A startup building optical links for AI accelerators has closed an $88 million Series A round, betting that memory bandwidth rather than raw compute is the binding constraint on inference.
Volantis Inc. said the round was led by angel investor Lachy Groom and Abstract Ventures, with more than half a dozen others joining, among them Kleiner Perkins chair John Doerr and Naveen Rao, formerly head of Intel’s AI products group. The founding team includes engineers from Nvidia and Broadcom whose past work took in the first commercial implementation of CoWoS packaging.
The problem Volantis targets is distance. On today’s accelerators, the wires tying memory to processing cores stretch only about five millimeters, which caps how many memory modules fit around a die. Volantis says its interconnect reaches beyond 200 millimeters and can support more than 220 memory chiplets, yielding more than 30 times the bandwidth of current parts.
The links are optical, carrying data as light from vertical-cavity surface-emitting lasers. Volantis argues VCSELs are cheaper and easier to make than the lasers used in networking gear, and can be built from gallium arsenide.
The chips arrive inside a data center appliance called the A-1, roughly a third of a standard rack, with 10 terabytes of memory and 250 terabits per second of bandwidth. The company estimates it can process up to 10,000 tokens a second on a 20-trillion-parameter model.
Chief executive Tapa Ghosh wrote that the design aims at real-time frontier inference, imagining a coding agent that finishes in 30 seconds rather than 30 minutes.
