Sam Altman-backed Volantis, founded in 2022, is developing an AI accelerator called A-1 that uses a photonic interposer to move memory farther from the compute die. By replacing electrical interconnects with optical waveguides, the company says it can overcome the physical shoreline limit that caps memory capacity and bandwidth on conventional GPUs. It claims the design can support upwards of 10 TB of memory at speeds up to 240 TB/s, enough to serve a 20-trillion-parameter model at 10,000 tokens/s per user.
The startup's differentiation is the interposer itself: it integrates micro-VCSELs directly into the package rather than relying on external lasers. CTO Roy Meade says the design favors many slower links over a few fast ones, similar to how HBM achieves high bandwidth. Volantis is targeting 1 picojoule per bit per 24 Gbps lane and around 20kW total system power for its first product.
Unlike Lightmatter or Celestial AI, which have explored similar photonic interposer ideas, Volantis is not licensing its technology to other chipmakers. Instead, it plans to build a full inference chip, licensing compute and memory IP from partners. The company raised $88 million this week, but CEO Tapa Ghosh says it will need significantly more capital to reach production. It is aiming for an MVP in 12 to 18 months.