Chip startup Volantis has raised $88 million to replace short-range electrical connections inside AI servers with laser links, and it claims a single GPU in its design could connect to as many as 220 memory chips.

The San Francisco-based company wants to swap copper interconnects for optical links built on vertical-cavity surface-emitting lasers (VCSELs) — the same class of component that smartphone makers already produce at scale for features like Face ID. Volantis is aiming to deliver a chip next year.

Why is memory the bottleneck?

AI accelerators from Nvidia and Advanced Micro Devices often hit a simple wall: the computing cores can only work as fast as they can fetch data from memory, where the model and its "working set" sit. The industry's fix so far has been high-bandwidth memory (HBM) stacked right next to the GPU.

The constraint is reach. Today's electrical connections are so short that even top-end designs typically pair only about eight HBM chips per GPU. That forces memory to sit tightly around the compute chip and pushes designers toward the most complex, expensive packaging available.

What does the laser approach change?

Volantis says optical links using VCSELs move data with light, so memory no longer has to be packed tightly around the compute chip. In its design, that longer reach lets one GPU talk to far more memory chips — up to 220 — easing the bottleneck without relying solely on advanced packaging.

The company also argues VCSELs are comparatively manufacturable because the industry already produces them at scale for smartphone features such as Face ID. That existing supply base underpins its target of shipping a chip next year.

What could this mean for AI hardware economics?

Volantis' 220-chip pitch takes aim at a major cost driver in AI servers. More memory per GPU can change the economics of building and running AI systems:

• Many workloads are memory-bound, so operators add GPUs partly to get enough memory, not just more compute.
• They then pay extra time and energy to move data between chips.
• If optical links let memory sit farther away without sacrificing bandwidth, some deployments could need fewer GPUs for the same memory-heavy job.

That shift would change how budgets get split between accelerators, premium HBM packaging, and the networking that stitches GPUs together.

The technology would not dethrone Nvidia or AMD overnight. But if it proves out at scale, it could reshape where pricing power sits across the AI hardware stack — moving value away from tightly coupled HBM packaging and toward whoever supplies the optical interconnect layer.