Volantis raises $88M to link AI chips and memory with light

Volantis is building A-1, a data-center inference system for organizations running very large AI models. Its processors would reach a broader bank of memory over light-based connections instead of depending only on modules packed tightly beside each chip. The San Francisco semiconductor startup announced an $88 million Series A co-led by Lachy Groom and Abstract Ventures.

During inference, processors repeatedly retrieve a trained model’s parameters from memory to generate each response. High-bandwidth memory can move that data quickly, but conventional electrical links require the modules to sit close to the processor. That limits both the memory available to a model and the rate at which processors can receive its parameters.

The constraint is physical as much as computational. SiliconANGLE reported that conventional links keep memory within about five millimeters of processor cores. Volantis says its optical connections can extend beyond 200 millimeters, creating room for more than 220 memory chiplets. Adding memory in that layout is also intended to increase the bandwidth feeding the processors.

The optical design uses microscopic VCSEL lasers. Reuters reported that the same laser family has an established consumer-electronics supply chain, including use in facial-recognition features on Apple devices. Volantis plans to package its chips inside A-1 rather than sell only a component; SiliconANGLE reported that the appliance would occupy about one-third of a standard server rack.

Volantis plans to deliver its first integrated inference engines to customers in 2027. It says A-1 is being designed with 10 terabytes of memory and 250 terabits per second of memory bandwidth, with a target of up to 10,000 tokens per second per user on models exceeding 20 trillion parameters. Those remain projections for a product that is not yet shipping. The round will fund engineering, commercialization and the move toward customer deployments.