ChemSec PFAS Data Center Warning Exposes AI Cooling’s Toxic Tradeoff

ChemSec says its review of 10 major producers found widespread PFAS expansion, with AI infrastructure helping drive demand despite mounting health and regulatory concerns. The ChemSec PFAS data center warning turns an obscure supply-chain issue into a direct challenge for the companies building AI capacity.

The conflict reaches beyond chemicals factories. PFAS, or per- and polyfluoroalkyl substances, support semiconductor fabrication and some advanced cooling systems because they resist heat, water, and corrosion. Those same qualities allow many PFAS to persist in the environment.

That creates an uncomfortable tradeoff. AI operators need denser computing systems, while chipmakers and cooling suppliers need materials that tolerate increasingly demanding conditions. ChemSec argues that expanding persistent chemical production is the wrong answer, even when suppliers frame those materials as essential to efficient infrastructure.

The immediate opponents are not two technology companies. They are the efficiency promise surrounding next-generation AI infrastructure and the pollution risks embedded in the materials that make it possible.

ChemSec’s PFAS Data Center Warning Targets a Growing Supply Chain

ChemSec’s central claim is that AI demand is giving PFAS producers a new reason to expand when regulators are considering tighter controls.

The Swedish nonprofit published its updated analysis of major PFAS manufacturers on September 13, 2026. The review examined corporate disclosures, public documents, product portfolios, and expansion announcements.

ChemSec concluded that most leading producers were increasing capacity or pursuing growth tied to data centers, semiconductor production, and lithium-ion batteries. Its list covered companies operating across North America, Europe, and Asia.

The following day, an established-news production survey detailed several projects and corporate statements behind the warning. The report described AI as a leading source of new demand, although it is not the only one.

ChemSec’s own top-producer review presents a more complicated market than the headline suggests. Semiconductor manufacturing, batteries, cables, refrigerants, and industrial coatings all contribute to growth.

That distinction matters. It would be misleading to treat every planned PFAS production line as a direct response to generative AI. The stronger conclusion is that AI infrastructure adds momentum to an already expanding fluorochemical market.

ChemSec identified AGC, Arkema, Chemours, Daikin, Orbia Fluor & Energy Materials, Solstice, and Syensqo among companies pursuing relevant growth. It also discussed producers in China, India, Russia, and other markets.

Several manufacturers have explicitly connected their investments to semiconductor capacity, advanced data centers, or both. Those statements give campaigners evidence that AI is influencing capital spending far upstream from cloud platforms.

Arkema opened a new refrigerant production unit in Kentucky after describing data center cooling as one source of rising demand. Its official Kentucky expansion announcement said the facility would increase production of Forane 1233zd.

Chemours has also promoted specialized heat-transfer fluids for advanced data center cooling. Its liquid-cooling venture includes a manufacturing agreement with India-based Navin Fluorine.

Daikin has connected fluorochemical growth to semiconductor demand and has promoted a dedicated data center business strategy. Other suppliers see related opportunities in battery binders, coatings, and refrigerants.

ChemSec treats these projects as parts of one expansion cycle. The organization warns that partial restrictions will struggle if demand keeps creating new production incentives.

However, ChemSec is an advocacy group seeking broad PFAS restrictions. Its analysis should be read as a documented campaign intervention, not an independent production census.

ChemSec also acknowledges an important data gap. Public information does not provide a complete picture of global production volumes, particularly across parts of Asia.

The organization therefore ranks producers using the number of substances associated with them, public company information, and its own databases. That method can reveal direction, but it cannot calculate AI’s precise share of worldwide PFAS output.

The event still changes the environmental debate around AI. It connects data center growth to manufacturing decisions that happen long before a server reaches a cloud facility.

Until now, public scrutiny has concentrated on electricity, water, construction, and backup generators. The ChemSec PFAS data center warning adds chemical persistence and end-of-life management to that list.

AI Chips Depend on Chemicals That Regulators Increasingly Question

AI’s PFAS exposure begins in semiconductor factories, where demanding production steps require materials that survive extreme chemical and thermal conditions.

Modern chips contain structures measured at nanometer scales. Their production involves repeated coating, cleaning, etching, deposition, rinsing, and patterning processes.

Fluorinated materials can serve as processing aids, membranes, tubing, seals, filters, coolants, and components in specialized equipment. Their stability makes them valuable where contamination or material failure can destroy expensive wafers.

This is not a minor operational detail. A data center operator can change a cooling design, but it cannot easily remove chemicals embedded throughout a global semiconductor supply chain.

Some industry estimates describe PFAS as present across hundreds of chipmaking steps. The exact count varies by process and by how the chemical class is defined.

That variation is one reason broad claims need caution. PFAS includes thousands of substances with different uses, properties, exposure pathways, and toxicity evidence.

The category is not one interchangeable chemical. A fluoropolymer used in equipment does not behave exactly like a mobile compound found in contaminated drinking water.

Manufacturers use this diversity to oppose treating all PFAS identically. Environmental groups respond that extreme persistence creates a shared problem, even when individual toxicology profiles differ.

Both points matter. Regulation that ignores chemical differences can disrupt essential production without securing the fastest pollution reductions. Regulation that examines each compound in isolation can allow persistent substitutes to replace restricted substances indefinitely.

AI intensifies this policy problem because demand is arriving faster than substitution efforts. More accelerators require more leading-edge fabrication, advanced packaging, memory, networking equipment, and supporting infrastructure.

That demand does not stop with the most visible graphics processors. It reaches foundries, packaging plants, chemical suppliers, component makers, and waste contractors.

The result is a supply chain with weak public visibility. Major cloud providers publish carbon and water goals, yet customers rarely see complete inventories of process chemicals used to manufacture their computing hardware.

Chipmakers disclose some hazardous-material targets, but reporting methods differ. Suppliers can also protect detailed formulations as confidential business information.

This opacity leaves enterprise buyers in a difficult position. They can purchase lower-carbon electricity or choose a more efficient region, but they cannot easily compare the chemical footprint of two AI services.

The problem becomes larger when hardware refresh cycles accelerate. Older servers, cooling equipment, cables, and components eventually enter reuse, recycling, landfill, or incineration systems.

PFAS-containing materials are difficult to manage because conventional disposal does not necessarily eliminate the underlying compounds. Treatment can move contamination between air, water, soil, and waste streams.

That lifecycle is why the ChemSec PFAS data center warning matters to technology buyers. The environmental footprint of an AI workload includes more than power consumed during inference.

It also includes the materials used to fabricate chips, maintain equipment, control heat, and retire hardware. Those impacts remain poorly reflected in cloud sustainability dashboards.

The pressure now falls on chipmakers and major cloud customers. They must either demand clearer chemical reporting or accept that environmental claims cover only part of their infrastructure.

Efficient AI Cooling Can Shift the Environmental Burden

The central tradeoff is that a cooling system can use less water and electricity while introducing persistent chemical risks elsewhere.

AI accelerators concentrate substantial computing capacity inside each rack. That density creates heat that traditional air cooling cannot always remove efficiently.

Operators increasingly use direct-to-chip liquid cooling, which circulates fluid through cold plates attached to processors. Another approach immerses hardware in a nonconductive fluid.

Two-phase immersion cooling uses a fluid with a low boiling point. Heat from the electronics turns the liquid into vapor, which rises, condenses, and returns to the bath.

That cycle can move heat efficiently because boiling absorbs substantial energy. It can also reduce reliance on evaporative cooling systems that consume water.

Chemours says its Opteon 2P50 developmental fluid can support two-phase immersion designs with lower cooling energy and minimal water consumption. Those figures remain company claims tied to particular system assumptions.

The benefits explain why the technology attracts attention. Data center developers face opposition in places where new facilities compete with residents, farms, or industry for water.

A cooling system that nearly eliminates operational water use can look like a straightforward environmental improvement. The chemical used to obtain that benefit complicates the picture.

Environmental organizations identify Opteon 2P50 as PFAS under commonly used broad definitions. They argue that leaks, maintenance losses, manufacturing releases, breakdown products, and disposal deserve more scrutiny.

Chemours says closed-loop designs limit emissions and that different PFAS should not be treated as if they share one risk profile. That response highlights the core dispute.

A closed loop reduces routine releases, but it does not guarantee zero loss. Connections can leak, equipment needs servicing, and fluid must eventually be recovered or replaced.

The production stage also sits outside the data center’s fence line. Even if a facility contains its coolant successfully, manufacturing can create waste or emissions near chemical plants.

End-of-life management creates another gap. Operators need documented recovery procedures, verified destruction or reclamation routes, and clear responsibility when equipment changes hands.

Public reporting rarely answers those questions. Data center permit applications often emphasize electricity, land, noise, and water without identifying every cooling fluid or refrigerant.

This is where “water efficient” can become an incomplete description. It measures one environmental input while leaving persistence, toxicity, and disposal outside the comparison.

The issue does not mean every liquid-cooling system uses PFAS. Water-based direct-to-chip systems, dielectric oils, single-phase immersion fluids, and other technical paths have different material profiles.

Nor does it mean two-phase cooling inevitably contaminates nearby drinking water. Public evidence connecting operating data centers to measured PFAS releases remains limited.

A 2025 investigation found that testing around facilities was scarce and chemical reporting requirements were inconsistent. That absence of data supports concern, but it does not prove widespread contamination.

The most defensible conclusion is narrower. Rapid deployment is moving ahead without a mature public accounting system for fluorinated cooling chemicals.

That leaves operators unable to prove their strongest sustainability claims. It also leaves communities unable to determine which risks are present before facilities begin operating.

The efficiency promise therefore depends on transparency. A lower-water design is only a clear improvement when its chemical inventory, leak controls, recovery plan, and lifecycle impact withstand examination.

The Health Evidence Is Serious, but the Data Center Link Has Gaps

PFAS health concerns are well established for some compounds, while exposure from specific AI data center systems remains insufficiently measured.

The carbon-fluorine bond gives many PFAS their unusual resistance to heat, oil, water, and chemical attack. It also allows many compounds to break down extremely slowly.

The US Environmental Protection Agency says research associates exposure to certain PFAS with reproductive effects, developmental changes, immune impacts, higher cholesterol, hormonal interference, and some cancers.

Its health risk summary also stresses uncertainty. Thousands of PFAS exist, but researchers have detailed evidence for only a limited subset.

That distinction prevents a simple conclusion that every fluorinated data center fluid carries identical risks. Toxicity, mobility, persistence, and exposure routes can differ substantially.

Campaigners focus on persistence because uncertainty does not make emissions reversible. Once highly persistent compounds spread through water or soil, remediation can become technically difficult and expensive.

Manufacturers emphasize chemical specificity. They argue that regulation should evaluate actual hazards and exposure rather than assign the same profile to an entire class.

The policy challenge is deciding who carries the burden of uncertainty. Communities often want safety demonstrated before large-scale production begins. Suppliers generally seek approval based on available substance-specific evidence.

In September 2025, the EPA announced that it would prioritize reviews of new chemicals intended for data center projects or covered components. The agency presented the policy as support for American manufacturing and AI infrastructure.

Its priority review policy did not eliminate statutory chemical reviews. It changed which submissions could move to the front of the queue.

Environmental organizations fear that priority can become pressure for rapid approval. Industry sees long review backlogs as an obstacle to deploying more efficient technology.

That tension surfaced around Chemours’ Opteon 2P50. A coalition of 17 environmental groups asked the EPA to reject or more closely examine the application.

The groups challenged the available toxicology and climate analysis. Chemours responded that critics made overly broad generalizations and overlooked the closed-loop nature of the proposed systems.

No article can resolve that scientific dispute from public statements alone. The decision will depend on exposure assumptions, confidential data, operating controls, and the EPA’s final assessment.

Europe is taking a broader approach. Authorities have been evaluating a proposal covering a wide range of PFAS uses, with possible time-limited derogations for applications lacking alternatives.

The European Chemicals Agency’s restriction process reflects the same balance. Regulators are weighing persistence against economic and technical dependence.

A broad restriction would pressure semiconductor and cooling suppliers to document essential uses. It would also increase demand for substitutes that meet strict reliability requirements.

Exemptions, however, can persist when industries argue that alternatives are unavailable. AI demand may strengthen those arguments by turning data centers into strategic infrastructure.

The largest immediate weakness is measurement. Regulators and communities need facility-specific data on chemical inventories, releases, worker exposure, waste, and environmental concentrations.

Without those measurements, supporters can point to closed-loop engineering while critics point to unresolved lifecycle risks. Neither side can fully describe real-world performance.

The ChemSec PFAS data center warning should therefore be treated as an early supply-chain alarm. It is not proof that every AI facility is releasing dangerous concentrations.

Its strongest evidence concerns business direction. Chemical companies see growing demand, are adding capacity, and increasingly connect fluorinated products to chips and data centers.

Its largest uncertainty concerns exposure. Public data does not yet reveal how much PFAS AI infrastructure releases, where those releases occur, or which communities bear them.

PFAS Alternatives Will Test the Industry’s Sustainability Claims

The clearest test is whether AI suppliers invest in substitution and disclosure before regulators force them to act.

ChemSec contrasts expanding producers with companies moving away from parts of the PFAS market. BASF has said it plans to exit most products formulated with PFAS by 2028.

Archroma has promoted PFAS-free alternatives in selected product categories. 3M previously announced an exit from PFAS manufacturing, although transition details and legacy liabilities remain significant.

These examples do not prove that every semiconductor or cooling use has an immediate substitute. They do show that withdrawal is commercially possible in some applications.

Substitution is rarely a simple chemical swap. A replacement must meet performance, safety, compatibility, availability, cost, and manufacturing-quality requirements.

In semiconductor plants, even small material changes can affect contamination, yield, or equipment life. Qualification may take years because failures can disrupt high-value production.

Data center cooling has more design flexibility, but reliability standards remain strict. Fluid conductivity, boiling behavior, corrosion, flammability, climate effects, and service procedures all matter.

A substitute that removes fluorinated chemistry but increases energy use can create another environmental tradeoff. One that degrades quickly might produce unknown byproducts or require frequent replacement.

The answer is not to abandon comparison. It is to use a lifecycle standard that measures multiple impacts instead of optimizing one headline number.

Cloud providers can start by requesting complete chemical inventories from cooling vendors. They can also require suppliers to disclose PFAS definitions, quantities, leak assumptions, and end-of-life plans.

Hardware buyers can ask chipmakers how they track fluorinated materials across fabrication and packaging. Those questions should extend to contracted factories and chemical suppliers.

Investors have a related role. Expansion announcements often describe AI as a growth opportunity, yet financial disclosures may say less about future regulation, cleanup, or litigation.

PFAS liabilities have already produced major settlements across the chemical industry. Historical cases do not establish the cost of newer compounds, but they show why persistence becomes financially material.

Data center developers also need community-facing disclosure. Residents should not have to identify cooling fluids through technical attachments after a permit receives approval.

A credible application would name relevant substances, expected volumes, storage controls, leak detection, emergency procedures, waste routes, and environmental monitoring plans.

It would also distinguish direct facility use from upstream semiconductor manufacturing. Combining those pathways under one label can obscure where intervention is most effective.

For cloud customers, the issue belongs in procurement rather than marketing. Enterprise sustainability teams increasingly collect data about energy sources and operational emissions.

Chemical dependency should join that assessment when providers deploy immersion systems or commission specialized hardware at scale. Buyers do not need proprietary formulas to request meaningful risk information.

The industry also needs common definitions. US regulators, European agencies, campaign groups, and manufacturers do not always classify the same fluorinated compounds identically.

A supplier can therefore call a product environmentally preferable under one metric while critics classify it as PFAS under another framework. Both statements can circulate without resolving the underlying risk.

Transparent chemical identities and lifecycle data would narrow that gap. They would also help engineers compare technologies without relying entirely on vendor claims.

The central question is not whether AI can operate without every fluorinated material tomorrow. It is whether rapid growth will delay substitution and lock in another generation of persistent chemistry.

ChemSec believes expansion is doing exactly that. Manufacturers argue that their products enable essential technologies while newer designs reduce water, energy, or climate impacts.

The next stage must move beyond competing descriptions. Performance data, emissions measurements, and enforceable recovery requirements will determine which environmental claims survive.

Three Signals Will Show Whether the PFAS Warning Changes AI Infrastructure

Regulatory decisions, supplier disclosures, and real-world monitoring will reveal whether the warning produces reform or only another sustainability debate.

The first signal is the EPA’s handling of new data center chemicals, particularly applications for advanced cooling fluids. A detailed public risk determination would clarify how the agency separates persistence, toxicity, exposure, and infrastructure benefits.

Strict conditions on manufacturing, leaks, recovery, or disposal would strengthen ChemSec’s case that current safeguards are incomplete. Rapid approval with limited disclosure would weaken public confidence, even if the agency concludes risks are manageable.

The second signal is movement in Europe’s broad PFAS restriction process. Semiconductor and data center exemptions will show whether policymakers consider these uses essential, transitional, or replaceable.

Short, conditional derogations would pressure suppliers to develop alternatives. Broad or lengthy exemptions would protect existing materials and support further production investment.

The third signal is facility-level disclosure. Operators should begin reporting which fluorinated cooling chemicals they use, how much they hold, and what happens during leaks or decommissioning.

Actual environmental testing matters even more. Measurements near chemical plants and operating data centers would replace assumptions with evidence about exposure pathways.

These signals should be read together. A chemical can pass a regulatory review yet still require careful operational monitoring. A facility can avoid local releases while depending on polluting upstream production.

The AI industry has spent years improving the visibility of power use and carbon emissions. Water consumption is now receiving similar attention in many proposed data center markets.

PFAS presents a harder accounting problem because impacts stretch across manufacturing, facility operation, maintenance, and disposal. Responsibility is divided among suppliers, operators, waste companies, and regulators.

That division can become an excuse for inaction. It can also become a procurement opportunity if large cloud buyers use their leverage across the supply chain.

The ChemSec PFAS data center warning does not show that AI alone created the forever-chemical market. It shows that AI demand is reinforcing production at a politically sensitive moment.

Developers, enterprise buyers, and AI users should now ask a sharper question: does an infrastructure efficiency claim include its chemical costs, or merely move them elsewhere?

The answer will depend on evidence that is still missing. Watch regulatory conditions, substitution timelines, chemical inventories, leak reporting, and environmental measurements.

If companies disclose those details voluntarily, the industry can compare cooling and chipmaking options before new capacity becomes locked in. If they do not, regulators and communities will set the terms instead.