Photo: Steve A Johnson / Pexels
Volantis raises $88M Series A to attack the AI memory bottleneck with light
The San Francisco chip startup is betting optical links between compute and memory can make AI inference faster and bigger
AI models keep getting larger, and the chips running them keep waiting on memory. Volantis just raised $88 million to try fixing that problem with lasers.
The San Francisco semiconductor startup closed a Series A on October 1, 2026, bringing its total funding to $97 million. Its pitch: replace part of the electrical plumbing between AI chips and memory with optical connections.
The round and the roster
Lachy Groom and Abstract Ventures co-led the round. John Doerr and VXI Capital also participated.
That follows a $9 million seed round in 2025, backed by names including Alex Wang and Trevor Blackwell.
The leadership team brings semiconductor and photonics experience from NVIDIA, AMD, Broadcom, and Ayar Labs. CEO Tapa Ghosh leads the company.
How Volantis wants to fix memory
Compute chips have gotten dramatically faster, but feeding them data from memory has become a limiting factor for AI inference, the stage where a trained model actually answers your prompts.
Volantis is attacking this with photonic interconnects. The core component is a custom micro-VCSEL, short for vertical-cavity surface-emitting laser.
AI, tech, and the markets they move—in one daily briefing.
Daily. Free. Join 34,000+ readers across crypto, finance, and policy.
These are tiny lasers that shoot light straight up off a chip’s surface. Volantis uses them to create optical links between compute chips and memory.
The company’s architecture is designed to connect a GPU to as many as 220 memory chips, versus around 8 chips in traditional setups.
The VCSEL technology is already proven in consumer devices, including the iPhone’s Face ID. Volantis is adapting it and pointing it at a data center problem.
According to the company’s approach, this could allow memory to be aggregated more efficiently without requiring radical changes to packaging or manufacturing.
Meet A-1, the first system
Volantis is building its first system, called A-1, with the following targets:
- Designed to handle models with more than 20 trillion parameters
- Up to 10,000 tokens per second per user
- About 10 terabytes of memory capacity
- 240 terabytes per second of bandwidth
Volantis plans to deliver A-1 integrated inference engines to customers by 2027.
Why the big chipmakers are the benchmark
Volantis is positioning its optical approach as a way past the memory limits that conventional accelerators from NVIDIA and AMD face. Several of Volantis’s own team members came from those same companies.
Volantis is not necessarily trying to out-muscle the incumbents on raw compute. Its bet is narrower: that memory access, not processing power, is where the next big gains in inference will come from.