Moving memory centimetres away instead of millimetres
Volantis, a San Francisco chip startup founded in 2022, announced an $88m Series A on 1 October and then set out what it intends to build with it. Its planned accelerator, the A-1, replaces much of the electrical wiring between compute and memory with light, using optical waveguides and micro-VCSEL lasers built directly into the interposer the chips sit on.
The problem it is aiming at is the memory wall. On-chip SRAM has enormous bandwidth and almost no capacity. High-bandwidth memory stacked beside a GPU has far more capacity but runs into a limit on how much data can cross the edge of the die — what the industry calls the shoreline. Volantis’s argument is that optics removes the distance constraint: memory can sit centimetres from the compute rather than millimetres, and still be reached fast.
The numbers are targets, not measurements
Volantis says a single A-1 package should hold more than 10 terabytes of memory at up to 240 terabytes per second of bandwidth, enough to serve a 20-trillion-parameter model at 10,000 tokens per second per user. It also claims 15 times better tokens per dollar than Nvidia’s Rubin platform, six times better efficiency for low-latency mixture-of-experts models above a trillion parameters, and 30 times lower latency on large-model inference.
None of that has been built. The Register reported that chief executive Tapa Ghosh described 12 to 18 months as an aggressive target for a minimum viable product, not production volume; unite.ai reported the company plans to deliver its first integrated inference engines to customers in 2027. Every figure above is Volantis’s own projection for silicon that does not yet exist.
Who is backing it
The round was led by Lachy Groom and Abstract Ventures and brings total funding to $97m. Other investors named include John Doerr, VXI Capital, Triatomic and Susa Ventures, with angel cheques from Dwarkesh Patel, Naveen Rao and Sholto Douglas. Earlier backers include Sam Altman, Jeff Dean and SemiAnalysis’s Dylan Patel.
That investor list is the most informative thing in the announcement. It is not a bet on a product; it is a bet that memory bandwidth, not arithmetic, is what constrains inference economics over the next few years — a position several of those people have argued publicly.
Ghosh told The Register the company intends to license as much of the rest of the design as possible and keep its own work on the optical interposer. The thing to watch is whether a working interposer appears inside the stated window, and whether anyone independent gets to measure it.
