The startup says optical links could let AI processors reach far more memory than today’s electrical interconnects, turning a familiar smartphone laser technology into a high-stakes bet on faster inference.
The latest attempt to loosen one of artificial intelligence’s hardest hardware constraints is coming from a relatively small semiconductor company with a very large claim. San Francisco-based Volantis said on October 1 that it had raised $88 million to develop an optical interconnect system intended to move data between processors and memory far more efficiently than the short electrical links used in today’s accelerators. The financing was reported by Reuters, which described the company’s plan to use vertical-cavity surface-emitting lasers, or VCSELs, to connect compute with a much larger pool of memory.
The importance of the idea is less about the size of the funding round than about the problem Volantis is targeting. Modern AI systems are often discussed in terms of raw arithmetic throughput, but increasingly the limiting factor is how quickly processors can obtain the data they need. Large language models are stored in memory, and every inference request forces the system to move enormous quantities of weights and intermediate results between memory and compute. When that movement becomes the bottleneck, adding more arithmetic units delivers diminishing returns.
The memory wall has become an AI infrastructure problem
High-bandwidth memory has become one of the most valuable components in the AI supply chain because it places fast memory close to graphics processors and other accelerators. Nvidia, AMD and other chip designers have built increasingly sophisticated packaging around this principle. But proximity brings physical constraints. Reuters reported that current leading systems can place only a limited number of high-bandwidth memory stacks around a processor because the electrical connections are short and the package itself has finite area.
Volantis is trying to replace part of that electrical path with light. Its premise is that optical connections can travel farther without the same signal losses that constrain copper wiring at extreme data rates. The company says that this could allow a processor to communicate with a dramatically larger number of memory chiplets, creating a system in which memory capacity and bandwidth can scale more independently from the dimensions of a single package.
That is a technically ambitious proposition, but it addresses a well-established industry problem. The rapid growth of AI models has pushed memory capacity, bandwidth and energy consumption into the center of data-center design. Training remains compute-intensive, but inference is becoming a larger commercial market as enterprises deploy models continuously. Inference workloads often require large models to remain immediately accessible, which means that the cost and speed of moving data can matter as much as the processor’s peak calculation rate.
Why Volantis is using a technology already found in consumer electronics
One of the most notable parts of the Volantis strategy is its reliance on VCSEL technology rather than an entirely novel light source. These tiny lasers are already produced at large scale for sensing applications, including facial-recognition systems in smartphones. Reuters noted that Apple’s investment in the VCSEL supply chain helped establish a mature manufacturing base over the past decade.
That existing supply chain could matter because many photonics projects fail not at the level of laboratory performance but at the point where they must be manufactured cheaply, consistently and in very large volumes. A component that works in a prototype can still be unsuitable for a data center if yields are poor, packaging is fragile or production depends on materials that cannot scale. Volantis is effectively betting that it can combine known optical components with advanced semiconductor packaging in a way that reduces those commercialization risks.
The company is not claiming that packaging will be simple. Optical links introduce their own demands in alignment, thermal control, power delivery and testing. The challenge is to make the full system economical rather than merely demonstrating that light can carry the signal. That distinction will determine whether Volantis becomes an important supplier or remains one more promising photonics project that struggles to move beyond early hardware.
The 220-memory-chip claim is the number investors will watch
According to Reuters, Volantis says its architecture could allow as many as 220 memory chips to be connected around a GPU, compared with the much smaller number supported by conventional short-reach electrical interconnects. The figure is a company projection rather than an independently validated production result, and that distinction matters. Yet it provides a useful measure of what the startup is trying to change: the amount of memory a processor can reach without turning every additional byte into a packaging crisis.
If a system can place much more memory within a high-speed domain, developers could run larger models without splitting them across as many separate accelerator nodes. That could reduce some communication overhead, simplify certain inference workloads and make very long context windows more practical. The gains would depend heavily on software, network architecture and the actual latency of the optical system, but the direction is clear: more directly accessible memory would give system designers another lever besides simply adding more GPUs.
This is particularly relevant because AI infrastructure spending is increasingly judged by utilization rather than headline chip counts. An expensive accelerator that spends time waiting for memory is an underused asset. Cloud providers and model developers therefore have strong incentives to improve the balance between compute, memory and networking. Volantis is entering a market where even a modest improvement in that balance can be worth billions if it scales across large fleets.
The startup is entering a crowded photonics race
Optical interconnects are not a new idea, and Volantis will face competition from established semiconductor companies, specialized photonics startups and research groups pursuing different approaches to co-packaged optics. The broader industry has already begun moving optical links closer to processors because electrical signaling becomes increasingly expensive and power-hungry at the highest bandwidths.
What distinguishes the Volantis pitch is the emphasis on memory rather than primarily on connecting servers or switches. The company wants light to extend the effective reach between compute and high-bandwidth memory. SiliconANGLE reported that Volantis plans an inference system called A-1 built around its photonic memory architecture, with the company targeting customer deployments after development and validation.
The startup’s timing reflects a broader shift in the AI hardware market. The first phase of the generative-AI boom rewarded whichever suppliers could deliver accelerators fastest. The next phase is increasingly about the efficiency of the entire computing stack: memory, optical networking, cooling, power systems, packaging and software. That opens room for companies that do not compete directly with Nvidia on the processor core but can remove a bottleneck around it.
Funding gives Volantis time, but not proof
The $88 million round was led by former Stripe executive Lachy Groom and Abstract Ventures, with investors including John Doerr, VXI Capital, Triatomic and Susa Ventures, according to Reuters. The investor list gives the company credibility and enough capital to expand engineering, tape out hardware and build systems, but semiconductor development consumes cash quickly. Packaging experiments, wafers, test equipment and specialist hiring can absorb tens of millions before meaningful revenue appears.
The central milestone will therefore be technical validation. Volantis needs to show that the optical interconnect works at the promised bandwidth and distance, that the system can be manufactured with acceptable yields, and that the total power and cost profile is competitive with alternative architectures. It will also need software support so developers can use the additional memory without redesigning every application.
A second milestone is schedule. Reuters reported that Volantis aims to deliver a chip next year. In semiconductors, a one-year target can compress several difficult stages into a short period: design completion, fabrication, packaging, bring-up, debugging and customer qualification. Any one of those stages can introduce delays. Investors will be watching not just whether a chip exists, but whether it performs consistently enough to move into production.
The economics of inference are becoming as important as benchmark speed
AI companies increasingly need to answer a simple commercial question: how much does it cost to serve one useful response? The answer depends on model size, utilization, electricity, memory, networking and the amount of hardware required to meet latency targets. Faster memory access can lower that cost if it allows processors to spend more time computing and less time waiting.
That makes memory architecture a business issue, not merely an engineering detail. Cloud platforms price inference services in ways that eventually reflect infrastructure cost. Enterprises deciding whether to deploy a larger model must weigh better capabilities against higher serving expense. If photonic memory links can materially improve efficiency, the effect could spread from semiconductor design into cloud pricing and the economics of AI applications.
But there is an important counterpoint. A breakthrough at one layer often shifts the bottleneck elsewhere. More memory bandwidth can expose limits in compute, networking or power delivery. A larger memory pool can encourage developers to build even larger models, consuming the efficiency gain. The history of computing is full of improvements that reduce one constraint only to reveal the next.
A test of whether photonics can move from network edge to memory fabric
The most interesting question raised by Volantis is whether optical technology can move deeper into the internal architecture of AI systems. Photonics already plays a central role in data-center networking, especially between racks and clusters. Extending it into the memory fabric would mark a meaningful architectural shift because it would change how designers think about the physical boundary of an accelerator.
If successful, such systems could weaken the assumption that memory must sit only millimeters from compute. That would create more flexibility in how chiplets are arranged and how capacity is upgraded. It could also make specialized inference systems more attractive, with architectures designed around very large model memory rather than general-purpose acceleration.
For now, however, Volantis remains at the stage where engineering claims must become working silicon. The company has capital, experienced backers and a problem that every major AI infrastructure operator recognizes. What it does not yet have is proof at production scale.
The next year will show whether the memory wall can be turned into an optical opportunity
The AI hardware industry has spent the past several years racing to build faster processors. The next competitive advantage may come from making those processors wait less. Volantis is betting that a laser technology familiar from consumer devices can be repurposed into a high-bandwidth bridge between compute and memory, potentially allowing much larger memory pools to sit within reach of a single accelerator.
That is a plausible direction, but the gap between plausible and manufacturable remains large. The company must prove bandwidth, reliability, power efficiency, packaging and economics at the same time. If it does, the $88 million round could look like an early investment in a new layer of AI infrastructure. If it does not, the memory wall will remain one of the most expensive constraints in the industry.
Either way, the funding signals where attention is moving. AI’s next hardware contest is no longer only about who can build the fastest processor. It is increasingly about who can connect processors to memory, networks and power systems efficiently enough to keep them useful. Volantis has chosen one of the hardest of those bottlenecks, and the market will now get a chance to see whether light can carry more than data — whether it can carry the economics of the next generation of AI inference.
