The Pulse
Volantis Raises $88M for Photonic AI Memory System
Volantis announced an $88 million Series A to develop a photonic system for AI inference, with its A-1 design targeting models above 20 trillion parameters. The company says it plans to deliver its first integrated inference engines to cust
AI.info Team ·
“Today's hardware forces a tradeoff between running the largest, most sophisticated models and running them fast. We started Volantis to eliminate that tradeoff.”
Tapa Ghosh, CEO and co-founder of Volantis
Volantis announced an $88 million Series A on October 1 to develop a photonics-based system for AI inference, betting that optical links between processors and memory can ease a hardware constraint on serving large models. The San Francisco semiconductor startup says its planned A-1 system will support models with more than 20 trillion parameters at up to 10,000 tokens per second per user. Those are company targets, not reported results from a shipping product.
Volantis aims light at the memory connection
AI accelerators need to move model data from memory to compute quickly. Volantis argues that current systems face a compromise: on-chip SRAM can deliver high bandwidth but has limited capacity, while systems using high-bandwidth memory offer more capacity but cannot feed data to processors at the same rate. The startup’s proposed optical fabric is intended to expand both capacity and bandwidth by connecting many memory chips into a shared pool.
The company says its design uses custom micro-VCSELs—vertical-cavity surface-emitting lasers—instead of external lasers. It says those components can draw on a gallium arsenide supply chain and avoid constraints associated with indium phosphide. Volantis also claims its end-to-end optical links can consume less than one picojoule per bit; the announcement does not include an independent comparison or a customer deployment demonstrating that figure.
A-1’s speed claims are still targets
Volantis says A-1 is being 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 company’s announcement presents that speed as a way to shorten agent workflows, including coding tasks. It plans to deliver its first integrated inference engines to customers in 2027, leaving the performance claims to be tested against hardware, software and production requirements.
The underlying idea is not simply to use optics in a data center. Volantis says chip-to-memory connections move far more data over shorter distances than other optical links, demanding different energy and cost tradeoffs. Reuters reported that Ghosh described advanced packaging as difficult but not unprecedented, and said the company expects to deliver a chip next year. That interview offers a grounded view of the engineering challenge behind the company’s product pitch.
The round’s size differs across Volantis announcements
The October 1 release calls the financing an $88 million Series A, co-led by Lachy Groom and Abstract Ventures. John Doerr, VXI Capital, Triatomic and Susa Ventures also participated, alongside angel investors Dwarkesh Patel, Naveen Rao and Sholto Douglas. Volantis says the money will fund A-1 development, commercialization and engineering-team growth.
A company post dated September 29 described the Series A as $80 million and said Volantis had raised $97 million in total. The October 1 release instead gives the round as $88 million.
Commercialization will test the architecture
The funding gives Volantis resources to build toward its 2027 customer deliveries, but the central test is whether its optical links can deliver the capacity and bandwidth gains the company promises in a complete inference system. The company’s release does not identify customers or provide a shipping date more specific than 2027. For now, the concrete announcement is an $88 million financing and a development target: A-1, a photonic inference system that Volantis says will connect compute and memory with light.