The secret ingredient keeping AI running: advanced materials science
Chips and software grab the headlines, but a quieter race is under way in the labs developing the physical materials that let AI hardware exist at all.
Key points
- Specialty materials, including perfluoroelastomers (heat-resistant seals used inside chip-making machines) and heat-transfer fluids, set hard physical limits on how powerful AI chips can become.
- Belgian chemicals company Syensqo has developed a next-generation perfluoroelastomer made without fluorosurfactants, a class of long-lasting chemical compounds that are difficult to break down in the environment.
- Syensqo is using Microsoft Discovery, an AI-assisted research platform, to identify candidate molecules for new cooling fluids faster than traditional lab methods allow.
- Cooling techniques borrowed from electric-vehicle engineering are being adapted for high-density AI server racks, which generate intense heat in a small space.
- New materials must pass years of qualification testing before chip makers will accept them, so performance and reliability still come before everything else.
Every time a new AI model launches, the conversation turns to algorithms, data, and the billions spent on new chips. Rarely does it turn to the seals inside the machines that print those chips, or the fluids that stop them from overheating. That is a gap worth closing.
Manufacturing a modern semiconductor chip requires thousands of tightly controlled steps. Tiny temperature swings or chemically unstable materials create microscopic defects, cutting the number of working chips per wafer and pushing costs up. As each new chip generation packs in more transistors, those tolerances tighten further.
Perfluoroelastomers are one example of where materials science sits at the sharp end of this problem. These dense, flexible seals hold the reactive gases and plasma inside chip-fabrication equipment. They must survive conditions that would destroy most industrial materials. Syensqo, a specialty chemicals company spun out of Solvay in 2023, now makes a version of these seals using a manufacturing process that avoids fluorosurfactants. The company says the new material performs better and is produced more responsibly, without forcing chip makers to choose between the two.
Does this affect ordinary AI users?
Yes, indirectly but concretely. When materials fail or fall short, chip yields drop, hardware costs rise, and the price of running AI services climbs with them. Better materials mean more reliable chips, more affordable cloud computing, and AI tools that reach more people.
The pressure is not only on the fabrication floor. Data centres running large AI workloads now pack computing hardware so densely that conventional air cooling struggles to keep up. Higher-voltage power systems are becoming standard, and that brings a fresh set of materials challenges around electrical insulation and heat management.
Syensqo says it is drawing on work it has already done for electric vehicles to address this. Cooling fluid systems designed for EV battery packs share enough engineering DNA with liquid-cooled AI server racks that lessons transfer across, cutting development time.
Then there is the question of how new materials get discovered in the first place. Traditionally, chemists would form a hypothesis, synthesise a compound, test it, and iterate, a cycle that takes months per candidate. Syensqo now uses Microsoft Discovery, an AI-assisted research platform, to screen molecular candidates for new heat-transfer fluids early in the process. The AI does not replace the chemists. It narrows the field so they spend laboratory time on the most promising options rather than on long-shots.
None of this moves quickly. A material identified in a lab today could take several years to qualify for use in commercial chip-making equipment. Manufacturers change materials only when a new one genuinely solves a problem.
The piece originally appeared as sponsored content published by MIT Technology Review. That context matters: it is a company making its case, not independent analysis. The engineering principles it describes, however, reflect real constraints that shape what AI hardware can and cannot do.
