Volantis Unveils Photonics Plan to Break AI Chip Memory Wall

What happened

Volantis, an AI hardware startup backed by OpenAI's Sam Altman, has set out plans to use photonics — light-based interconnects — rather than traditional copper wiring to link memory to compute inside AI accelerators. The approach is designed to address the so-called "memory wall," the bandwidth and latency bottleneck that occurs when data must move between processing units and memory.

According to reporting from The Register, Volantis's proposition is to embed optical connections directly into the architecture of AI chips, replacing the electrical pathways that currently constrain how quickly accelerators can read and write data. The company frames this as a structural rethink of how memory and compute are wired together, rather than an incremental tweak to existing chip designs.

Why it matters

The memory wall has become one of the central constraints on scaling AI workloads: as models grow, the time and energy spent shuttling data between memory and processors increasingly limits performance gains, even as raw compute power continues to climb. A viable photonics-based alternative to copper interconnects would mark a meaningful shift in how AI infrastructure is engineered, potentially easing one of the bottlenecks that shapes training speed, inference latency and energy consumption at scale.

For organisations building or buying AI infrastructure, this is a reminder that the next wave of AI capability gains may come as much from hardware architecture as from model design. Decisions about chip and interconnect technology, made years before a model ships, increasingly determine what AI systems can deliver — and at what cost and speed — once they reach end users.

The Renascence take

It's tempting to file this under "deep tech, not our problem." But the physical plumbing of AI — how fast data moves inside a chip — is exactly what determines whether the AI experiences people actually touch feel instant or sluggish, cheap or expensive to run.

Most experience and transformation leaders treat AI performance as a model problem, obsessing over prompts, fine-tuning and guardrails while ignoring the infrastructure underneath. But latency, cost-per-query and the feasibility of real-time AI features are all downstream of decisions like this one. Operators planning AI-driven service or support experiences should start asking their vendors not just "which model" but "on what hardware, and with what constraints" — because the memory wall, not the model, is often the real limit on what feels fast, affordable and scalable to customers.