Volantis Uses Photonic Interconnects to Address AI Hardware Memory Bandwidth Limitations

Volantis, a startup funded by Sam Altman, is developing an AI accelerator chip that leverages optical interconnects instead of traditional electrical wiring to significantly expand memory bandwidth and capacity. The company's A-1 chip design aims to achieve up to 240 TB/s memory bandwidth with 10 TB capacity, reportedly sufficient to serve massive language models at high throughput rates. By focusing development on the optical interposer while licensing other components from established IP providers, Volantis seeks to overcome the physical constraints that currently limit memory scaling in modern GPUs.

The memory bandwidth bottleneck in current AI hardware stems from physical constraints on how much high-speed memory can connect to processor cores. Traditional electrical connections require memory modules to sit extremely close to compute dies—within millimeters—to maintain signal quality and minimize power waste. This proximity limitation, defined by the chip's perimeter or "shoreline," caps practical memory capacity and speed. Volantis's optical approach overcomes this by using light-based interconnects through silicon waveguides, allowing memory to be positioned several centimeters away while maintaining vastly higher data transfer rates.

The company's strategy differs from prior photonic interposer research by integrating micro-VCSELs directly into the interposer rather than relying on external light sources, and by focusing on building complete inference chips rather than licensing technology to other manufacturers. By outsourcing compute architecture and memory specifications to established partners, Volantis concentrates engineering resources on the novel optical interconnect layer—a focused approach intended to reduce development risk while accelerating time to market.