Skip to content
Shipyard

Carbon

What it costs the planet to ship a feature here, measured for every request that went live, using the Software Carbon Intensity method (ISO/IEC 21031:2024).

"g CO₂e" means grams of carbon dioxide equivalent: one unit that puts different greenhouse gases on the same scale as CO₂. Shipyard measures it to show what shipping each feature costs the planet.

Per shipped feature, average of 15

13.1 g CO₂e

Up to 112.0 g if AI model calls cost ten times our estimate. About 22 minutes of CI and 79 model calls each.

  • CI runners
    1.88 g, 14%
  • AI model calls
    10.99 g, 84%
  • Hardware
    0.26 g, 2%

What the design avoided

Routing decisions made in plain code, not by an AI model

360

About 50.1 g CO₂e avoided across these features.

Requests declined at triage, before any code was written

11

About 125.0 g CO₂e avoided: each one cost a triage run instead of the full build.

Every launched feature

Serving a page

Every visit has a cost too. The pipeline measures what each page transfers on every release and applies the Sustainable Web Design Model v4: 122 g CO₂e per GB.

Page First visit Repeat visit mg CO₂e, first visit
/ 15.0 kB 4.0 kB 1.82
/roadmap 16.1 kB 5.2 kB 1.96
/changelog 13.1 kB 2.2 kB 1.60
/carbon 14.9 kB 3.9 kB 1.81
/incidents 13.1 kB 2.2 kB 1.60
/requests/1 13.6 kB 2.6 kB 1.65

The live roadmap checks for changes every 20 seconds. When nothing changed, the answer is an empty "not modified" reply; when something did, 3.5 kB, so an hour of watching costs at most about 76 mg CO₂e. Pages are compressed, styles and scripts are cached for a year, and the pipeline fails any change that makes a first visit heavier than 50 kB, so agents can't quietly bloat the site.

How it's measured

SCI = (E × I + M) per R

R, the unit
One feature request that reached production.
E, energy
Runner time of every pipeline the feature caused, read from GitLab: agent runs, merge request pipelines and the pipeline that deployed it. Each runner has 2 vCPUs and 8 GB, drawing 8.92 W at 50% utilisation, including a PUE of 1.1. AI model calls are estimated at one per 5 seconds of agent run time, 0.24 Wh each.
I, grid intensity
579.44 g CO₂e per kWh, location-based, for Google Cloud us-east1 where GitLab's runners run. The same value is assumed for model inference, whose location isn't published.
M, embodied emissions
The runner's share of the server's manufacturing footprint: 1200 kg CO₂e per 96-vCPU server over a 4-year life.
Outside the boundary
Serving visitors, because it grows with traffic, not with shipping. It's measured separately, per page view, above.
Serving, per page view
Sustainable Web Design Model v4: bytes transferred × energy per GB for data centres, networks and devices, operational and embodied. The data-centre share uses Montréal's grid (11 g CO₂e per kWh), where the site runs; networks, devices and all embodied energy use the global average (494 g).

The weakest number is the AI model energy: GitLab doesn't report it per run, so it is estimated and shown with a ten-times upper bound. Everything else is measured.

Sources