ChemSec PFAS Report Links AI Data Center Growth to a New Chemical Supply Boom

ChemSec has linked the AI infrastructure boom to expanding PFAS production, despite mounting health concerns, lawsuits, and proposed restrictions on these persistent chemicals. The ChemSec PFAS report says AI data centers, semiconductor manufacturing, and battery materials are driving new capacity among major producers.

That finding adds a less visible cost to the race for computing power. Public debate has focused on electricity, water, emissions, and local power prices. ChemSec argues that the physical supply chain also increases demand for fluorinated materials used in chips, cooling equipment, cables, and batteries.

The conflict is not simply AI companies against environmental advocates. Chemical and semiconductor manufacturers say some PFAS remain difficult to replace in demanding industrial applications. The central question is whether rapid infrastructure construction will lock in more production before regulators, buyers, and engineers can establish safer alternatives.

The ChemSec PFAS Report Tracks an Expanding Supply Chain

The report’s most important finding is that AI demand now appears in the investment logic of companies producing fluorinated chemicals.

ChemSec, formally the International Chemical Secretariat, published its updated review of major PFAS producers on September 13, 2026. PFAS means per- and polyfluoroalkyl substances, a large chemical family defined by highly durable carbon-fluorine structures.

The nonprofit examined ten producers, including AGC, Archroma, Arkema, BASF, Bayer, Chemours, Daikin, Orbia Fluor & Energy Materials, Solstice Advanced Materials, and Syensqo. It also discussed 3M because of the company’s historically important position in the industry.

ChemSec based its list on substances produced or imported in the European Union and United States. It supplemented those records with public corporate documents and its own databases. The organization acknowledges that the industry remains opaque, especially outside those two markets.

That limitation matters. The research does not provide a complete global production total, an emissions inventory, or a direct measurement of releases from AI data centers. It instead documents corporate strategies, identified substances, capacity projects, and the markets used to justify expansion.

The producer survey concludes that most of the companies are expanding PFAS capacity. It identifies AI and data center infrastructure, semiconductor manufacturing, and lithium-ion battery materials as the three main demand sources.

Those sectors overlap. Data centers require processors, memory, networking hardware, electrical systems, cooling equipment, backup power, and large supporting supply chains. PFAS can appear in the manufacturing processes, components, coatings, seals, coolants, and refrigerants connected to that infrastructure.

ChemSec highlights Daikin as one clear example. According to the report, Daikin plans to more than triple its fluoropolymer production capacity by 2027. The company expects semiconductor demand to support that expansion.

ChemSec also points to Solstice Advanced Materials, the company spun out of Honeywell in 2025. The report says at least half of Solstice’s revenue comes from PFAS, particularly refrigerant gases.

Solstice executives have described AI-related demand as a generational growth opportunity. That commercial framing is significant because it ties chemical investment decisions to expectations about sustained computing infrastructure growth.

Chemours presents another important case. ChemSec says the company produces or uses 57 listed PFAS and derives between half and two-thirds of its revenue from PFAS production. The report also identifies the company as a supplier to semiconductor and data center markets.

These examples do not prove that every planned AI facility will release PFAS into surrounding water or soil. They show that AI infrastructure has become part of the business case for producing more fluorinated materials.

That distinction keeps the story grounded. The immediate change is an expansion signal upstream, not a verified global spike in contamination measured at data center property lines.

The signal still deserves attention. Chemical capacity built for an expected AI market can operate for decades, while persistent pollution can remain after the original equipment becomes obsolete.

Why AI Data Center PFAS Demand Extends Beyond the Server Hall

AI data center PFAS demand runs through the entire hardware system, from chip fabrication to cooling and fire protection.

The servers inside an AI facility are only the final visible layer. Every accelerator contains semiconductor components that pass through complex fabrication processes before installation. Those processes depend on materials able to withstand heat, corrosive chemicals, and exceptionally strict purity requirements.

The semiconductor trade group SEMI says PFAS support nearly every part of the semiconductor value chain. Its industry explainer identifies uses in production equipment, fluid-management systems, facility infrastructure, and process chemicals used for photolithography.

Photolithography is the process that transfers microscopic circuit patterns onto a wafer. Some fluorinated substances provide the chemical stability, surface properties, and contamination control required during that work.

PFAS semiconductor manufacturing also extends to etching, deposition, chamber cleaning, seals, tubing, filters, pumps, and protective equipment. Not every application involves the same compound, exposure path, or health evidence.

That diversity creates a policy problem. PFAS describes thousands of substances rather than one uniform material. Their persistence, mobility, toxicity, molecular size, and industrial functions can differ substantially.

A broad exemption for the semiconductor sector can therefore shelter many uses from scrutiny. A broad prohibition can also disrupt applications for which qualified substitutes are not yet available.

Industry representatives emphasize this technical difficulty. Battelle’s manufacturing assessment says PFAS are deliberately engineered into semiconductor production and that many cannot be replaced on a near-term schedule.

The challenge does not end when the chips leave the factory. Some data center designs use fluorinated refrigerants or heat-transfer fluids. Fire-suppression equipment and other infrastructure can introduce additional fluorinated materials.

Two-phase immersion cooling receives particular attention because it places computing equipment inside a dielectric fluid that does not conduct electricity. Heat causes the fluid to boil, and a condenser returns the vapor to liquid form.

This design can move heat efficiently from dense racks. That makes it attractive for AI accelerators, which concentrate substantial computing activity and heat within limited physical space.

However, “two-phase fluid” does not automatically mean one specific PFAS product. Cooling vendors use different chemistries, and technology choices continue to change. Operators also use air cooling, direct-to-chip liquid cooling, and single-phase immersion systems.

The ChemSec PFAS report therefore supports a supply-chain warning, not the claim that every AI data center immerses servers in the same hazardous chemical. Treating those statements as interchangeable would exaggerate the evidence.

Direct releases from operating facilities remain difficult to quantify. Closed-loop cooling systems are designed to contain their working fluids, but leaks, maintenance, disposal, equipment failures, and end-of-life handling still require controls.

The indirect footprint is clearer. More chips require more fabrication, and more facilities require additional cooling and electrical equipment. Production sites can release PFAS through air emissions, wastewater, waste handling, or accidental loss.

An analysis from the Environmental and Energy Study Institute describes this distinction between direct and indirect PFAS pollution. It notes that direct cooling releases may be limited while upstream manufacturing creates significant concerns.

That is why the data center boundary can mislead. A facility may report modest on-site chemical use while depending on PFAS-intensive production elsewhere.

AI buyers rarely receive a complete chemical ledger for every accelerator, networking component, cooling unit, and manufacturing step. Without that information, procurement teams cannot easily compare one infrastructure design with another.

The emerging pressure falls on cloud providers, chip designers, fabrication companies, cooling vendors, and chemical producers together. Each controls only part of the system, but each can shift risk to another participant.

The Efficiency Promise Collides With Persistent Pollution

The core tradeoff is that materials selected for exceptional stability can remain exceptionally difficult to contain, destroy, or replace.

PFAS became useful because carbon-fluorine bonds resist heat, water, oil, and chemical attack. Those properties support precise manufacturing and reliable equipment. They also help many PFAS persist after entering the environment.

This persistence is the basis of the “forever chemicals” label. However, it does not mean every compound remains unchanged forever or carries an identical cancer risk.

Health language requires particular care. Researchers have associated exposure to certain PFAS with several adverse effects, but evidence varies by compound, dose, exposure route, and studied population.

The International Agency for Research on Cancer classifies PFOA as carcinogenic to humans and PFOS as possibly carcinogenic to humans. Those classifications should not be extended automatically to every member of the PFAS family.

The US National Institute of Environmental Health Sciences says PFAS exposure has been associated with altered metabolism, fertility changes, immune effects, and increased risks of some cancers. Its health overview also stresses that scientists are still studying the effects of different mixtures and exposure levels.

The headline phrase “cancer-causing forever chemicals” captures a serious concern but compresses a complicated evidence base. A more accurate conclusion is that some well-studied PFAS have cancer links, while thousands of others lack equivalent toxicological data.

The absence of complete data does not establish safety. It also does not justify treating every compound as equally carcinogenic. Effective policy must manage both uncertainty and variation.

ChemSec takes a class-based approach and argues for universal restrictions with temporary, narrowly defined exemptions. The group believes partial rules will be overwhelmed by increased production tied to AI and other expanding markets.

Manufacturers and semiconductor organizations generally favor differentiated regulation. They argue that essential industrial uses should receive exemptions until technically qualified substitutes exist.

That disagreement forms the main opponent structure in the ChemSec PFAS report. One side sees broad restrictions as necessary because substance-by-substance regulation moves too slowly. The other sees broad restrictions as a threat to critical manufacturing.

The AI boom raises the stakes because it compresses decision timelines. A chipmaker cannot replace a process chemical by substituting a similar liquid during normal production.

New materials must meet purity, performance, compatibility, safety, and reliability requirements. They must then pass testing and customer qualification without damaging manufacturing yields.

Cooling systems face their own qualification demands. A fluid must remove heat, remain compatible with electronic components, limit fire risks, and operate predictably across thousands of hours.

These constraints explain why some companies call certain PFAS essential. They do not eliminate responsibility for emissions, disclosure, recovery, or investment in substitutes.

The environmental case also extends beyond the final material. Some fluoropolymers may be relatively immobile during normal product use, yet their production can involve hazardous processing aids, emissions, or waste streams.

End-of-life handling adds another gap. Retired servers, cooling fluids, filters, pipes, and industrial equipment can enter waste systems that were not designed to identify every fluorinated compound.

Destroying PFAS can require high temperatures, specialized treatment, or energy-intensive processes. Conventional wastewater treatment does not reliably remove or destroy the entire class.

This leaves communities facing costs that are not visible in a data center’s electricity contract. Testing wells, treating drinking water, managing contaminated soil, and monitoring emissions can continue long after construction ends.

The efficiency promise is therefore incomplete when evaluated only inside the server hall. A cooling design can reduce water or electricity use while shifting chemical risks upstream or into future waste management.

That does not prove the design is worse overall. It means operators need life-cycle evidence before presenting efficiency gains as an unqualified environmental benefit.

Production Plans Are Moving Faster Than Disclosure

The largest verification gap is not whether PFAS appear in digital infrastructure, but how much is produced, released, recovered, and destroyed.

ChemSec explicitly says reliable global production volumes are unavailable. Companies can also protect chemical identities or quantities as confidential business information, limiting independent analysis.

That opacity makes precise attribution difficult. Researchers may identify PFAS near a manufacturing site without possessing complete records of every substance used, transformed, or emitted there.

Data center operators disclose substantial information about energy, carbon, and sometimes water. Comparable reporting on fluorinated chemicals remains uncommon.

A credible AI data center PFAS inventory would need to separate several categories. These include chemicals used at the site, chemicals embedded in equipment, refrigerants, fire suppressants, upstream fabrication inputs, and waste sent elsewhere.

It would also distinguish purchased volume from environmental release. A contained chemical inside a sealed component does not create the same exposure pathway as an air emission or wastewater discharge.

The United States has begun regulating selected PFAS through several legal channels. In 2024, the Environmental Protection Agency established drinking-water limits for six PFAS and designated PFOA and PFOS as hazardous substances.

Federal policy has continued to shift. A 2025 Government Accountability Office review noted that EPA leaders planned to retain limits for PFOA and PFOS while reconsidering several others.

The same federal review found gaps in how Department of Energy sites documented historic and current PFAS use. GAO recommended deadlines for completing those reviews and identifying additional investigation needs.

That example matters for AI infrastructure because it illustrates a broader governance problem. Regulators cannot manage releases effectively when facilities lack complete chemical histories.

Europe has considered a much wider PFAS restriction under its chemicals framework. The proposal has produced extensive debate over exemptions, transition periods, essential uses, and the availability of alternatives.

The ChemSec PFAS report enters that debate with a warning about timing. If manufacturers expand capacity before restrictions take effect, regulators will confront larger investments and stronger claims of economic dependence.

The report also contains evidence that an exit is possible in parts of the market. 3M completed its planned departure from PFAS manufacturing at the end of 2025, according to ChemSec.

BASF has said it will phase out most products formulated with PFAS by 2028, excluding pesticides. Archroma has reduced identified PFAS production and markets alternatives in several product categories.

These examples do not establish that advanced chip fabrication can immediately eliminate every PFAS use. They show that production strategies are choices rather than an unavoidable consequence of chemistry.

Other producers are moving in the opposite direction. ChemSec describes capacity additions linked to semiconductors, batteries, refrigerants, and thermal management.

That divergence creates a market test. If buyers demand chemical disclosure and safer substitutes, companies reducing PFAS dependence may gain an advantage. If buyers prioritize rapid delivery alone, expansion plans receive stronger support.

Cloud companies have considerable purchasing influence. Their technical specifications can shape chip packages, cooling architectures, materials, and supplier reporting requirements.

Yet procurement pressure works only when customers ask consistent questions. A general sustainability pledge does not reveal which PFAS are present or where releases occur.

Operators need substance-level inventories where possible, total fluorine screening where identities remain uncertain, and clear procedures for leaks and disposal. Suppliers also need measurable reduction targets.

Public reporting should identify the methodology and its limitations. A company should not claim “PFAS-free” status when its assessment excludes upstream semiconductor production or confidential supplier formulations.

At the same time, campaigners should avoid presenting every proposed data center as a confirmed contamination site. Site-specific testing and verified release data remain essential.

The strongest case for action rests on what is already established. PFAS are deeply integrated into semiconductor supply chains, several producers are expanding, and comprehensive disclosure remains weak.

Semiconductor Makers Face the Hardest Substitution Test

PFAS semiconductor manufacturing is where the demand story becomes both technically credible and hardest to resolve.

Advanced chips depend on repeated processing steps performed at extremely small scales. Contamination that appears insignificant elsewhere can ruin wafers, lower yields, or damage expensive equipment.

Fluorinated materials can supply low surface tension, chemical resistance, thermal stability, electrical insulation, and high purity. Those properties are difficult to reproduce simultaneously.

A substitute must perform its intended function without creating another serious hazard. Replacing a persistent material with a more toxic or volatile chemical would not represent meaningful progress.

Engineers also need to determine whether an alternative increases water, energy, or material use. Life-cycle comparisons should capture those tradeoffs instead of focusing on a single chemical label.

This technical challenge supports targeted transition periods. It does not support indefinite exemptions without disclosure, reduction plans, or research milestones.

The semiconductor industry has already organized research around measurement and treatment. Industry groups are examining total fluorine methods, waste streams, process changes, and potential substitutes.

Measurement remains complicated because fabrication facilities use complex chemical mixtures. PFAS can appear at very low concentrations across liquid, solid, and gaseous waste.

Some compounds can transform during manufacturing or treatment. Others are difficult to detect using standard targeted tests, which search only for substances already included in a laboratory method.

This creates the risk of regrettable substitution. A company can replace a regulated PFAS with a less studied fluorinated substance while retaining similar persistence concerns.

Class-wide screening can reveal more total fluorine, but it may not identify individual compounds or their toxicity. Targeted analysis provides specificity but can miss unknown materials.

Manufacturers therefore need both approaches. They also need mass-balance accounting that compares chemicals entering a facility with those recovered, destroyed, shipped as waste, or released.

The AI connection intensifies this challenge through scale. Even if chemical use per chip declines, total demand can increase when the number and complexity of accelerators grow faster.

Newer chips can also improve computing efficiency while encouraging additional deployment. This rebound effect means better performance does not guarantee lower total material consumption.

Large customers can influence the transition by asking foundries for verified chemical inventories and abatement data. They can also fund qualification work for safer materials.

Chip designers do not directly control every fabrication chemical, especially when manufacturing is outsourced. However, they control supplier selection, contracts, design requirements, and long-term capacity commitments.

Data center operators face a similar choice. They can evaluate cooling systems by total environmental performance, including fluid production, leakage, recovery, and end-of-life treatment.

Regulators can support this work by defining comparable reporting rules. Without consistent definitions, one company may report only direct releases while another includes purchased substances and upstream impacts.

Researchers need access to real operational data. Models based on confidential or incomplete inventories cannot reliably estimate community exposure.

Communities near fabrication and chemical plants need monitoring before expansion begins. Baseline measurements allow future changes to be detected rather than disputed after contamination appears.

Workers also require attention. Occupational exposure can differ from community exposure because employees may handle concentrated chemicals, maintenance waste, or contaminated equipment.

The hardest substitution cases should receive focused research, not a permanent shield. Time-limited exemptions can require companies to demonstrate why a use remains essential and what alternatives they tested.

That approach would place the burden of evidence on the user seeking an exemption. It would also give suppliers a clearer reason to invest in replacement technologies.

The central risk is lock-in. Once new plants, processes, and service contracts depend on a material, changing them becomes more expensive and politically difficult.

Three Signals Will Show Whether the Warning Changes the AI Build-Out

The next test is whether the ChemSec PFAS report produces measurable changes in disclosure, procurement, and regulation.

The first signal is substance-level reporting from major cloud, chip, and cooling companies. Investors and communities need more than general statements about responsible chemical management.

A meaningful disclosure would identify direct PFAS use, supplier coverage, exclusions, release pathways, waste handling, and measurement methods. It would also distinguish verified data from estimates.

Such reporting would strengthen ChemSec’s warning if it reveals rising purchases or significant untracked use. It would weaken the broadest claims if operators document limited use, effective containment, and declining upstream demand.

The second signal is a binding procurement response from large AI infrastructure buyers. Cloud providers can require suppliers to disclose fluorinated substances, test alternatives, and document recovery systems.

A procurement rule would matter because it changes demand before regulation finishes. Suppliers respond quickly when their largest customers make chemical performance part of qualification.

The strongest policies would cover semiconductor fabrication, cooling equipment, refrigerants, fire suppression, maintenance, and disposal. Narrow rules limited to on-site coolants would miss much of the footprint.

If large buyers continue signing capacity agreements without chemical conditions, PFAS producers will have little reason to revise their expansion plans. Voluntary sustainability language alone will not counter that market signal.

The third signal is the design of regulatory exemptions in the United States and Europe. The key issue is whether exemptions remain temporary, application-specific, transparent, and tied to substitution work.

Open-ended sector exemptions would support ChemSec’s concern that industrial expansion can outrun restrictions. Carefully reviewed essential-use exemptions would create pressure without demanding unsafe, immediate substitutions.

Regulators should also require emissions and waste reporting during any exemption period. Otherwise, an exemption becomes a blind spot rather than a managed transition.

Readers should resist two premature conclusions. The report does not show that every AI data center is contaminating nearby drinking water. It also does not support treating chemical supply chains as somebody else’s problem.

The verified concern sits between those extremes. AI infrastructure depends on material systems with persistent chemical risks, while production decisions are moving faster than public accounting.

Developers and AI users cannot audit a fabrication plant each time they run a model. They can still ask providers for evidence about hardware sourcing, cooling fluids, supplier standards, and disposal.

Enterprise buyers have greater leverage. Requests for proposals can include chemical disclosure, refrigerant management, leak detection, recovery plans, and supplier reduction targets.

Investors can examine whether chemical producers treat litigation and regulation as material risks. They can also compare expansion plans with spending on alternatives and pollution controls.

Policymakers should demand baseline monitoring around new chemical and semiconductor facilities. Waiting for a disputed contamination pattern makes prevention more expensive and evidence harder to reconstruct.

The ChemSec PFAS report has shifted the data center debate from operational resources to industrial chemistry. That shift will matter only if buyers and regulators convert it into verifiable requirements.

Ask your cloud or infrastructure provider one direct question: which PFAS does this system depend on, and where are they released? If the answer is unavailable, that information gap is itself a risk signal.