Most AI silicon asks buyers to pick a side. Fast on-chip SRAM holds too little to fit a large model or a long context window, while the bigger HBM pools that solve capacity serve data more slowly. Volantis, a San Francisco semiconductor company, says four years of work produced hardware that sidesteps that tradeoff, and on October 1 it closed an $88M Series A to carry the design to market.
Lachy Groom and Abstract Ventures co-led the round. VXI Capital, Triatomic and Susa Ventures joined, along with angels Dwarkesh Patel, Naveen Rao and Sholto Douglas. Sam Altman, Jeff Dean and Dylan Patel are among the earlier backers, and the company puts total funding at $97M.
The centerpiece is the A-1 system, which pairs a photonic interconnect with off-chip memory. Where optics usually link one chip to the next, Volantis turns them on memory instead. Many memory dies are gathered into a single fabric, so each addition of capacity also adds bandwidth. Volantis is aiming at models beyond 20 trillion parameters, with as much as 10,000 tokens per second for every user, 10 TB of capacity and 240 TB/s of bandwidth; the rack unit draws 20kW and stands 15U tall. The company also claims 15 times better tokens per dollar than Nvidia’s Rubin on some low-latency workloads.
Chief executive Tapa Ghosh argues that speed has become a business metric for agents, since a coding agent that finishes in two minutes rather than 30 lets developers iterate far more often. Systems will not ship immediately, though: Volantis expects its first integrated inference engines to reach customers at some point in 2027, with more architecture details to come as the product nears commercialization.
