Volantis Raises $88M to Build New AI Inference Architecture

Volantis has secured $88 million in Series A funding to advance its innovative AI inference architecture. The round was co-led by Lachy Groom and Abstract Ventures, and included participation from notable investors such as John Doerr, VXI Capital, Triatomic, Susa Ventures, Dwarkesh Patel, Naveen Rao, and Sholto Douglas. Based in San Francisco, the company aims to tackle the memory bottleneck in modern AI infrastructure by developing a new architecture that improves both memory capacity and bandwidth concurrently.

Large AI models necessitate substantial memory capacity during inference processes, while also requiring high bandwidth to transfer data into compute systems efficiently. Traditional architectures face a trade-off, with on-chip SRAM offering high bandwidth but limited memory and GPU systems relying on high-bandwidth memory to provide greater capacity, but at a cost to speed.

Volantis' A-1 system seeks to bridge this gap by supporting models with over 20 trillion parameters and achieving up to 10,000 tokens per second per user. The architecture promises to reduce inference costs per token, thus benefiting the development of complex AI agents that could complete tasks much faster. The company's photonic architecture is a key component, utilizing scalable optical links to connect compute and memory systems.

This approach allows for increased bandwidth and memory capacity to grow together as more memory chips are added.

The photonic interconnect employs custom micro-VCSELs, which utilize an established gallium arsenide manufacturing process to ensure compact, temperature-stable, and energy-efficient components. These micro-VCSELs are designed to connect up to 220 memory chips around a GPU, consuming less than one picojoule per bit for end-to-end links.

With the new funding, Volantis will focus on developing and commercializing the A-1 system, with plans to deliver integrated inference engines to customers by 2027. The founding team, composed of veterans from NVIDIA, AMD, Broadcom, and Ayar Labs, brings extensive experience in packaging, VCSELs, and silicon photonics. The company's approach leverages proven technologies rather than untested breakthroughs, positioning A-1 as a viable alternative to address AI memory supply chain challenges.