Volantis raises $88 million in a Series A to develop A-1, an AI inference system that uses optical connections between compute and memory. The San Francisco semiconductor startup says the round brings its total funding to $97 million.

 

The A-1 plan rests on three ambitious targets:

  • Connect more than 220 memory chiplets in one pool.
  • Serve models exceeding 20 trillion parameters.
  • Deliver initial customer inference engines in 2027.

 

Related Research

 

Volantis Raises $88 Million for A-1

The financing was co-led by Lachy Groom and Abstract Ventures, according to the company's funding announcement. John Doerr, VXI Capital, Triatomic and Susa Ventures participated, alongside individual investors including Dwarkesh Patel, Naveen Rao and Sholto Douglas.

 

Volantis plans to use the money to develop and commercialize A-1, expand its engineering team and move the system toward customer deployments. Its founding team includes veterans of Nvidia, AMD, Broadcom and optical-interconnect company Ayar Labs.

 

The company's later October 1 release describes A-1 as designed for models exceeding 20 trillion parameters and speeds of up to 10,000 tokens per second per user. Those figures are design targets published by Volantis, not results from independent customer benchmarks.

 

Reuters reported that Volantis aims to deliver a chip next year and interviewed founder and chief executive Tapa Ghosh about the manufacturing approach. The company separately says its first integrated inference engines should reach customers in 2027.

 

Optical Fabric Connects Compute to Memory

Modern AI accelerators need to move model weights and intermediate data between processors and memory continuously. That movement becomes a bottleneck when electrical connections cannot reach far enough to add more memory packages without compromising bandwidth or latency.

 

Volantis proposes replacing those short electrical links with an optical fabric built for chip-to-memory communication. Its architecture uses integrated micro vertical-cavity surface-emitting lasers, or micro-VCSELs, and optical waveguides instead of external lasers and conventional optical fiber.

 

The company's technical overview says electrical interposer wires typically travel about two to five millimeters, while its optical waveguides extend beyond 200 millimeters. Volantis says that reach allows more than 220 memory chiplets to form a single, uniform-latency pool.

 

The design aggregates bandwidth as memory is added and uses lower-cost off-chip memory rather than depending only on tightly packed high-bandwidth memory stacks. Volantis reports more than 200 terabytes per second of bandwidth, sub-five-nanosecond latency and link energy below one picojoule per bit.

 

Why Memory Limits AI Inference

AI inference is often described as a compute problem, but large models also require enough memory capacity to hold their parameters and enough bandwidth to keep processing units supplied with data. Improving only one side can leave the other as the limiting factor.

 

On-chip SRAM offers high bandwidth and low latency but limited capacity. High-bandwidth memory provides more capacity near a GPU or accelerator, yet physical reach, packaging complexity, power and cost constrain how many stacks can be placed around the processor.

 

Volantis argues that a longer optical connection changes that geometry. A larger memory pool could support bigger models and longer contexts without forcing all memory onto the same crowded package edge, while greater aggregate bandwidth could reduce the time agents spend generating long outputs or completing multi-step work.

 

The approach also relies on a mature component family. VCSELs are already produced at large scale for sensing applications, including facial-recognition hardware. Volantis is adapting that supply chain to a far denser data-movement problem inside AI systems.

 

Customer Hardware Must Validate the Claims

A-1 is not yet a generally available product, and Volantis has not published independent benchmarks, customer deployments or a full system power profile. Its parameter scale, token speed and cost claims therefore describe the company's intended performance rather than demonstrated results in production.

 

Packaging remains another engineering challenge. The architecture must integrate lasers, waveguides, memory and compute while maintaining yield, thermal stability and reliable links across a complex interposer. Familiar components can reduce supply-chain risk without eliminating the difficulty of assembling them at scale.

 

The 2027 delivery target provides a clear test. Customer silicon will need to show that optical reach produces usable capacity and bandwidth gains after accounting for software, manufacturing, cooling, error rates and total system cost.

 

If Volantis meets those requirements, the company could attack one of inference's most persistent constraints without asking customers to abandon established memory manufacturing. Until then, the $88 million round funds an unusually concrete architectural bet whose most important evidence will come from working A-1 systems.