Syensqo is making the case that the next bottleneck in AI infrastructure is not confined to model design, chip architecture, or the construction of more data centers. In company-produced material published by MIT Technology Review, the materials company said advanced polymers, elastomers, specialty fluids and related materials are being pushed harder as AI systems demand more compute, memory, power density and reliability.
The claim is not glamorous, which is probably why it gets less attention than GPUs and billion-dollar fab announcements. But the mechanism is straightforward: faster chips and denser servers create harsher physical conditions. Manufacturing equipment has to tolerate aggressive chemicals, plasma and tight thermal control. Data centers have to move more heat, manage higher voltages and keep connectors, capacitors, drives and other components working under heavier load.
Chipmaking leaves little room for sloppy materials
Syensqo said modern semiconductor manufacturing involves thousands of tightly controlled process steps. Small shifts in temperature or unstable chemistry can introduce defects, reduce yield and raise costs, according to the company.
That puts pressure on materials used inside fabrication tools. Syensqo said chipmakers need higher purity, stronger resistance to chemicals and plasma, and better stability as each chip generation tightens tolerances and operating conditions. The company framed this as incremental engineering rather than reinvention: materials suppliers have to keep improving the supporting materials as semiconductor processes become more demanding.
The same argument extends to AI data centers. Syensqo said rising compute density is changing server and facility design, including cooling systems, power delivery, storage and high-speed data transmission. It pointed to direct liquid cooling for AI servers as one area where know-how from semiconductor fluid circulation and automotive coolant systems can be reused.
Syensqo also said higher-voltage data center designs create materials problems similar to those seen in electric vehicles. That is a company claim, not an independent benchmark, but the comparison is technically plausible at the level Syensqo described: more power in tighter spaces means insulation, cooling and reliability become harder to fake with marketing slides.
Perfluoroelastomers get a manufacturing change
Syensqo highlighted perfluoroelastomers, which are used as seals in semiconductor manufacturing equipment. The company said these materials must withstand high temperatures, plasma and reactive chemicals.
The company said its next generation of perfluoroelastomers is made with a fluorosurfactant-free manufacturing process. Syensqo presented the change as an attempt to improve performance while making production more responsible. It did not provide performance figures, qualification results or customer names in the material.
Syensqo also cautioned that new materials do not enter chipmaking or data center supply chains quickly just because they are new. Qualification can take years, the company said, and manufacturers tend to change materials only when they solve a specific engineering problem or enable a new process.
AI is also being used to find the materials
Syensqo said it is using AI tools, including Microsoft Discovery, to identify and evaluate molecular candidates for next-generation heat-transfer fluids used in semiconductor manufacturing and data centers.
According to the company, these tools help researchers screen candidates against desired properties before committing as much laboratory time. Syensqo said that can reduce the number of physical experiments needed in early discovery, though it said lab testing, scientific review and customer qualification remain necessary before any material is deployed.
The result is a tidy feedback loop: AI infrastructure needs better materials, and materials companies are using AI to search for them. Syensqo’s version still needs the usual proof points, including qualification and field performance. The company’s broader point is harder to dismiss: the AI stack has a physical underside, and it is getting less forgiving.
This story draws on original reporting from MIT Technology Review.
