Volantis raises $88M to put optical memory links inside AI systems

Why it matters

Volantis is raising to move photonics from a component-level promise into an AI inference system. Its 2027 delivery target makes manufacturing and customer validation the next test for claims about memory capacity, bandwidth and inference speed.

Volantis raised an $88 million Series A on October 1st to build optical links between AI processors and memory, betting that moving data faster inside a system can ease a bottleneck that more computing power alone does not solve. The San Francisco semiconductor startup says its first system, A-1, is designed for models exceeding 20 trillion parameters and up to 10,000 tokens per second per user. Those are company targets, not reported customer results; Volantis plans its first integrated inference-engine deliveries for 2027.

Founder and CEO Tapabrata "Tapa" Ghosh (@semiDL) has been working on the data-movement problem since before the current AI hardware rush. As a teenager, Ghosh built a quadcopter to explore deep learning for collision-free drones, then co-founded chip startup Vathys. In a 2017 Stanford talk, he argued that data movement, rather than computation, was the source of inefficiency in deep-learning processors. Volantis is pursuing a related thesis with a different tool: using light to connect compute and memory.

The Series A was co-led by Lachy Groom and Abstract Ventures, with John Doerr, VXI Capital, Triatomic and Susa Ventures also participating, according to Volantis' funding announcement. Angel investors named in the round include Dwarkesh Patel, Naveen Rao and Sholto Douglas. Volantis says it has raised $97 million in total, including the new round. The funding is earmarked for A-1 development and commercialization, engineering hires and preparation for customer deployments.

Volantis' system is intended to change how much memory a processor can reach, and how quickly. In current designs, SRAM offers high bandwidth but limited capacity, while high-bandwidth memory (HBM) provides more capacity with a bandwidth tradeoff. Volantis says its optical fabric pools memory chips and adds bandwidth as capacity grows, using optical waveguides and custom micro-VCSEL lasers rather than relying on external lasers. The company's technology description claims optical links can travel more than 200 millimeters across an interposer, compared with a few millimeters for electrical wires, and consume less than one picojoule per bit.

Those are ambitious system-level claims. The company's website describes measured optical links and wafer-scale results, while the headline performance figures for A-1 - including 10,000 tokens per second per user and support for models above 20 trillion parameters - remain targets for a system that has not yet reached its planned customer-delivery date. Reuters reported that the design could connect as many as 220 memory chips around a processor; that is a proposed configuration, not a shipped product. Volantis says an iteration of its technology is taped out, a step in fabrication that precedes evaluation of the resulting silicon.

The engineering pitch rests on a team with experience in several relevant chip technologies. Co-founder and CTO Roy Meade led Micron's HBM program and was a vice president at Ayar Labs, according to Volantis' team page. The company credits its staff with prior work on CoWoS packaging, silicon-photonics co-packaged optics and VCSEL lasers at companies including NVIDIA, AMD, Broadcom and Ayar Labs. Ghosh's earlier Vathys work and the team's semiconductor backgrounds help explain why investors are funding a hardware program years ahead of planned delivery.

Volantis emerged from stealth in June 2025 with a $9 million seed round, then described a broader photonic-computing platform. Its current financing is tied to a more specific commercial bet: A-1 as an inference system that uses optics for compute-to-memory connections. That places the company in a crowded effort to use photonics in AI infrastructure, while distinguishing its target as the connection between processor and memory rather than primarily links between separate chips or systems.

The immediate commercial test is whether Volantis can turn its optical interconnect into a manufacturable system with the promised capacity, bandwidth, cost and reliability. The company's announced funding buys time and engineering capacity to reach that test. Until customer deliveries begin, the proposed tokens-per-second and model-size gains remain a forecast about what its architecture may enable.