MIT Pairs Photonics With Electronics to Break AI Data Centers' Energy Barrier

A team of MIT researchers has announced progress in the FUTUR-IC program, aimed at solving one of the most pressing problems facing the artificial intelligence industry: the energy cost of moving data between processors, accelerators and memory. Rather than building a fully optical processor, the researchers developed hybrid couplers in which electronics handle computation and photonics handle communication.

The key achievement described by MIT is a set of optical couplers, the first practical solutions for transmitting light between components built in conventional silicon semiconductor foundries. The team, led by Anu Agarwal, developed a GRIN coupler that operates across a wider range of wavelengths, and an evanescent coupler that is simpler and cheaper to manufacture.

Our solutions will enable a jump from hundreds of terabits per second to more than 1 petabit per second - Anu Agarwal, director of the MIT FUTUR-IC program

A Problem Outpacing Compute Growth

Modern language models and AI systems no longer suffer only from a lack of raw chip compute, but from the cost of moving data between chips. The larger the model and the more accelerators running in parallel, the more energy is consumed by inter-chip communication rather than by the computation itself. That bottleneck limits further scaling of the training clusters used by the largest AI labs.

Transmitting data with light is physically far more energy-efficient than traditional electrical connections, since photons don't generate heat losses the way current flowing through metal traces does. The catch is that integrating photonics with existing chip manufacturing infrastructure has so far been costly and hard to scale within existing semiconductor foundries.

How It Works in Practice

The MIT team's approach doesn't replace electronics with photonics wholesale, but combines both technologies within a single chip package. Silicon electronics still handle computation, while communication between chips, memory and accelerators runs on light carried through the new couplers. That split lets manufacturers use existing electronics foundries and chip packaging infrastructure instead of building new photonics production lines from scratch.

One of the couplers, the evanescent design, made the cover of Advanced Engineering Materials, while the GRIN coupler was described in the March 2026 issue of Journal of Physics: Photonics. Juejun Hu and Lionel Kimerling of MIT are also working on the project, and Hu's team previously published related results in Laser & Photonics Reviews.

What It Means for the AI Industry

For companies building infrastructure to train large models, from cloud providers to accelerator makers, every percentage point of energy saved in chip-to-chip communication translates into real money and fewer power constraints in data centers. The forecast that data centers could account for up to 10 percent of global electricity demand by 2030 shows the scale of the problem FUTUR-IC is trying to solve before it becomes an insurmountable barrier.

The program is aimed at collaboration with chip packaging suppliers, materials manufacturers and the broader data center supply chain, not just individual research labs. That means if the technology proves out, it could reach commercial production faster than typical academic breakthroughs, since it was designed from the start for compatibility with existing foundries.

For Polish companies and research institutions working on computing infrastructure, the topic matters indirectly: energy and cooling costs for data centers directly affect the price of cloud and compute services used by domestic teams training their own models, including projects such as Bielik. More energy-efficient chip-to-chip communication means, in the long run, lower costs of access to compute for smaller players.

Sources: MIT News (news.mit.edu), Chip.pl (chip.pl), Optics.org (optics.org)