Volantis raises $88 million for lasers that link AI chips to 220 memory chips
Volantis says Face ID-style lasers could wire one GPU to 220 memory chips, against eight on Nvidia's top parts, though its first system is not due until 2027.
Volantis announced an $88 million Series A on 1 October, co-led by Lachy Groom and Abstract Ventures with John Doerr, VXI Capital, Triatomic and Susa Ventures. Angels include podcaster Dwarkesh Patel, chip veteran Naveen Rao and Anthropic researcher Sholto Douglas, according to the company's release.
The company wants to replace short electrical links between a processor and its memory with light. Volantis says it uses custom micro-VCSELs, the vertical-cavity lasers found in the iPhone's Face ID, rather than external lasers. Reuters notes that this draws on the gallium arsenide supply chain and avoids indium phosphide shortages.
Its pitch is the memory bandwidth limit. Volantis says the approach lets bandwidth grow as memory capacity grows, and that it could eventually connect as many as 220 memory chips around one GPU, against eight on Nvidia's top chips. It says its links use under one picojoule per bit. None of that has been independently measured.
The first product is the A-1 inference system, which Volantis says is designed for models above 20 trillion parameters at up to 10,000 tokens per second per user. It is due in 2027. CEO Tapa Ghosh said the company started out to remove the trade-off between running the largest models and running them fast.
The team comes from Nvidia, AMD, Broadcom and Ayar Labs, with past work on CoWoS packaging, tunable VCSELs and co-packaged silicon photonics, the release says. One analyst write-up, from FourWeekMBA, headlines the round as resting on unverified specifications, and the claims above are the company's own.