Where can PFAS enter the AI data-centre supply chain?
A peer-reviewed review maps potential releases from chipmaking, cooling and hardware disposal. AI growth increases infrastructure demand, but the paper does not calculate a new AI-attributable PFAS emissions total.
The short version
- What happened
- The review maps PFAS pathways across semiconductor manufacturing, cooling systems, wastewater, gaseous emissions and end-of-life hardware.
- Who it affects
- Workers and nearby communities may face different exposure pathways across manufacturing, operation and disposal; this review maps the routes but does not measure individual exposure or disease.
- What remains unproven
- Review of existing evidence rather than a new emissions inventory
Keep in mind
Review of existing evidence rather than a new emissions inventory
No measured AI-specific PFAS total
Application-specific material-flow data remain incomplete
Substitution feasibility varies by safety, performance and regulation
No site-level exposure or health-outcome study
Source analysis · 1 linked source. Primary analysed source date: 26 October 2026.
Read the full report · 4 min The evidence, methods, implications and limitations
The footprint begins before a server is switched on
Per- and polyfluoroalkyl substances are valued for chemical stability, heat resistance and surface properties—the same persistence that makes releases difficult to manage. The review follows them through three parts of data-centre infrastructure: semiconductor fabrication, cooling equipment and the recycling or disposal of hardware. Possible pathways include manufacturing wastewater and solvents, coolant evaporation or leakage, and residues released when PFAS-containing components are dismantled.
AI matters because larger computing fleets increase demand for chips, cooling and hardware replacement. That makes AI an infrastructure multiplier in the paper, not a chemical fingerprint. The authors do not report a new measured total for PFAS emitted specifically by AI workloads, and a data centre’s electricity or water footprint cannot be converted into PFAS releases without application-specific material-flow evidence.[1]
Controls differ by waste stream
Dilute wastewater may be treated with activated carbon adsorption or ion exchange; concentrated liquids and solids can require high-temperature destruction; gaseous streams may call for thermal or catalytic oxidation. Each option moves through capture, concentration and verified destruction, and can add energy use or secondary emissions. A treatment that removes PFAS from water has not solved the problem if the captured material is later released.
Source reduction is therefore central. The review describes process-compatible substitutes in chip fabrication, PFAS-free cooling options, lower coolant inventories, recovery and reuse, simpler fabrication steps, more efficient computing and longer hardware life. But substitution is not a single switch: alternatives must meet performance, fire-safety and compatibility requirements, while regulations may retain exemptions for applications without validated replacements.[1]
The distinction that matters
A lifecycle, not a single exhaust pipe
PFAS can move through several physical forms and facilities before data-centre operation.
Manufacture
Photoresists, solvents and component processes
Wastewater, solvent and gaseous pathways
Operate
Cooling inventories and leakage control
Recovery and PFAS-free design options
End of life
Reuse, recycling and disposal
Residual coolants and components need controlled handling
Procurement can ask for flows before claiming reductions
For communities and workers, the relevant questions sit across the supply chain: which PFAS are used, in what quantities, where losses occur, how waste is captured and whether destruction is verified. Operators purchasing AI infrastructure can request material declarations, coolant-loss records, end-of-life handling and phase-out roadmaps rather than relying on a broad claim that a facility is efficient.
The review is peer reviewed and integrates engineering, treatment and policy evidence, but it is not an exposure study and does not measure health outcomes around a particular fab or data centre. Its central evidence gap is itself practical: more application-specific flow data and supply-chain transparency are needed to identify hotspots and compare interventions. Measured inventories and independent release monitoring would change the assessment most.[1]