Volantis Raises $88M Series A for Photonic AI Inference

Volantis raised an $88 million Series A for a photonic AI inference architecture designed to run models exceeding 20 trillion parameters at up to 10,000 tokens per second per user, according to the company’s October 1, 2026 PR Newswire release. The round was co-led by Lachy Groom and Abstract Ventures.

The numbers on the wire

• $88 million Series A; no valuation disclosed.
• Co-leads: Lachy Groom and Abstract Ventures.
• Participating: John Doerr, VXI Capital, Triatomic, Susa Ventures; angels Dwarkesh Patel, Naveen Rao, Sholto Douglas.
• Target system A-1: models > 20 trillion parameters; up to 10,000 tok/s per user (company design targets).
• Photonic links claimed at < 1 pJ/bit end-to-end via custom micro-VCSELs.
• First integrated inference engines planned for customer delivery in 2027.

What Volantis is building

Volantis frames today’s inference hardware as stuck on a memory-capacity vs. bandwidth tradeoff: on-chip SRAM is fast but small; HBM-class GPUs hold more, but bandwidth caps how fast large models serve. The company says A-1’s photonic fabric connects many memory chips into one pool so capacity and bandwidth rise together, aiming for nearly two orders of magnitude gains on both axes versus that tradeoff curve.

The interconnect is purpose-built for chip-to-memory, not just chip-to-chip optics already used in data centers. Volantis uses custom micro-VCSELs on gallium arsenide rather than external indium-phosphide lasers, citing supply-chain and power reasons. Founding engineers came from NVIDIA, AMD, Broadcom and Ayar Labs; CEO and co-founder is Tapa Ghosh.

What the release leaves out

No valuation, revenue, customer names, tape-out status, process node, or silicon availability dates beyond “2027.” The 20T-parameter / 10,000 tok/s figures are design targets in a company release, not published third-party benchmarks. Company site: volantissemi.ai.

Why it matters

Inference cost and latency for long agent workloads are becoming the bottleneck after training. A photonic memory architecture that claims to break the capacity/bandwidth tradeoff is a bet that next-gen serving will not look like bigger HBM stacks alone — if Volantis can ship silicon on the 2027 timeline it set.