Retrieve embodied water estimates from BoaviztAPI

[jnioche]

Embedded (manufacturing) phase — wu has data

In factors.yml, manufacturing impact factors for components include wu. For example, every component type (CPU, RAM, SSD, HDD, etc.) has a wu impact entry:

• CPU: wu: { impact: 3.8931 } m3 eq.
• RAM: wu: { impact: 1.0459e+01 } m3 eq.
• SSD: wu: { impact: 1.6173e-02 } m3 eq.

So embedded/manufacturing water impact is supported and will return values.

curl "http://localhost:5000/v1/cloud/instance?provider=aws&instance_type=m5.xlarge&verbose=true&criteria=wu"

[jnioche]

Although in practice this might not be doable just yet. A bit of Clauding on the BoaviztAPI repo

In the PEF (Product Environmental Footprint) methodology, water use (wu) can be negative. A negative value means the product's lifecycle results in a net return of water to the environment — typically from end-of-life recycling
processes that release water previously consumed during raw material extraction. The ADEME NegaOctet dataset (the source) includes end-of-life credits, and for servers/electronics, the recycling credits for water often exceed the water
consumed during manufacturing.

This is also true across the board — almost every device-level wu factor is negative: SERVER (-3710), LAPTOP (-820/-812), DESKTOP (-4490), SMARTPHONE (-107), TABLET (-84.6), etc.

How you're hitting the -3710 value specifically

Looking at cloud_impact_embedded (line 821-889): when computing embedded impacts per-component, if any component's compute_single_impact returns None (line 864), the code falls through to the except NotImplementedError block (line 873)
which uses the generic SERVER fallback factor:

impact = Impact(
value=get_impact_factor(item="SERVER", impact_type="wu")["impact"], # -3710.0
...
)

This gets allocated by duration / life_time and then multiplied by the vCPU allocation ratio. For an EC2 instance that gets, say, ~4% of the server:

-3710.0 × (duration / lifetime) × allocation ≈ -154.6

So your -154.6 is the generic server wu factor (-3710) prorated by lifetime and vCPU share. It means the per-component wu calculation failed for at least one component (likely because cpu, ram, or ssd don't have individual wu factors in
factors.yml), causing the entire embedded calculation to fall back to the coarse generic server-level factor, which is negative.