6 October 2026

 

An AI coding assistant starts working through a request. It has a powerful processor behind it, but the answer still takes time to appear. Some of that time goes into calculations. Some may be spent waiting for the information those calculations need.

That second problem is where Volantis sees an opening. The Californian semiconductor company is developing a system that uses light to move data between processors and memory. The ambition is to give AI access to much more information without making it wait longer for that information to arrive.

On 1 October, Volantis announced an $88 million Series A, co-led by Lachy Groom and Abstract Ventures, with participation from John Doerr, VXI Capital, Triatomic and Susa Ventures. It plans to deliver its first integrated AI inference systems to customers in 2027. 

Inference is the work a trained AI model does when it answers a question, writes code or analyses a document. For businesses paying to run these services, faster answers are only part of the attraction. Volantis also hopes to reduce the cost of producing them.

A fast processor needs a steady supply of information

A large language model relies on enormous collections of numerical values learned during training. These must be available when the model generates an answer, alongside the working data associated with the request.

Two things matter here. The system needs enough memory to hold the information, and it needs to retrieve it quickly enough to keep the processor busy. Having a large store of data is useful only if the machine can reach what it needs in time.

The pressure often becomes particularly noticeable when a model generates its answer, one fragment of text after another. At this stage, the speed at which information arrives from memory can limit performance. Other stages, such as processing the initial prompt, can place greater demands on calculation instead.

Chipmakers address this by placing fast memory very close to the processor. The arrangement works well, but there is only so much room around a chip. Connecting more memory becomes a problem of space as well as speed.

One response is to spread a large model across additional processors, gaining access to the memory attached to each. That brings more equipment to buy and power, and more communication between processors.

Volantis wants to make a larger pool of memory available through optical connections that reach farther than the short electrical links used around conventional accelerators. The potential gain is straightforward: more information within easy reach of the machinery doing the work.

 

From an early idea to an AI product

 

Volantis was founded in 2022 and emerged from stealth in June 2025 with a $9 million seed round. 

Chief executive Tapa Ghosh leads the company with co-founder and chief technology officer Roy Meade. Meade previously led high-bandwidth memory development at Micron and was an engineering vice-president at optical interconnect company Ayar Labs.

That experience is relevant to the task ahead. Volantis must bring together technology for storing information, technology for moving it and the engineering needed to assemble everything into a dependable machine.

Its first product, A-1, combines the company’s optical connections with established processing technology licensed from other suppliers. The light moves the data; electronic processors still perform the calculations. 

The company’s plans have therefore become more concrete since its prototype announcement: an integrated product intended to run large AI models. Volantis says the latest investment brings total funding to $97 million and will support engineering expansion and commercialisation. 

 

Small lasers, a larger memory pool

 

Volantis uses tiny lasers from a familiar family of devices: vertical-cavity surface-emitting lasers, usually shortened to VCSELs. Versions of these lasers already appear in facial recognition systems. An established manufacturing base gives Volantis a starting point for sourcing and producing its own customised devices.

Its approach uses many optical connections in parallel. Sharing the traffic across numerous links is intended to move large amounts of information without requiring each link to operate at an extreme speed or consume excessive energy. 

The company says its architecture could connect more than 220 memory chiplets in one pool. Reuters reported its comparison with eight high-bandwidth memory stacks around a GPU in NVIDIA’s leading offerings. The devices are not equivalent, so this does not mean 27 times the memory or 27 times the speed. It shows the scale of the change Volantis is attempting. 

For users, the value would come from what that extra access enables. A coding assistant working with a large codebase, for example, needs both room to hold relevant information and the ability to use it quickly. Volantis is targeting that combination.

 

The machine must earn its place

 

A-1 is being designed as a substantial data-centre system, rather than a small component that customers simply plug into an existing graphics card. Its published plans include ten terabytes of memory and an enclosure occupying fifteen standard units of rack height. 

Those figures describe the intended product. Volantis has reported measurements from optical links, but these do not establish how well a complete A-1 system will run customers’ models.

The practical questions are how quickly it produces useful answers, how much electricity it consumes and what it costs to buy and operate. The optical components must also survive the heat and continuous use of a data centre, and be manufactured reliably enough to support commercial deliveries.

 

Other routes to faster AI

 

Volantis is part of a wider effort to use light inside AI infrastructure. Ayar Labs and Lightmatter are developing optical connections that help processors exchange information. Their products and designs differ from A-1, but they address the same broad pressure: moving data efficiently as AI systems grow. 

Avicena uses tiny LEDs rather than lasers for its optical links. In August, it announced shipments of evaluation kits for applications including connections between processors and memory. These allow customers to test the technology; they are not complete AI inference systems. 

Established chipmakers are also improving memory access. NVIDIA’s Rubin platform includes faster memory and changes intended to help processors use it effectively. Volantis will have to demonstrate an advantage against that continuing progress. 

Its opportunity is to give customers a more economical way to keep large models supplied with data. If A-1 can do that, operators could need fewer processors simply to obtain enough fast memory. The next step is to show that the savings remain worthwhile once the optical technology, assembly and cooling are included in the bill.

 

Further reading on MoveTheNeedle.news:

How Xscape Photonics plans to solve AI data centres’ networking bottleneck with silicon photonics

The startup betting on a more energy efficient general purpose processor