The gases that never break down, on the one bill of materials that never counts them.
A single AI accelerator carries fluorine through six companies before it reaches a rack. It is etched with C₄F₈ and CF₄, its chambers cleaned with NF₃, its wafers shipped from fabs whose abatement uptime nobody outside the fab has ever verified.
It is then cooled by HFC blends, switched by SF₆ gear, and protected by fluorinated fire suppressant — before an AI PC repeats a smaller version of the same story on a desk.
None of this is illegal. Almost none of it is visible. This is the plan to make it visible, then make it small, then make it unnecessary.
F-gases are a fraction of a percent of hardware tonnage and can be a double-digit percentage of a fab's climate footprint. Global warming potential is a multiplier, and the multipliers here are absurd. AR6 GWP-100 values, logarithmic scale.
Read the right-hand column twice. CF₄ emitted while etching a 2026 accelerator will still be warming the atmosphere in the year 52,000. There is no removal technology for it, no sink, no decay path that matters on a human timescale. Abatement at the tool is not one option among several — it is the only intervention that exists.
Follow one accelerator and one notebook mainboard through the chain. At every tier the gas is real, the emission is real, and the accountability quietly changes hands — usually into a gap.
Acid-grade CaF₂ mining (China, Mexico, Mongolia, South Africa) → anhydrous HF → F₂ → on-site NF₃ and PFC synthesis.
Linde, Air Liquide, Air Products, Resonac, Kanto Denka, SK Materials, Central Glass — plus fluorinated heat-transfer fluid makers.
Logic and HBM fabs, panel makers, OSAT. Plasma etch, dielectric deposition, remote-plasma chamber clean, wet-bench fluoropolymers.
Server ODMs, PSU and UPS makers, chiller and CDU vendors, MV switchgear suppliers, clean-agent fire suppression integrators.
Chiller plant, CRAH/CRAC, liquid-cooling loops, MV/HV switchgear, clean-agent suppression, and the maintenance contractors who touch all of it.
Millions of devices; vapour-chamber and thermal materials; the recovery event that decides whether decades of charge is reclaimed or vented.
Nobody is lying. The number simply doesn't exist yet. The actual state of F-gas Scope 3 in 2025
Before commissioning anything, mine what is already disclosed. Six bodies of evidence, read against each other, get you to a defensible first baseline in one quarter. All sources are linked in section 07.
EU Regulation (EU) 2024/573 replaced the 2014 F-gas rule with a steeper HFC phase-down to zero placing-on-market by 2050, product bans stepping through 2026–2035, and a schedule that ends SF₆ in new medium-voltage switchgear. The US AIM Act drives an 85% HFC phase-down by 2036 with sector-specific Technology Transitions and Refrigerant Management rules. The Kigali Amendment binds both. ECHA's universal PFAS restriction dossier, meanwhile, targets the same molecular family from the other direction.
→ Read as a design constraint, not a compliance task: it tells you which of today's equipment is a stranded asset in 2031.
TSMC, Intel, Samsung, Micron and SK hynix all publish F-GHG or PFC emissions and abatement coverage in annual ESG reports, typically citing destruction-and-removal efficiency (DRE) targets and point-of-use abatement installation rates. Intel and TSMC both report multi-year absolute or normalised F-GHG reduction programmes.
→ Stops at: site totals, no per-process or per-product allocation, no third-party verification of abatement uptime as distinct from installation.
Google, Microsoft, Meta and AWS all publish Scope 1, and refrigerant fugitives are increasingly broken out. Colocation providers vary widely. Comparing disclosed fugitive tCO₂e against disclosed IT load reveals order-of-magnitude differences that are methodological, not physical.
→ Stops at: leak rates derived from factors rather than measured charge reconciliation; contractor-held service data.
Apple's supplier F-GHG abatement programme is the existence proof that a downstream buyer can compel upstream fluorinated-gas abatement contractually, at scale, across semiconductor and display suppliers. Dell, HP and Lenovo publish supplier engagement and product carbon footprints for notebooks.
→ Stops at: programme-level claims rather than gas-level, audited, per-supplier abatement performance data anyone can re-compute.
The World Semiconductor Council has run industry-wide PFC emission reduction goals since 1999. SEMI's Semiconductor Climate Consortium convenes 100+ members across Scope 3 and process-gas workstreams. The Open Compute Project and iMasons Climate Accord carry the datacenter hardware side. CIGRE holds the deep technical record on SF₆-free switchgear.
→ Stops at: voluntary, normalised, aggregate goals — with no mechanism to bind a specific supplier to a specific buyer's product.
NOAA GML and AGAGE measure actual atmospheric concentrations of NF₃, SF₆, CF₄ and HFC-23. Peer-reviewed top-down studies have repeatedly found global emissions exceeding the sum of reported bottom-up inventories for several of these gases.
→ Use it as your reality check: if the chain's reported total can't reconcile with the atmosphere, your inventory is optimistic.
Each wall is stated honestly, then given the one move that turns it from a blocker into a workstream. No wall is solved by asking a supplier for a spreadsheet.
You buy an accelerator from a systems vendor, who buys a package from an OSAT, who receives die from a foundry. Three contracts away, with confidentiality at every hop, your leverage is nominally zero — and the foundry has hundreds of customers.
Fab F-GHG is a site-level output of hundreds of process steps shared across many products. Allocating it to your die requires mask-layer-level process data that is the foundry's core IP.
A fab can truthfully report 95% abatement coverage while real destruction efficiency drops with burner degradation, bypass during maintenance, low-flow conditions and unmonitored tool events. DRE is measured in qualification, not in production.
Thermal point-of-use abatement burns fuel or electricity and can produce NOx and acid gases requiring scrubbing. In a grid-constrained region this competes directly with your energy targets, and plasma abatement adds load.
Normalised emission rate per wafer can fall 30% while your absolute chain F-GHG doubles, because AI capacity is doubling faster. Every intensity-based goal in the industry currently hides this.
The best low-GWP switchgear alternatives, the best two-phase immersion fluids, and much of the wet-process hardware are fluorinated. A universal PFAS restriction and a GWP phase-down can point in opposite directions for the same purchase.
A new chamber-clean chemistry, a new immersion fluid, or a new refrigerant in a chiller does not enter production because sustainability asked. It enters after yield, reliability, safety-code and warranty qualification — three to five years.
At decommissioning, venting a charge is free and fast; recovery costs money and time. The technician's incentive and the operator's incentive both point the wrong way, and no one audits the difference.
A decade is not one plan. It is a visibility programme that must succeed before an abatement programme can be enforced, which must succeed before substitution has anywhere to land. Each horizon names its moves, its evidence, and the number it moves.
Objective: a defensible, verified, product-attributed F-GHG baseline covering ≥80% of hardware spend — and contract language that makes the data non-optional from the next award onward.
Extend the existing BOM/PLM schema with a fluorine flag at part level and a gas register at site level.
A single, machine-readable F-GHG data request that Tier 1 flows to Tier 2, aligned to existing standards so it is answerable.
Stop reporting fugitives from default leak factors. Move to purchase-minus-inventory reconciliation across all sites.
Insert F-GHG terms into every master supply agreement renewing in 2026, before pricing is locked.
A single buyer cannot change a foundry's roadmap. Four aligned buyers with one specification can.
An unverified baseline will be litigated later. Get limited assurance on the F-GHG line specifically.
Objective: absolute attributed F-GHG below the 2026 baseline despite capacity growth — bought with procurement leverage, abatement performance floors and a hard exit from SF₆ and high-GWP refrigerant in everything new.
Apply the internal carbon price to attributed F-GHG in every sourcing decision and capex approval.
Move from "abatement installed" to guaranteed destruction performance, verified in production.
The regulation is arriving regardless; front-run it and avoid a decade of stranded switchgear.
Every new thermal system designed around low-GWP working fluid and lower charge, with liquid cooling as the structural win.
Abatement retrofit in an older fab or an OSAT is real capex against thin margin. Share it, and take the reductions.
Recovery becomes a paid, evidenced, audited deliverable across every decommissioning event.
Objective: fluorine becomes a design decision made at architecture stage, not a facilities problem discovered at commissioning. The chain's residual F-GHG is small, measured, and shrinking in absolute terms.
The qualification cycles funded in 2026–2028 come to production. This is where the deep reductions live.
Give silicon and system architects the number at the point they can still change it.
Anticipate restriction rather than react to it, and never re-qualify the same component twice.
By now the near tiers are managed. The remaining unpriced risk is in gas synthesis and feedstock.
The dual-lock target is the whole point: capacity can grow, the attributed absolute number cannot.
A methodology only one company uses cannot verify anything. Give it away and let it become the norm.
Every large upstream change in this industry has moved through the same four mechanisms. None of them is persuasion.
One buyer asking a foundry for per-layer F-GHG data is a nuisance request. Four buyers representing a large share of leading-edge wafer starts, asking with one identical specification, is a roadmap item. The cost of answering collapses when the question is asked once.
You cannot make a supplier fear your sustainability team. You can make them plan around the EU F-gas phase-down schedule, the AIM Act, and a PFAS restriction. Align your contractual dates just ahead of the regulatory ones and the argument stops being about ambition.
Abatement retrofit, SF₆-free switchgear and fluid substitution are capex against thin supplier margin. Co-investment tied to multi-year offtake converts a cost centre into a financeable project — and gives you a defensible claim on the reduction.
Nothing unlocks upstream disclosure faster than a buyer who has already disclosed its own measured leak rate, including the year it was terrible. It removes the fear of asymmetric exposure and converts the request from an audit into a shared problem.
Primary regulation, accounting standards, association method work, company disclosure hubs and atmospheric measurement. Read the first group before writing a single contract clause.