Volantis raises $88 million to link GPU and laser memory and house 220 memory chips around a processor

The San Francisco-based start-up is relying on micro-VCSELs to replace copper and remove the memory bottleneck of inference. First deliveries are scheduled for 2027.

Volantis has closed an $88 million Series A round, co-led by Lachy Groom and Abstract Ventures, with John Doerr, VXI Capital, Triatomic and Susa Ventures. The San Francisco startup, which now has raised $97 million, is developing A-1, an inference architecture where the links between computation and memory are made through light rather than copper tracks.

The starting point is well known: in the inference of large models, it is often not the calculation that limits, but the transfer of data between processor and memory. By freeing itself from the distance constraints of copper, Volantis claims to be able to place 220 memory chips around a single GPU, nearly 28 times what Nvidia's most advanced current chips allow.

Lasers already mass-produced

Rather than external lasers, the architecture uses custom-designed micro-VCSELs, a family of lasers that is already manufactured on a large scale (they are found in iPhone Face ID modules, for example). The company intends to rely on the gallium arsenide sector and avoid the supply tensions that affect indium phosphide lasers. The first integrated inference engines are due to be delivered to customers in 2027.