ChemSec PFAS AI Warning Puts the Data Center Boom on a Chemical Collision Course
ChemSec has issued a PFAS AI warning after finding that most of the world’s 10 largest producers plan to expand forever-chemical production. The group connects those plans directly to rising demand from semiconductor factories, AI hardware, and data center cooling systems.
The warning exposes a less visible cost of the computing boom. AI companies promise more efficient infrastructure, while their suppliers prepare to manufacture additional chemicals that persist in the environment. Some PFAS have also been associated with serious health effects.
The resulting conflict is larger than one disputed cooling technology. Chemical producers and semiconductor companies argue that many fluorinated materials remain essential. Campaigners counter that new capacity will lock in pollution while governments are still deciding how tightly to regulate the entire chemical class.
PFAS Producers Are Expanding Around AI Demand
The immediate change is a coordinated expansion of fluorinated materials across several parts of the AI hardware supply chain.
ChemSec published its updated assessment of major PFAS producers on September 13, 2026. Its research found that most companies in the group were increasing capacity or pursuing related growth opportunities.
PFAS are per- and polyfluoroalkyl substances, a broad chemical family defined by exceptionally durable carbon-fluorine bonds. Those bonds provide resistance to heat, water, oil, corrosion, and aggressive industrial chemicals.
The same durability creates the environmental problem. Many PFAS persist for very long periods after entering water, soil, air, waste streams, or living organisms. This persistence produced the widely used label “forever chemicals.”
The ChemSec producer survey ties current expansion to three demand centers. These are AI and data center infrastructure, semiconductor manufacturing, and lithium-ion battery materials.
The companies discussed span Europe, Asia, and North America. They include AGC, Arkema, Chemours, Daikin, Orbia Fluor & Energy Materials, Solstice, and Syensqo. ChemSec also identifies other large producers with different strategies or degrees of dependence.
Several corporate announcements reinforce the connection to computing infrastructure. Daikin has discussed a data center hub built around fluorine products. The company also plans to expand capacity serving semiconductor applications.
ChemSec says Daikin intends to more than triple fluoropolymer capacity by 2027. Fluoropolymers are polymeric PFAS used in coatings, seals, pipes, cables, equipment, and other demanding applications.
Arkema has taken another route into the same market. The French chemical company opened a new production unit in Kentucky for a fluorinated refrigerant used in thermal-management applications.
Arkema described the investment as a response to expanding data center cooling needs. The facility involved an announced investment of $60 million.
Chemours is promoting liquid cooling products for advanced data centers and AI hardware. It says its approach can reduce energy, water, space, maintenance, and infrastructure requirements.
Those benefits remain company claims, not a complete environmental accounting. A cooling fluid can improve efficiency inside a data center while creating risks during manufacturing, leakage, maintenance, disposal, or chemical breakdown.
The original investigation also identified AGC, Dongyue, Gujarat Fluorochemicals, HaloPolymer, Orbia, Solstice, and Syensqo among producers increasing output.
That geographic spread matters. This is not a single factory expansion that one local permit can settle. It is an international supply response serving a distributed market for chips, cooling equipment, batteries, and electronic components.
The trend also extends beyond chemicals consumed directly inside operating data centers. AI demand first reaches chip designers, cloud providers, and server manufacturers. It then travels upstream into fabrication equipment, specialty gases, coatings, filters, pipes, seals, and process chemicals.
Every additional layer makes responsibility harder to assign. A model developer may never purchase PFAS directly. Its infrastructure plans can still encourage suppliers several tiers away to expand their production.
The ChemSec PFAS AI warning therefore concerns induced demand. AI growth is changing investment decisions throughout the physical supply chain, including decisions that will outlast the current generation of accelerators.
This creates the article’s central tension. The industry treats fluorinated chemistry as an enabling material for efficiency and reliability. Campaigners treat the same expansion as preventable pollution with long-term public costs.
Why AI Hardware Depends on Forever Chemicals
PFAS demand rises because modern computing depends on materials that tolerate conditions ordinary plastics and fluids cannot reliably survive.
Semiconductor fabrication requires extraordinary chemical control. Manufacturers repeatedly coat, expose, etch, clean, deposit, rinse, and inspect silicon wafers while maintaining extremely low contamination levels.
PFAS can appear in process chemicals and in the equipment surrounding those chemicals. Applications include photoresists, photoacid generators, plasma processes, heat-transfer fluids, filters, gaskets, valves, pumps, tubing, and corrosion-resistant pipe linings.
Photolithography transfers microscopic patterns onto a wafer. Some fluorinated substances help control surface behavior during that process, supporting the precision needed at increasingly small dimensions.
Plasma etching removes selected material to form chip structures. Fluorinated gases and resistant equipment components can withstand or participate in the harsh chemical environment required by that step.
Fluoropolymers also line systems that move ultrapure water and aggressive chemicals. Their stability helps prevent corrosion, contamination, or component failure inside expensive fabrication plants.
An investor note cited in the Guardian estimated that forever chemicals can appear across as many as 1,000 distinct chipmaking steps. That figure illustrates the integration challenge, although individual fabs use different processes and materials.
Removing one substance is not equivalent to replacing a household coating. A substitute must work with tightly connected equipment, materials, safety systems, and process recipes without reducing yield.
Yield measures the share of manufactured chips that meet required specifications. A small decline can waste wafers, chemicals, energy, water, and production time.
The Semiconductor Industry Association says high-volume fabrication requires near-perfect performance at each stage. Its members argue that a material change can take years of research, validation, and production qualification.
According to the semiconductor consortium, viable substitutes remain unavailable for many applications. The organization supports restrictions on nonessential uses but seeks exemptions where alternatives cannot yet meet manufacturing requirements.
That position represents the strongest counterargument to ChemSec. A rapid blanket prohibition could disrupt chip production before qualified replacements reach the market.
AI intensifies this dependency because its computing systems use large quantities of advanced logic, memory, networking, and power-management hardware. More accelerators also require supporting servers, switches, storage, and electrical equipment.
The chemical footprint consequently begins before a server reaches a data center. It includes the upstream manufacturing infrastructure used to fabricate and package its components.
Cooling adds another route. High-density AI racks concentrate substantial computing activity into limited space. Removing their heat becomes a design constraint rather than a routine facilities task.
Liquid cooling moves heat more effectively than air in many dense configurations. Systems can circulate coolant through cold plates attached to components or place equipment within an electrically nonconductive fluid.
Two-phase immersion cooling uses a fluid with a low boiling point. Hardware heats the liquid until it becomes vapor, carrying thermal energy toward a condenser.
The condenser cools the vapor back into liquid, which returns to the tank. This phase change can remove heat while limiting direct water consumption inside the primary cooling loop.
Some fluorinated fluids fit this use because they resist ignition, conduct little electricity, and remain stable around electronic components. Those properties can reduce equipment risks and support compact designs.
However, not every liquid-cooling system uses PFAS. Direct-to-chip systems commonly use water-based liquids in closed loops, while other immersion designs use hydrocarbon-based dielectric fluids.
Treating all advanced cooling as a single PFAS market would therefore overstate the connection. The concern is that particular vendors and operators can select fluorinated fluids for demanding applications.
There is also an important distinction between fluoropolymer components and mobile fluorinated fluids. A solid gasket, a process gas, and an immersion coolant have different exposure and release pathways.
The OECD life-cycle review emphasizes this complexity. It calls for better data on production, processing aids, product use, degradation, and environmental emissions.
That life-cycle perspective changes the efficiency calculation. Operators cannot judge a coolant only by electricity or water saved during service.
A complete assessment must include chemical production, transport, leakage, equipment servicing, end-of-life recovery, destruction, and persistent transformation products. Public data for several of those stages remain limited.
The ChemSec PFAS AI Warning Challenges the Efficiency Story
AI infrastructure can become more efficient inside the facility while shifting pollution and liability toward chemical plants and surrounding communities.
Data center operators often present cooling decisions through operational metrics. These include power consumption, water use, equipment density, uptime, and the amount of heat removed.
Those measures matter, but they describe only part of the system. A cooling method with favorable facility metrics can still carry a larger upstream chemical burden.
This is the core tradeoff behind the ChemSec PFAS AI warning. The chemical industry highlights the functional performance of fluorinated materials. Campaigners focus on persistent releases across the material’s entire life.
Chemours offers a clear example of that collision. The company describes its liquid cooling technology as a way to lower operating costs and resource demands.
At the same time, Chemours has faced extensive litigation connected to historical PFAS pollution. In June 2026, authorities announced a $450 million settlement concerning discharges from facilities in North Carolina and West Virginia.
The settlement does not establish that every current Chemours product presents the same risk. It does show why communities and investors question assurances based only on controlled product use.
ChemSec says Chemours remains highly dependent on PFAS-related business. Its assessment identified 57 PFAS associated with the company’s production or use in relevant databases.
The campaign group also says Chemours has expanded refrigerant capacity and is developing products aimed at data center cooling. That strategy places AI growth beside a continuing legal and environmental burden.
Producers reject the idea that all PFAS should receive identical treatment. Their preferred distinction separates certain fluoropolymers from smaller, mobile, or bioaccumulative PFAS.
An industry-backed AI infrastructure paper describes fluoropolymers as chemically inert and non-bioaccumulative. It argues that broad restrictions could weaken reliability, fire safety, cooling efficiency, and semiconductor supply.
That argument has practical force. PFAS describes thousands of substances with different molecular structures, physical forms, applications, and exposure profiles.
A solid fluoropolymer installed inside closed equipment should not be discussed as though it were identical to a soluble processing aid. Risk depends on production methods, emissions, degradation, and disposal.
Yet the category debate can also obscure shared persistence. Stable end products require manufacturing, and that manufacturing may involve other fluorinated substances or produce emissions.
Waste handling adds another problem. A material that stays stable during normal use can become difficult to destroy after equipment reaches the end of its useful life.
Incineration, landfilling, recycling, and specialized destruction technologies produce different outcomes. Reliable public information about the final destination of data center cooling fluids remains sparse.
The AI industry’s layered procurement structure makes this information harder to obtain. Cloud companies buy systems from equipment vendors, which depend on component and chemical suppliers.
Public sustainability reports often disclose electricity, carbon emissions, or water consumption. They rarely provide comparable inventories for PFAS embedded in hardware or used during chip fabrication.
That gap prevents meaningful comparisons. Buyers cannot easily determine whether one server design, semiconductor process, or cooling system creates fewer persistent chemical releases.
Companies also use “eco-friendly” or similar language when discussing reduced energy or water demand. Such descriptions become misleading if they exclude manufacturing pollution and end-of-life management.
The correct question is not whether liquid cooling is good or bad. It is whether a specific configuration reduces total harm compared with available alternatives.
That comparison requires a defined boundary. It should include the semiconductor factory, chemical plant, data center, maintenance chain, waste contractor, and affected water systems.
AI developers also influence this calculation through model and hardware choices. Greater computational demand can require more accelerators even when each new chip performs more work per unit of energy.
Efficiency gains can therefore coexist with greater total resource consumption. A similar rebound can occur with cooling if denser, cheaper operation encourages larger installations.
The current evidence supports a supply-chain warning, not a precise estimate of AI-attributable PFAS pollution. ChemSec identified corporate expansion plans and their stated markets.
It did not provide a complete global production-volume dataset. The organization acknowledges that chemical-sector opacity limits visibility into output, substances, and regional activity.
That limitation matters. Expansion announcements show strategic direction, but they do not reveal exactly how much additional PFAS will enter commerce because of AI.
Companies serve multiple growing markets. Batteries, medical equipment, transportation, defense, renewable-energy systems, and industrial processing can use the same chemical families.
Still, corporate language provides evidence of intent. When producers repeatedly identify AI, semiconductors, and data centers as growth markets, those sectors cannot treat chemical expansion as unrelated activity.
Regulation Faces an Essential-Use Test
Regulators must distinguish genuinely irreplaceable applications from convenient uses without creating permanent exemptions that eliminate pressure to develop alternatives.
Europe offers the most visible test. Authorities have been evaluating a broad restriction proposal covering the PFAS family under the European Union’s chemicals framework.
The proposal reflects a central regulatory concern. Assessing thousands of persistent substances individually can leave governments reacting long after contamination has spread.
A class-based approach can prevent one restricted chemical from being replaced by a closely related substance with similar persistence. Industry groups answer that the class is too diverse for uniform treatment.
The European chemicals overview notes that several PFAS have raised cancer and other health concerns. It also describes active restrictions and the continuing universal proposal.
Semiconductor manufacturers want time-limited or continuing exemptions for uses without qualified replacements. They argue that sudden loss of critical materials could constrain production, safety, and technological competitiveness.
This concern is not merely rhetorical. Chip fabrication uses integrated processes in which material changes must preserve purity, reliability, equipment compatibility, and performance.
The industry has also demonstrated that selected substitutions are possible. Semiconductor companies previously eliminated intentional uses of PFOS and later completed a phaseout of intentional PFOA uses in specified processes.
Those changes required coordination and long development periods. They show both the difficulty of substitution and the danger of declaring current chemistry permanently indispensable.
ChemSec favors universal restrictions with strictly time-limited derogations. A derogation allows a temporary exception for a defined use while an industry develops alternatives or emission controls.
The time limit is crucial. An exemption without milestones can protect existing production more effectively than it encourages safer chemistry.
BASF presents one alternative strategy. The company has said it plans to exit most products formulated with PFAS by 2028, excluding pesticides from that commitment.
Archroma has reduced part of its PFAS portfolio and markets PFAS-free alternatives in selected categories. Neither example proves that advanced semiconductor uses already have direct substitutes.
They do challenge the assumption that every producer must pursue expansion. Corporate strategy, product mix, and exposure to legal risk can produce different decisions.
Daikin, Arkema, Chemours, and other expanding suppliers are placing a different bet. They expect demand and essential-use arguments to preserve significant markets.
That creates policy pressure from two directions. Regulators risk disrupting critical infrastructure if restrictions move faster than substitution. They risk encouraging another generation of contamination if exemptions remain broad.
The strongest regulatory response would separate urgency from indiscriminate prohibition. Governments can demand disclosure, measurable emission reductions, recovery plans, and funded substitution research before every replacement is available.
Production reporting is an immediate need. Regulators and communities require substance-level information, volumes, release pathways, waste destinations, and monitoring results.
Facilities should also disclose which pollution-control technologies capture PFAS and what happens after capture. Moving contamination from wastewater into sludge or filters does not eliminate it.
Procurement rules can extend that pressure downstream. Cloud providers and semiconductor buyers can require suppliers to report chemical content and demonstrate reductions in nonessential applications.
Essentiality should depend on the use, not simply the chemical category or economic value of the final product. An application can support AI without automatically qualifying for an unlimited exemption.
Regulators also need to test claims about regrettable substitution. A replacement that introduces flammability, toxicity, or equipment failure would not represent progress.
That concern supports comparative hazard assessment and staged qualification. It does not justify indefinite reliance without transparent research targets.
The health evidence must be described carefully. PFAS exposure is associated with different outcomes depending on the substance, dose, timing, and route.
The EPA risk summary links exposure to certain PFAS with reproductive effects, developmental effects, immune impacts, cholesterol changes, and some cancers.
Those findings do not mean every fluoropolymer component creates the same direct exposure. They strengthen the case for identifying releases instead of assuming safe containment.
The decisive policy question is therefore operational. Can industry document where PFAS enters, remains, escapes, transforms, and ends its life across the AI supply chain?
Without that evidence, claims of essential use remain incomplete. A material can be technically necessary today while its uncontrolled emissions remain unacceptable.
Three Signals Will Show Whether the Warning Changes AI Infrastructure
The next stage will be decided by disclosure, regulatory conditions, and credible alternatives rather than competing slogans about innovation or prohibition.
The first signal is whether major AI infrastructure buyers publish PFAS procurement requirements. Cloud providers have enough purchasing power to push chemical disclosure through server, cooling, and semiconductor suppliers.
A meaningful policy would identify covered PFAS, demand life-cycle information, and set reduction expectations for nonessential uses. It would also distinguish solid components, process chemicals, gases, and mobile fluids.
Generic supplier codes will not resolve the issue. Buyers need measurable requirements supported by facility data, independent testing, and end-of-life documentation.
Such disclosures would strengthen ChemSec’s argument by showing that AI companies recognize induced chemical demand. Silence would leave campaigners dependent on supplier announcements and incomplete public databases.
The second signal is how regulators define essential semiconductor and cooling uses. The crucial details will be exemption scope, duration, reporting duties, and pollution-control conditions.
A narrow, time-limited exception would preserve current production while maintaining pressure for substitution. A broad exception without deadlines would weaken the claim that restrictions will change industry behavior.
Cooling deserves particular scrutiny because alternative architectures already exist for some workloads. Regulators should ask whether each proposed fluorinated fluid solves a genuinely unique technical problem.
Semiconductor applications require a more granular assessment. A substance used in photolithography can face different substitution barriers from a polymer used in piping.
The third signal is whether suppliers demonstrate qualified alternatives or closed-loop controls at commercial scale. Announcements alone will not settle the debate.
A credible alternative must meet technical requirements and present a better environmental profile. It must also survive manufacturing qualification without transferring risk into another chemical category.
Where substitution remains unavailable, companies should demonstrate capture, monitoring, recovery, and destruction. They should report releases rather than relying on claims that normal use is contained.
Watch Daikin’s planned capacity growth, BASF’s 2028 exit commitment, and the development of non-PFAS cooling options. Together, these paths provide a real-world comparison between expansion, withdrawal, and substitution.
The outcome will affect more than chemical companies. Chip designers, cloud providers, enterprise buyers, and software teams increasingly depend on AI infrastructure whose physical impacts extend beyond electricity.
Enterprise procurement teams should ask vendors what cooling architecture they use and how fluids are recovered. They should also request semiconductor supply-chain disclosures when those become available.
Developers cannot personally redesign a fabrication plant. They can still treat computation as a material resource instead of an abstract service with no upstream footprint.
Product teams can examine whether model size, inference frequency, retention policies, or hardware utilization create avoidable demand. Those choices will not eliminate PFAS, but they affect infrastructure growth.
Investors should compare growth claims with legal liabilities and regulatory exposure. A producer’s position in the AI supply chain can create revenue while increasing remediation, monitoring, and substitution costs.
Communities near chemical plants and fabrication sites need access to emissions data before expansions begin. Monitoring after contamination appears transfers too much risk to residents.
The ChemSec PFAS AI warning does not prove that every AI request releases forever chemicals. It shows that the industry’s expansion is reshaping demand for persistent materials upstream.
The burden now shifts toward companies claiming that those materials are essential. They must explain which uses lack alternatives, how releases are controlled, and when safer options will arrive.
AI’s environmental accounting has expanded from energy and water into chemistry. The next useful step is to ask providers for specific material disclosures and measurable reduction plans.
Will the companies building AI infrastructure publish those answers before new PFAS capacity enters service, or only after regulators and communities force the issue?
