh
−24% vs Apr
HOURS ON GREEN POWER
5of 24h
5h green · 8h mixed · 11h fossil-heavy
May: 62% of compute-hours on green power · 38% not
EST. SCOPE 3 · MAY
847tCO₂eq
audit-ready · CSRD
FLEX SCHEDULER · 14:002 jobs routed (0 complete) · avg −48% intensity · ~680 g CO₂eq avoided (sim)
CARBON INTENSITY & REQUEST VOLUME · DAILY · MAY 2026 intensity volume
Anti-rebound view: intensity is falling, but volume is rising. Both are surfaced so efficiency cannot disguise growth.
GREEN WINDOW USE · DAY
46% of 84k requests ran while the grid was green
46% green · 36% mixed · 18% fossil-heavy
2 flexible jobs routed to green windows · 0 complete (sim)
ran in green windows46%
potential71%
potential = every deferrable job shifted to a green window, inferred by the system (sim · illustrative)
BY TEAM
R&D782k
Eng621k
Mktg428k
Ops275k
BY MODEL
Sonnet 4.71.5M
Haiku 4.5744k
Opus 4.7211k

Concept prototype · illustrative data · one shared simulation; grid & dashboard figures are invented · chat runs on a live model API via a server-side proxy

The P-codes on each surface map to the ten design principles from the Design for a Warming World course (SP26 syllabus); I built the project against that framework.

open the prototype full-screen ↗

§ 05 · The opportunity · three forces

Why now, not five years ago.

01

Regulatory pressure

All demand vendor-level emissions data. None specify the technical standard.

02

Procurement demand

Cloud got per-service carbon dashboards. AI didn't.

03

Reusable infrastructure

Three layers already exist. The fourth is the work.

§ 06 · Named, not hidden · if it works

What this project cannot pretend.

The biggest lever for AI's footprint is supply-side: siting, PPAs, grid decarbonization. Greener Hours operates on a derivative lever: user-facing visibility pressuring the supply side. The bet is contestable; the HTTPS analog suggests the pattern works, but the timeline is uncertain.

Open standards get captured. Three guardrails: an open-source reference implementation, third-party verification of disclosure claims, and an anti-rebound layer that surfaces total consumption alongside efficiency. None is sufficient alone.

The sharpest critique I got at midterm, and the answer is structural, not defensive: Tier 1 is not a behavior tool. Behavior change happens at Tier 2 (a one-click default) and Tier 3 (institutional decisions). Tier 1's job is to make the cost visible, and visible costs create supply-side pressure.

Greener Hours is a timing-and-disclosure intervention, not a sufficiency one. It does not advocate using less AI. Tier 3 makes absolute usage visible, but visibility is not advocacy.

When compute disclosure becomes a normal procurement field.

  • 3 of 5Adoption · 24 mo major AI providers implement the spec, on competitive and procurement pressure.
  • 60%Coverage of enterprise AI API calls return standardized carbon data by default.
  • 40%Behavior · Tier 2 of deferrable batch tasks shift to cleaner grid windows.
  • 5+Institutional · Tier 3 Fortune 100 buyers name the spec in AI RFPs.
  • 1Regulatory EU AI Act or California SB 253 references the spec: the mandate moment.
  • Cultural “What's your compute-carbon disclosure?” becomes a normal sales question.

Directional, not predictive: these are the moves I'd watch for, not measured outcomes.

Make the invisible legible, and the providers move.

Project
Greener Hours · v0.1 final
Author
Rishabh Salian
Course
Design for a Warming World · SP26