Syensqo technology chief: AI workloads are pushing chips and data centers to physical limits

The next constraint on AI may not be a missing algorithm or an undersized chip order, but the chemistry of the machines themselves. In a Business Lab episode published by MIT Technology Review on September 16, Mike Finelli of Syensqo argues that AI workloads are driving semiconductors and data centers up against hard physical limits, and that the materials they are built from increasingly shape what the technology can do. The episode was produced in partnership with Syensqo; Finelli is the company's chief technology and innovation officer and chief North America officer.

Stacking requirements

According to Finelli, the demands placed on infrastructure materials are accumulating rather than trading off against one another. Chips and data center hardware must now handle higher temperatures, tighter purity requirements, better electrical performance, resistance to aggressive chemicals and plasmas, and long-term stability, all simultaneously.

He frames this with a pyramid image: commodity materials sit at the base, and every added requirement pushes the relevant material toward a narrow apex of high-performance specialty products. His argument is that advanced materials are no longer merely supporting AI progress but increasingly defining its boundaries.

Cooling, voltage and seals

The interview outlines where Syensqo, a specialty materials company, is directing its work: materials suited to high-voltage data center power architectures, advanced sealing materials for semiconductor manufacturing equipment, and thermal-management products that include fluids for direct immersion cooling.

Finelli also points to cross-industry reuse. Materials originally developed for electric vehicles, which already contend with high voltage and high energy density, can carry over to data center applications facing similar demands.

AI finds the molecules

The relationship runs both ways. According to MIT Technology Review, Syensqo is using AI agents to digitally synthesize millions of candidate molecular combinations, then predict both their technical performance and their sustainability profile before sending a shortlist to laboratory testing. Finelli says this lets researchers go "broader, deeper, and faster" and frees scientists to concentrate on genuinely difficult engineering problems rather than manual screening.

The company says 20% of its annual revenue comes from products and applications launched within the past five years, which it presents as evidence of an active innovation pipeline.

Sustainability from the start

Finelli says customers increasingly expect materials to meet technical requirements while lowering environmental impact, and he describes Syensqo's goal as eliminating the trade-off between the two. In practice, that means sustainability criteria enter the research process at the beginning rather than being layered onto an already-finished material. It also means the AI screening step predicts sustainability characteristics alongside technical ones.

A loop that feeds itself

Looking ahead, Finelli anticipates a reinforcing dynamic: AI helps discover better materials, those materials improve AI infrastructure, and improved infrastructure in turn accelerates the next round of materials discovery. He describes the result as an accelerating cycle of materials innovation that expands what future technologies can achieve.

Why it matters

The story reframes the AI capacity debate. Scaling compute is increasingly bottlenecked by physical engineering — power delivery, heat dissipation, purity and long-term reliability — rather than by model design alone. As data centers shift toward high-voltage architectures and immersion cooling, supply chains for specialty fluids, polymers and sealing materials become part of the AI capacity story, and slow qualification or shortages in that layer would translate directly into slower deployments.

At the same time, using AI agents to compress materials discovery from millions of candidates down to a testable shortlist is one of the more concrete industrial applications of the technology, with timelines that matter when infrastructure demand is measured in years, not decades.

One caveat is worth flagging: the source is sponsored content built around an interview with the sponsor's own executive, so the framing naturally emphasizes that company's market position. But the underlying tension it describes — between AI's growth curve and the physical limits of the materials underneath it — is a genuine constraint on how far compute infrastructure can go.