A team at MIT's Research Laboratory of Electronics has demonstrated on-chip wavelength multiplexing at 100 terahertz aggregate bandwidth — approximately 1000 times what current electrical interconnects can achieve at the same energy budget.
What Wavelength Multiplexing Actually Is
Standard chips send data over copper traces one bit at a time per wire. Optical interconnects can send many independent streams simultaneously by using different wavelengths of light on the same waveguide — the same principle used in long-haul fiber optic cables, but now shrunk to chip scale.
The MIT team demonstrated 64 independent wavelength channels packed into a single silicon nitride waveguide, each running at roughly 1.6 THz, for a combined aggregate bandwidth of 100 THz on a chip smaller than a thumbnail.
Why This Number Matters
The primary bottleneck in large AI clusters today is not compute — it is moving data between chips and memory fast enough to keep the compute units fed. A 100 THz on-chip interconnect changes that calculus entirely.
At that bandwidth, memory latency becomes nearly irrelevant for matrix operations. A GPU cluster that today requires complex software-level tiling to fit model weights into memory could be replaced by a smaller, simpler system with photonic interconnects that makes the memory wall invisible.
From Lab to Product
The MIT result is a research demonstration, not a product. The waveguide structures require fabrication using standard CMOS-compatible silicon nitride processes, which is a genuine advantage for commercialization.
Several of the researchers involved have previously spun out companies from related MIT photonics work. Industry contacts suggest active conversations with at least two major semiconductor foundries about licensing the fabrication process. A product timeline of 3–5 years from publication is considered realistic by photonics analysts.