Volantis Raises $88 Million to Replace AI Server Wiring With iPhone Face ID Lasers — 220 Memory Stacks per GPU

San Francisco-based semiconductor startup Volantis has raised $88 million to attack one of AI hardware's most stubborn constraints: the short electrical runway between a processor and its memory. The Series A was co-led by Lachy Groom and Abstract Ventures, with participation from investors including John Doerr and Naveen Rao, according to Reuters.

The company's pitch rests on a component most people already own. Volantis wants to use vertical-cavity surface-emitting lasers, or VCSELs — the tiny semiconductor lasers that make iPhone Face ID work — to carry data between AI compute chips and memory over light rather than electricity. VCSELs are cheap and proven: they already ship by the billions in smartphones and data-center optics, which means the supply chain for Volantis' core ingredient is mature.

The problem Volantis targets is the so-called memory wall. AI accelerators from NVIDIA and AMD are only as fast as they can fetch model weights and activations from high-bandwidth memory, and today's electrical connections are so short that even top-end GPUs pair with roughly eight HBM stacks. That proximity limit forces memory to crowd around the compute die, leaving expensive silicon idle while it waits for data, and pushing cloud providers to buy more GPUs simply to get enough memory capacity.

Because optical links tolerate far longer distances, Volantis says its architecture could let a single GPU connect to as many as 220 memory chips — a more than twenty-fold jump that remains, for now, a company claim that has not been independently verified in production systems. The startup describes its approach as the "A-1" photonic memory architecture, combining micro-VCSELs and photonic interconnect with semiconductor design, and says it is aimed at inference systems running models of more than 20 trillion parameters at speeds of up to 10,000 tokens per second. It intends to deliver a first chip next year.

If the numbers hold up in real hardware, the economics of AI serving could shift. Memory-bound workloads are a major cost driver for inference providers, and a design that lets memory sit farther from compute without sacrificing bandwidth could reduce how many GPUs a deployment needs for a given model — or change where the money flows among accelerators, premium HBM packaging, and the networking that stitches racks together. It would not dethrone NVIDIA or AMD overnight, but it could move pricing power to a layer of the stack they do not currently own.

Volantis is not alone. Optical input/output has become one of the busiest corners of AI infrastructure, with several startups and every large chipmaker exploring laser-based links to scale past the limits of copper, and hyperscalers have been funding the technology for years. The hard part is integration: fitting lasers into the tight, power-hungry, thermally punishing environment of an AI server is a different discipline from using them in networking gear or a phone's face scanner.

That an $88 million round went into the idea anyway is a signal of how valuable the bottleneck has become. With training clusters now measured in the hundreds of thousands of chips and inference demand compounding, even a modest improvement in memory reach would be worth billions across the industry. Whether Volantis can get from prototype to a shipping chip next year — and persuade server makers to rebuild their designs around light — will determine if this round looks prescient or expensive.