Volantis, a San Francisco semiconductor startup, has raised US$88 million in Series A funding to develop a new photonic architecture for AI inference. The round was co-led by Lachy Groom and Abstract Ventures, with participation from John Doerr, VXI Capital, Triatomic, and Susa Ventures. The capital will support the company's first system, A-1, which is designed to run models exceeding 20 trillion parameters at up to 10,000 tokens per second per user while reducing inference cost per token.
The AI Memory Wall
Running large AI models at high speeds requires enormous memory capacity to hold the model and enormous bandwidth to continuously feed data into the compute engine. On-chip SRAM provides high bandwidth but limited capacity, while HBM-based systems offer more capacity but remain constrained by bandwidth, cost, and energy requirements. Volantis argues that even emerging approaches such as 3D DRAM remain on the same underlying tradeoff curve, so its A-1 system is designed to increase memory capacity and bandwidth simultaneously.
A Faster Path for AI Agents
According to Tapa Ghosh, CEO and co-founder of Volantis, the speed of AI agents will increasingly determine how quickly companies can operate. Existing hardware forces a tradeoff between running the largest, most sophisticated models and running them fast, which became a key motivation for founding Volantis. The company believes that accelerating inference could allow a coding agent to complete a task in two minutes instead of 30, giving developers more opportunities to test ideas and iterate.
A New Photonic Interconnect Approach
Volantis is developing a photonic interconnect designed specifically to connect compute chips with memory, a different focus from existing chip-to-chip optical technologies. Its optical fabric connects large numbers of memory chips into a unified pool, aggregating bandwidth as memory is added to increase capacity and bandwidth together. The architecture uses custom micro-VCSELs instead of external lasers, leveraging the established gallium arsenide VCSEL supply chain while avoiding indium phosphide supply constraints.
Commercial Milestones and Product Targets
The A-1 system is being designed to support models exceeding 20 trillion parameters at up to 10,000 tokens per second per user, while reducing inference cost per token. These figures represent product targets as the company moves toward commercialization, not performance from a deployed system. Volantis plans to deliver its first integrated inference engines to customers in 2027, with the new funding supporting engineering expansion and customer deployments.
Competing in the AI Infrastructure Market
Volantis enters a growing AI infrastructure market that has attracted substantial investment. Lightmatter has raised US$400 million at a US$4.4 billion valuation, and Ayar Labs raised US$500 million in Series E funding at a US$3.8 billion valuation. Celestial AI has raised US$250 million at a US$2.5 billion valuation, while Volantis is approaching the problem from the memory side by betting that photonics can help AI systems scale both capacity and bandwidth.
With the Series A round in place, Volantis now faces the challenge of turning its photonic memory architecture into a commercially deployable inference platform. Its ability to reduce the tradeoff between memory capacity and bandwidth could make large models significantly faster and more cost-effective for AI agent workloads. The next milestone will be delivering integrated inference engines to customers by 2027 and demonstrating that the technology can meet its ambitious performance targets.