LCA & Sustainability · Product + Supply Chain Working plan · Horizon 2027–2030 AI Datacentre & AI PC

North star / the one sentence everything reports to

Carbon is a spec, not a story.

We build two things that the next decade will judge harshly: racks that turn electricity into intelligence, and laptops that carry a fab's worth of embodied carbon in a 1.2 kg shell. This is the plan to know the true number for both — per rack, per unit, per litre — and to make that number fall faster than anyone else's, on purpose, with receipts.

The plan below assumes nothing is credible until it is measured with primary data, modelled against a published method, and survivable under third-party assurance. Where the data does not yet exist, that is stated plainly rather than modelled away. Section 07 is the honest list of what we do not know.

2026 ground truth2027 coverage2028 reduction2029 circular2030 proof

01 / why the window is now

415 TWh · datacentres 2024

About 1.5% of global electricity, projected to roughly 945 TWh by 2030 in the IEA's base case — AI-optimised capacity is the fastest-growing slice.[1]

4.4 % of US grid, 2023

176 TWh in 2023, on a path to 6.7–12% of US electricity by 2028 (325–580 TWh) in LBNL's scenarios. Siting risk, not just carbon risk.[2]

62 Mt e-waste, 2022

Only 22.3% documented as formally collected and recycled; 82 Mt projected by 2030. Our take-back rate is a material claim, not a PR line.[3]

~80 % embodied, notebooks

For a typical notebook, manufacture + transport dominates cradle-to-grave CO₂e; the use phase is minor. Design decisions, not power contracts, decide it.[4][5]

Efficiency per token is falling. Absolute demand is rising faster. Only one of those numbers ends up in the inventory.

02 / north star

Every rack and every notebook leaves our dock with a verified cradle-to-grave number attached — and that number falls every single quarter.

The test of truthPrimary supplier data, ISO 14040/44 and 14067 conformant, PACT-exchangeable, limited assurance under ISAE 3000. If a number cannot be audited, we do not publish it.[6][7][8]
The test of pace1.5 °C-aligned near-term targets validated by SBTi, with an ICT-sector intensity pathway for compute and an absolute pathway for the enterprise.[9][10]
The test of valueEvery disclosure must earn its cost: winning regulated tenders, cutting energy and water opex, de-risking siting, and shortening enterprise procurement cycles.

03 / operating framework

Five layers. Each one feeds the next.

The order matters. Most programmes fail because they buy a reporting platform before they own their bill of materials. We build the ground truth first, model second, reduce third, close the loops fourth, and only then sell the result.

L1

Ground truth

A golden record joining item master → BOM → supplier site → commodity → emission factor, with a data-quality score on every row.

  • Every part number carries mass, material class, supplier site ID, process route and origin country — no BOM row without a material class by end of 2026.
  • Pedigree-matrix data quality scoring (reliability, completeness, temporal, geographic, technological) stored beside each factor, not in a footnote.
  • Site-level energy, water withdrawal and consumption, and waste streams metered at supplier tier 1 and all internal fabs of record; hourly grid data for owned datacentres via Redfish/DCIM.
  • Versioned emission-factor library (ecoinvent 3.11 + supplier-specific PCFs) with effective dates so a restated 2027 number is reproducible in 2031.
L2

Model & verify

One method stack, two speeds: full LCA for flagship SKUs, streamlined parametric models for the long tail.

  • Full ISO 14040/44 + 14067 studies with critical review for flagship AI PC and reference rack; PEF-aligned category rules for EU-facing claims.
  • Streamlined models for SKU proliferation: PAIA-style parametric device modelling and architectural carbon modelling (ACT) for silicon and accelerators.[11][12]
  • Compute-side: Software Carbon Intensity per unit of useful work (per million tokens, per training run, per inference SLA tier), published monthly by region.[13]
  • Assurance rehearsal every year from 2027 — walk the auditor through one product and one site before it is mandatory, not after.
L3

Reduce

Levers ranked by tCO₂e per engineering-week, reviewed at the same gate as cost and schedule.

  • Supply chain (the majority of our footprint): supplier renewable electricity clauses in tier-1 contracts, fab abatement of process gases and high-GWP F-gases, low-carbon aluminium and steel in chassis and racks.
  • Datacentre: hourly-matched carbon-free energy procurement rather than annual certificates; direct-to-chip and rear-door liquid cooling; higher supply temperatures per ASHRAE datacom guidance; heat reuse offtake where a district network exists.[14]
  • Silicon & system: node and packaging choices scored on embodied CO₂e/mm², memory-per-accelerator right-sizing, and power capping policies tuned to marginal grid intensity.
  • Product: reduce enclosure mass and part count, extend platform life, lift idle and low-load efficiency well past ENERGY STAR thresholds.[15]
L4

Circulate

Water, waste and materials are managed as loops with named owners, not as annual disclosures.

  • Basin-level water accounting using WRI Aqueduct stress classes; no potable water for cooling in extremely-high-stress basins; replenishment projects contracted in the same basin, not offset elsewhere.[16][17]
  • Chain-of-custody for recycled content: mass-balance certification (e.g. ISCC PLUS) for polymers, third-party recycled-content validation for metals, disclosed as physical vs allocated.[18]
  • Decommissioned accelerators and servers routed to redeploy → resell → harvest → recycle, with serial-level tracking and a hard landfill prohibition.
  • Design for disassembly: fastener count budget, no adhesive on batteries, published repair scores and 7-year spare availability.[19]
L5

Convert

Turn verified environmental data into revenue and permission to build.

  • A customer-facing PCF API: any enterprise buyer can pull the footprint of their own workload or fleet, formatted for their Scope 3 category 1 and category 11 inventory.
  • Digital Product Passport readiness ahead of ESPR delegated acts, so EU market access is a non-event rather than a scramble.[20]
  • Compliance as a by-product, not a project: CSRD/ESRS E1–E5, California SB 253/261, EU Energy Efficiency Directive datacentre reporting all fed from one dataset.[21][22][23]
  • Grid and community credibility: transparent water and load data is what wins the next interconnection and the next planning permission.

04 / the target ledger

What we sign up to, and how it is proved.

Every target below names an instrument. A target without a measurement method and a named owner is a slogan. Baselines are FY2024 unless stated; all figures are proposals for board sign-off, to be re-based once the 2026 ground-truth inventory closes.

Domain
2027 checkpoint
2030 target
Instrument
Scope 1 + 2
70% hourly-matched carbon-free energy across owned datacentres; F-gas abatement on 100% of owned process tools.
100% hourly-matched CFE; −50% absolute Scope 1+2 vs 2023.
Hourly EAC matching · GHG Protocol Scope 2 dual reporting · SBTi validation
Scope 3 — purchased goods
60% of category 1 spend covered by supplier-specific, PACT-exchangeable PCFs; top 100 suppliers on validated targets.
85% coverage and −30% absolute vs 2023, with primary data replacing spend-based factors.
PACT Pathfinder network · CDP Supply Chain · contractual data clauses
Compute intensity
gCO₂e per million tokens and per training PFLOP-hour published monthly, per region, location- and market-based.
−60% gCO₂e per unit of useful work vs 2025, with embodied carbon amortised into the number.
GSF Software Carbon Intensity · DCIM telemetry · hourly grid intensity feeds
AI PC product carbon
Published PCF for 100% of new SKUs; recycled content declared per material class on the datasheet.
−30% cradle-to-grave CO₂e per unit vs 2024, at constant performance class.
ISO 14067 · PEF category rules · parametric SKU model, critically reviewed
Water
Design WUE ≤ 0.20 L/kWh for all new builds; zero potable water for cooling in extremely-high-stress basins; consumption and withdrawal reported per site.
WUE ≤ 0.09 L/kWh on the new-build fleet; 120% replenishment of consumption in stressed basins; indirect (power-generation) water disclosed.
WRI Aqueduct basin classes · CDP Water Security · metered make-up water
Waste
≥90% diversion at all owned datacentre and manufacturing sites; hazardous streams tracked to final disposition.
Zero waste to landfill (≥95% diversion, certified); 100% of decommissioned servers reused or recycled with serial-level chain of custody.
Third-party landfill-free certification · R2v3/e-Stewards recyclers · asset serial tracking
Recycled & renewable materials
35% recycled content by mass in device enclosures; 100% recycled cobalt in batteries; plastic-free fibre packaging on all new SKUs.
50% recycled by mass overall; 75% recycled aluminium; 100% recycled rare-earth magnets in speakers and haptics; recycled copper in rack busbars.
ISCC PLUS mass balance · third-party recycled-content validation · declared physical vs allocated
Lifetime & repair
7-year spare-parts availability, public repair documentation, repairability score on every EU SKU.
Accelerator fleet life extended to a modelled optimum with ≥95% redeploy or resale at first retirement.
EU repair index · EPEAT criteria · fleet utilisation vs embodied-carbon break-even model

05 / data foundation & tooling

One dataset. Many audiences.

The architecture is deliberately boring: a lakehouse, versioned reference data, transformation as code, lineage on every field, and thin domain tools on top. We will not run sustainability off spreadsheets emailed between a supply-chain analyst and a consultant.

Layer A · modelling

LCA engines

Open, inspectable models first; commercial engines where regulator-facing rigour is required.

  • openLCA + ecoinvent 3.11
  • Sphera LCA FE / SimaPro
  • PAIA-style device parametrics
  • ACT for silicon & systems
  • Boavizta open device data
Layer B · supplier data

Value-chain exchange

Ask each supplier once, in a format they can reuse for every customer. Interoperability is the whole point.

  • WBCSD PACT Pathfinder conformant tooling
  • CDP Supply Chain campaign
  • EcoVadis / Manufacture 2030
  • RBA audit + smelter data
Layer C · corporate

Accounting & disclosure

The general ledger equivalent for emissions, water and waste — with restatement control and audit trail.

  • Watershed / Persefoni / Sweep
  • SAP Sustainability Footprint Mgmt
  • ESRS + SB 253 report packs
  • ISAE 3000 evidence vault
Layer D · operations

Live telemetry

Datacentre reality at hourly resolution, joined to marginal grid intensity and basin water stress.

  • Redfish + DCIM + BMS meters
  • Prometheus / Grafana pipelines
  • Hourly grid carbon intensity feeds
  • Cloud Carbon Footprint patterns
Layer E · platform

The lakehouse

Where product, procurement, logistics and facilities data finally meet. Owned by engineering, governed by finance-grade controls.

  • dbt models + tested contracts
  • Versioned emission-factor registry
  • Field-level lineage (OpenLineage)
  • Immutable restatement log
Layer F · outward

Passport & API

The same numbers, rendered for a regulator, a hyperscale customer and a consumer scanning a label.

  • ESPR Digital Product Passport schema
  • GS1 Digital Link identifiers
  • Customer PCF API (per workload)
  • Machine-readable EPD output

If two teams can produce two different footprints for the same laptop, we do not have a data foundation — we have opinions.

06 / sequence

Four years, in the order they have to happen.

Each year has one non-negotiable outcome. If that outcome slips, the following year's commitments are re-cut publicly rather than quietly missed.

H2 2026

Close the ground truth

  • BOM completeness sprint: material class, mass and supplier site on every active part number; gap list published internally with owners.
  • Stand up the lakehouse, factor registry and pedigree scoring; freeze the FY2024 recalculated baseline for carbon, water and waste.
  • Two reference LCAs completed and critically reviewed: one flagship AI PC, one reference AI rack (training and inference profiles).
  • Supplier data clauses drafted into standard T&Cs for all new tier-1 contracts.
2027

Coverage and credibility

  • 60% of category 1 spend on supplier-specific PCFs; PACT exchange live with the top 50 suppliers.
  • Monthly compute-intensity dashboard published per region; first limited assurance over Scope 1, 2 and the datacentre water dataset.
  • SBTi near-term targets validated; hourly CFE procurement contracted for the two largest campuses.
  • Design-gate rule in force: no product passes DVT without an approved PCF and recycled-content declaration.
2028

Reduction becomes visible

  • Liquid cooling standard on all new AI halls; heat-reuse offtake signed at a minimum of one campus.
  • Low-carbon aluminium and steel qualified for chassis and racks; first −15% cradle-to-grave notebook shipped.
  • Zero-waste-to-landfill certification achieved at all owned datacentres; supplier abatement programme covering top emitting fabs.
  • Customer PCF API in production; footprint data enters enterprise RFP responses as standard.
2029

Close the loops

  • Recycled-content chain of custody audited end to end; physical vs mass-balance claims separated in all public reporting.
  • Basin replenishment portfolio delivering measured volumes in every stressed basin where we operate.
  • Accelerator second-life programme at scale, with the retirement decision driven by the embodied-vs-operational break-even model.
  • Digital Product Passport pilot shipped on EU SKUs ahead of the compliance date.
2030

Prove it

  • Full target ledger externally assured; every claim traceable to a versioned dataset and a reproducible model run.
  • Reasonable — not merely limited — assurance on Scope 1, 2 and water; limited assurance extended over Scope 3 category 1.
  • Next-horizon commitments (2035) published with the same rigour, including whatever we failed to hit and why.

07 / open questions, risks, unknowns

The parts we cannot yet answer.

Leading means naming these before an auditor, a journalist or a customer does. None of the items below has a settled industry answer today. Each has a named owner and a review date, and each will be reported on whether or not it resolves in our favour.

OQ-01

Die-level embodied carbon is effectively unmeasurable today

Foundries publish site-level footprints, not per-wafer, and certainly not per-die. Allocation across a wafer with variable yield, across multi-die packages, HBM stacks and advanced packaging, has no agreed method. Our accelerator embodied numbers therefore carry wide uncertainty bands — and we will publish the bands, not just the midpoints.

No agreed method · high materiality
OQ-02

The Scope 2 rulebook may move under us

The GHG Protocol's standards update could change how market-based claims, certificates and hourly matching are treated. A revision that tightens deliverability or additionality would reset our baseline and our reported progress. We are building hourly, location-based and market-based views in parallel so that a rule change is a re-cut, not a rebuild.

Regulatory reset risk
OQ-03

Indirect water is probably larger than our on-site water

Reporting cooling make-up water while ignoring the water consumed generating our electricity understates the true basin impact. There is no consensus method for allocating power-sector water to a single consumer at hourly resolution. We will publish an indirect estimate with explicit assumptions rather than omit it.

Method gap · reputational exposure
OQ-04

When should an accelerator be retired?

Newer parts are more efficient per unit of work, but replacing them writes off embodied carbon. The break-even depends on utilisation, grid intensity and workload mix — all of which move. We do not yet have a defensible fleet-refresh model, and an over-eager refresh cycle could increase our total footprint while improving our intensity metric.

Unresolved · model in build
OQ-05

Efficiency gains may be swamped by demand growth

Intensity targets can improve every quarter while absolute emissions rise. If AI demand grows faster than we decarbonise, an intensity-only story becomes indefensible. We will report absolute and intensity side by side and accept the harder conversation.

Rebound effect · absolute vs intensity
OQ-06

Recycled content fights thermals, EMI and thinness

High post-consumer-recycled polymer loading affects flame retardancy, cosmetics and dimensional stability; recycled aluminium alloys can constrain anodising. Whether a 50%-by-mass target survives a premium thin-and-light thermal envelope is an open engineering question, not a procurement one.

Engineering feasibility unproven
OQ-07

Clean firm power may not arrive on schedule

Interconnection queues, transformer and turbine lead times, and the maturity of advanced nuclear and enhanced geothermal all sit outside our control. A 2030 CFE commitment that depends on projects energising in 2029 carries real slip risk, and interim gas capacity would visibly contradict the pathway.

Supply-side dependency
OQ-08

Assurance readiness is thinner than our ambition

Primary supplier data is sparse and inconsistently bounded; double counting between product footprints and the corporate inventory is easy to introduce and hard to detect. Our first assurance cycles will surface restatements. We would rather restate early and loudly than defend a number we cannot reproduce.

Expect restatements

08 / sources

References and standards

Figures quoted above are as published by the sources below at the time of writing; scenario ranges are the publishers', not ours. Targets in section 04 are internal proposals, not published commitments.

01IEA — Energy and AI (2025). Global datacentre electricity demand and 2030 projections.
iea.org/reports/energy-and-ai

02LBNL — 2024 United States Data Center Energy Usage Report.
eta.lbl.gov/publications/2024-united-states-data-center-energy

03Global E-waste Monitor 2024 (UNITAR / ITU).
ewastemonitor.info

04Apple — Environment & Product Environmental Reports.
apple.com/environment

05HP — Sustainable Impact reporting & product carbon footprints.
hp.com/us/en/hp-information/sustainable-impact.html

06ISO 14040:2006 — LCA principles and framework.
iso.org/standard/37456.html

07ISO 14044:2006 — LCA requirements and guidelines.
iso.org/standard/38498.html

08WBCSD PACT / Pathfinder Framework — interoperable product carbon footprint exchange.
carbon-transparency.org

09SBTi Corporate Net-Zero Standard.
sciencebasedtargets.org/net-zero

10SBTi ICT sector guidance.
sciencebasedtargets.org/sectors/ict

11MIT Materials Systems Laboratory — PAIA streamlined product carbon modelling.
msl.mit.edu

12ACT — Architectural Carbon Modeling Tool (Meta AI Research).
github.com/facebookresearch/ACT

13Green Software Foundation — Software Carbon Intensity specification.
sci.greensoftware.foundation

14ASHRAE Datacom series — thermal guidelines and liquid cooling.
ashrae.org/technical-resources/bookstore/datacom-series

15ENERGY STAR — Computers program requirements.
energystar.gov/products/computers

16WRI Aqueduct — water risk and basin stress mapping.
wri.org/aqueduct

17CDP Water Security disclosure framework.
cdp.net/en/water

18ISCC PLUS — mass-balance chain of custody for recycled materials.
iscc-system.org

19EPEAT / Global Electronics Council — sustainable IT product criteria.
epeat.net

20EU Ecodesign for Sustainable Products Regulation & Digital Product Passport.
commission.europa.eu — ESPR

21EU Corporate Sustainability Reporting Directive / ESRS.
finance.ec.europa.eu — CSRD

22California CARB — Climate Corporate Accountability (SB 253 / SB 261).
ww2.arb.ca.gov — climate corporate accountability

23EU Energy Efficiency Directive — including datacentre energy reporting obligations.
energy.ec.europa.eu — EED

24GHG Protocol — Corporate Value Chain (Scope 3) Standard.
ghgprotocol.org — Scope 3

25GHG Protocol standards update process (Scope 2 and market-based accounting revision).
ghgprotocol.org — update process

26ecoinvent — background life-cycle inventory database.
ecoinvent.org

27openLCA — open-source LCA modelling software.
openlca.org

28Boavizta — open environmental data for digital equipment.
boavizta.org/en

29Cloud Carbon Footprint — open-source cloud emissions estimation.
cloudcarbonfootprint.org

30Open Compute Project — open hardware, cooling and sustainability workstreams.
opencompute.org

31SEMI Semiconductor Climate Consortium.
semi.org — climate consortium

32imec — Sustainable Semiconductor Technologies and Systems.
imec-int.com — SSTS

33Gupta et al., "Chasing Carbon: The Elusive Environmental Footprint of Computing."
arxiv.org/abs/2011.02839

34Li et al., "Making AI Less 'Thirsty'" — water footprint of AI models.
arxiv.org/abs/2304.03271

35Luccioni et al., "Power Hungry Processing" — energy cost of inference.
arxiv.org/abs/2311.16863

36Masanet et al., Science (2020) — recalibrating global datacentre energy-use estimates.
science.org/doi/10.1126/science.aba3758

37Responsible Business Alliance — supply-chain audit and smelter due diligence.
responsiblebusiness.org

38Ellen MacArthur Foundation — circular economy frameworks.
ellenmacarthurfoundation.org

39Google — Environmental Reports (water stewardship & 24/7 CFE methodology).
sustainability.google/reports

40Microsoft — Sustainability Report (water positive & datacentre design disclosures).
microsoft.com — sustainability report