Interview review · Global sustainability, supply chain & product
Five years of program shipped. Five years now open — 2026 to 2030 — in the two loudest load curves on the planet: the AI PC on the desk and the AI datacenter behind it. This is a working roadmap: global goals, industry and market benchmarks, supply-chain data intelligence, LCA modelling, and the business decisions that actually move tonnes. Research and refresh only. Nothing confidential. Everything sourced.
No one decarbonizes alone. A roadmap only one company can read isn't a roadmap — it's a wish with a logo on it.
How to read it: the weights are illustrative planning weights for the review conversation, not commitments and not company data. The point of the shape is that the top wedge is grid-shaped — largely outside any single firm's factory — and the bottom wedge is a cost, not an achievement.
Site — Global goals
Paris and AR6 set the budget; SBTi turns it into a corporate rate; GHG Protocol decides what counts and how. The live edges for 2026–2030 are the Net-Zero Standard V2 debate, the Scope 2 market-based revision, and whether ICT keeps its own −45%-by-2030 trajectory.
Site — Industry benchmarking
PUE has flattened; the useful metrics now are perf-per-watt under a declared workload, WUE in stressed basins, and product carbon footprints computed the same way twice. Semiconductor Climate Consortium, OCP and MLCommons are where those definitions get argued in public.
Site — Market benchmarking
The AI PC refresh cycle put TOPS-per-watt on the spec sheet; enterprise and public tenders put hourly CFE, PCF per SKU and repairability beside price. Marketing claims that cannot survive an ESPR-style substantiation test are becoming a liability, not a differentiator.
Site — Supply chain data intelligence
The 2021–2025 program ran on averages. The next one cannot: primary supplier energy and process-gas data, PACT-conformant product footprints exchanged machine-to-machine, hourly certificates where the claim is hourly, and assurance that survives an auditor.
Site — LCA modelling & data
ISO 14040/44/14067 give the frame; ecoinvent and the JRC EF datasets fill the gaps; the gaps are exactly where AI hardware sits — advanced nodes, HBM stacks, substrate, and grid mixes that change every hour. Publish ranges and the cut-off rules, not single hero numbers.
Site — Business decision
A marginal abatement cost curve rebuilt on measured data, an internal carbon price with teeth at the capex gate, PPA versus on-site versus tariff structures, and product trade-offs stated openly: thinner, or longer-lived? Cheaper cell, or lower embodied carbon?
2026
The disclosure floor rises: first mandatory scope 1 and 2 reporting under California SB 253 for the covered year, climate-risk reporting under SB 261, EU reporting reshaped by the omnibus revisions, and annual datacentre energy and water reporting under the Energy Efficiency Directive. Baseline year re-cut on measured supplier data, with the restatement published.
Anchor: SB 253 · SB 261 · EED datacentre reporting2027
PCF per SKU produced by the top suppliers on the same rules, exchanged in a PACT-conformant format instead of a spreadsheet attachment. Ecodesign work-programme measures and digital product passport groundwork turn product data from a marketing asset into a compliance object.
Anchor: ESPR working programme · PACT Pathfinder2028
Priority sites move from annual matching to hourly carbon-free energy with granular certificates; flexible AI load becomes a grid service rather than a grid problem. Embodied carbon moves onto the product spec sheet next to weight and battery life.
Anchor: EnergyTag · 24/7 CFE Compact · EPRI flexibility work2029
Low-carbon aluminium, responsibly-produced steel, recovered-content plastics and fibre packaging under multi-year contracts — because supplier fab decarbonization needs demand certainty before it needs a pledge. Abatement cost curve rebuilt on three years of measured data.
Anchor: ResponsibleSteel · ASI · ELYSIS2030
Near-term target date. Whatever abatement did not reach gets stated as a number, not absorbed into a claim — then addressed with durable removals under a public MRV standard, priced and disclosed separately from reductions.
Anchor: Puro.earth · Isometric · FrontierMove one
Re-cut the 2025 close on measured supplier data, publish the restatement and the cut-off rules, and state uncertainty as a band. A roadmap on a stale baseline optimises the wrong wedge for five years.
Move two
One footprint format, one exchange protocol, one assurance route — so a number can travel from a fab to a tender response without a human retyping it.
Move three
Marginal abatement cost curve, internal carbon price at the capex gate, and an explicit residual line. Decide in 2026 what 2030 will cost.
Move four
Put the modelling assumptions, benchmark definitions and open questions where peers, suppliers and customers can argue with them. Shared method is faster than shared pledges.
Grouped the way the roadmap is built. Figures quoted on this page come from the IEA, ITU and Uptime Institute sources below; everything else is method, standard or benchmark material.
Read peer reports for method, not for score. The interesting differences are in boundary, matching and data quality statements — not in the headline percentage.