Volantis raises $88M to break AI's memory bottleneck with photonics
What's the deal? San Francisco-based VolantisDealroom has a profile for this one. Try Dealroom → has raised $88 million in a Series A round to build AI inference hardware that boosts memory capacity and bandwidth at once. The round was co-led by Physical Intelligence co-founder Lachy Groom and Abstract Ventures.
Who else is in? Participants include venture capitalist John Doerr, VXI CapitalDealroom has a profile for this one. Try Dealroom →, TriatomicDealroom has a profile for this one. Try Dealroom →, and Susa Ventures. Angel investors include Sam Altman, Scale AIDealroom has a profile for this one. Try Dealroom → co-founder Alex Wang, Anthropic engineer Sholto Douglas, and Unconventional AI founder Naveen Rao.
What's the endgame? Volantis is building an inference architecture that uses optics to run models exceeding 20 trillion parameters at up to 10,000 tokens per second per user, while cutting cost per token. "We are solving AI's memory bottleneck by using optics, enabling chips with huge amounts of fast & cheap memory," chief executive Tapa Ghosh said on X.
Why it matters: Large AI models need substantial memory to store parameters and context, while high-speed inference requires bandwidth to move data between memory and compute. Volantis aims to deliver both — the longer-term goal is memory systems that scale by orders of magnitude across hardware generations.
Who's building it? Founded in 2022 by Ghosh, former Micron engineer Roy Meade, and former Nvidia engineer Inderjit Singh, the company draws engineers from Nvidia, AMD, Broadcom, and Ayar Labs. Ghosh is a Thiel Fellow and Y Combinator alumnus who previously founded semiconductor startup Vathys.
What's next? Ghosh said the technology could initially accelerate AI agents, including coding systems that complete tasks in minutes or seconds. The company plans customer deployments in 2027.
The signal: Investor appetite for the picks-and-shovels layer of AI remains strong, with a roster of marquee technologists backing hardware that targets inference cost and speed — the bottleneck as models keep scaling.
Read more: citybiz.co, finance.yahoo.com, prnewswire.com
Image credit: IBM Research
