Volantis' Photonics Strategy Positions Startup as Hyperscaler Acquisition Target

In the race to scale artificial intelligence, the primary constraint is no longer just raw compute, but the physical limits of how memory connects to that compute. This 'memory wall' is defined by the 'shoreline'—the physical perimeter of the compute die—which limits the capacity and bandwidth of modern GPUs. As The Register first reported, AMD's MI455X currently caps at approximately 432 GB.

Volantis, a San Francisco-based startup founded in 2022 and backed by Sam Altman, is positioning itself as the solution to this bottleneck. Rather than attempting to reinvent compute architectures or memory technologies, Volantis is focusing exclusively on the optical interposer. By utilizing optical waveguides instead of electrical interconnects, the company aims to move memory a few centimeters away from the compute die rather than a few millimeters, effectively eliminating the shoreline bottleneck.

From a deals perspective, Volantis is executing a lean integration strategy. CEO Tapa Ghosh told The Register that the company intends to license almost everything other than the interposer from IP providers to mitigate risk. This approach allows Volantis to focus its resources on its primary differentiator: an interposer that integrates micro vertical-cavity surface-emitting lasers (micro-VCSELs) directly into the design. While VCSELs have historically faced reliability concerns, CTO Roy Meade told The Register that a 'wide and efficient' philosophy—prioritizing more links over raw link speed, similar to the HBM playbook—can achieve significantly higher reliability.

The technical ambitions are massive. Volantis claims its A-1 accelerator could support up to 10 TB of memory at speeds reaching 240 TB/s on a single package, which the company says would be sufficient to serve a 20 trillion parameter model at 10,000 tok/s per user. Regarding efficiency, the company is aiming for a total system power of around 20kW and 1 picojoule per bit per 24 Gbps lane.

While other firms like Lightmatter and Celestial AI have explored silicon photonic interposers, Volantis is distinguishing itself by developing a full inference chip rather than acting as a pure IP play. This shift toward a shippable product makes the company a highly attractive strategic asset for hyperscalers who are desperate to solve the memory bottleneck at scale. The precedent for such a move exists; The Register notes that Marvell's $5.5 billion acquisition of Celestial AI was likely driven by the potential for 'HBM-at-a-distance.'

However, the path to a production-ready chip remains capital-intensive. Volantis recently raised $88 million with participation from Lachy Groom, Abstract Ventures, and other investors. CEO Tapa Ghosh acknowledged that significantly more venture capital will be required to move from optical demonstrators to a final product. The company is currently targeting a minimum viable product (MVP) within 12 to 18 months.