As Nvidia and Coherent signal a shift toward all-optical networking to solve the AI power crisis, Redmond-based Lumotive has announced a milestone that could redefine data center scalability. In an interview with DIGITIMES Asia, Gleb Akselrod, Co-founder and CTO of Lumotive, detailed the success of the world's first programmable 2D optical beamforming chip.
Solving the GPU networking crisis
The breakthrough comes at a critical juncture for AI infrastructure. According to Akselrod, "the biggest bottlenecks in data centers today are the networking. It's actually more than computing." While optics have long been a part of the data center, Akselrod says that the technology is now "coming closer and closer to the GPU."
Most data centers currently rely on "packet switching," which requires converting light signals back into electronics to be routed. Lumotive's approach is to keep data entirely in the optical domain. Akselrod says that "the vision of the future is that all that data stays in the optical domain... You don't convert it from optics to electronics back to optics, but you keep it all optical. But for that, you need to be able to route optical signals in a switch, from any input GPU to any output GPU. And for that, you need optical beam steering."
Therefore, the solid-state approach offers a massive advantage in network agility. The 2D chip can switch in microseconds, whereas MEMS takes milliseconds. This means Lumotive's approach is hundreds of times faster, enabling "faster reconfiguration of the network for dynamic AI workloads." This speed is critical for hyperscalers that need to re-route massive data flows instantly as LLM training demands fluctuate.
The scaling advantage
While current Optical Circuit Switches (OCS) typically max out at 320 ports due to the physical limitations of Micro-Electro-Mechanical Systems (MEMS), Lumotive's solid-state approach allows for unprecedented density. The company's architecture is reportedly scalable to 10,000 x 10,000 ports, a feat Akselrod attributes to a fundamental shift in physics.
The "insertion loss, meaning the loss in these switches, doesn't depend on port count," Akselrod said. "The more ports we have, we lose the same amount of light through the port. Whereas with some technologies, the bigger the number gets, the higher the loss becomes. So they cannot scale to large port counts."
The hub of the optical age
With the company recently opening a dedicated office in Taiwan, Lumotive is tightening its ties to the core of the global semiconductor market. As the company moves toward its product launch in the fourth quarter of 2026, the focus remains on delivering a software-defined platform for light that matches the reliability of the chips it connects.
By leveraging CMOS-compatible manufacturing, Lumotive is positioning itself as a more reliable, lower-cost alternative to legacy optics. "Because our technology is solid state, we can make a switch that's much smaller, much higher port density, and as a result, much lower cost per port," Akselrod concluded.
Article edited by Jack Wu
