San Francisco chip startup Volantis has raised $88 million to replace the short-range electrical links inside AI servers with optical connections, claiming its approach would let a single GPU address as many as 220 memory chips — roughly a 27-fold increase over the eight HBM stacks that top-end accelerators typically carry today.

The round backs a direct attack on the memory bottleneck constraining AI accelerators from Nvidia and Advanced Micro Devices. Compute cores can only run as fast as they can fetch data from memory, where the model and its working set reside. The industry's answer so far has been high-bandwidth memory stacked immediately next to the GPU die, but electrical connections limit reach so severely that even flagship designs pair roughly eight HBM chips per GPU.

Volantis says optical links built on vertical-cavity surface-emitting lasers (VCSELs) can move data as light, freeing memory from having to sit tightly packed around the compute chip. In the company's design, that longer reach lets one GPU talk to far more memory chips, easing the bandwidth bottleneck without leaning exclusively on the most complex and expensive advanced packaging.

The manufacturing argument is central to the pitch. VCSELs are already produced at scale for smartphone features such as Face ID, which Volantis argues makes the technology comparatively manufacturable rather than a laboratory curiosity. The company is aiming to deliver a chip next year.

The economics of memory-bound workloads

The 220-chip claim targets a major cost driver in AI server builds. Many AI workloads are memory-bound, meaning operators add GPUs partly to secure enough memory capacity — not just compute — and then pay additional time and energy shuffling data between chips.

If optical links genuinely allow memory to sit farther from the processor without sacrificing bandwidth, some deployments could need fewer GPUs for the same memory-heavy job. That would shift how AI infrastructure budgets split among accelerators, premium HBM packaging, and the networking that stitches GPU clusters together.

The disruption would not be immediate. Volantis' technology would not dethrone Nvidia or AMD overnight, given their entrenched positions in accelerators and the CUDA-adjacent software ecosystems around them. But if the technology proves out at scale, it could reshape where pricing power sits across the AI hardware stack — moving value away from tight co-packaging of HBM and toward optical interconnect suppliers and disaggregated memory architectures.

What to watch

The near-term test is execution: Volantis has committed to delivering a chip next year, and the burden of proof sits with its claim that VCSEL-based links can sustain bandwidth at distance inside real server environments rather than on test benches. Success would give system designers a new lever against HBM supply constraints and packaging costs; failure would leave the eight-stacks-per-GPU status quo — and its suppliers — firmly in place.