AI chips are getting faster, but moving data to memory is becoming one of the industry’s biggest constraints. San Francisco semiconductor startup Volantis is betting that the answer is not another larger GPU, but replacing short electrical connections between compute and memory with optical links.
The company announced an $88 million Series A on October 1, led by Lachy Groom and Abstract Ventures, to commercialise its photonic AI inference architecture. Reuters reports that Volantis is targeting the fundamental bottleneck between the processing cores that run AI models and the memory that feeds them.
The memory problem behind today’s AI boom
Modern AI accelerators rely heavily on high-bandwidth memory, or HBM, because large models constantly move weights and intermediate data between compute and memory. Increasing compute capacity alone does not guarantee proportional performance if the processor cannot receive data quickly enough.
That creates a difficult trade-off: putting more high-speed memory close to a processor increases bandwidth and capacity, but the physical connections themselves become a limiting factor. Reuters notes that leading accelerators from Nvidia and AMD currently use HBM around their compute dies, with the reach of tiny electrical connections constraining how much memory can be placed in the fastest-access region.
Volantis wants to replace wires with light
Volantis is developing what it calls photonic wires: optical interconnects integrated into the accelerator rather than conventional electrical traces. The company says its optical waveguides can extend more than 200 millimetres on an interposer, compared with roughly 2–5 millimetres for electrical connections.
Its technology uses integrated micro-VCSELs — vertical-cavity surface-emitting lasers — to convert electrical signals into light. VCSELs are already a mature component category; Reuters notes that similar laser technology is used in facial-recognition systems in many Apple devices. Volantis is attempting to use that established manufacturing ecosystem for a much more demanding AI-computing application.
According to Volantis’ technical description, the architecture can connect more than 220 memory chiplets into a single uniform-latency memory pool. The company says its optical links can deliver more than 200 TB/s of aggregate bandwidth while consuming less than one picojoule per bit.
Those figures are Volantis’ claims about its architecture, not independently demonstrated performance from a shipping product. That distinction matters because the company’s first commercial system has not yet reached customers.
The A-1 system targets enormous AI models
Volantis’ first planned system, called A-1, is designed around the photonic memory architecture. The company says it is targeting models exceeding 20 trillion parameters and up to 10,000 tokens per second per user.
Volantis says A-1 is designed to combine roughly 10 TB of memory with 240 TB/s of memory bandwidth in a 15U form factor. Its stated goal is to make very large models responsive enough for workloads such as coding agents and other long-running inference tasks.
The company plans to deliver its first integrated inference engines to customers in 2027. That timeline means the immediate significance of the announcement is less about a new accelerator users can buy today and more about where AI hardware designers think the next performance ceiling is emerging.
Why optical interconnects matter for AI infrastructure
AI infrastructure has spent years focusing on faster processors, more HBM and increasingly sophisticated packaging. Photonics attacks a different part of the system: the distance and energy cost of moving information.
If optical links can maintain high bandwidth over longer physical distances without an equivalent increase in power consumption, memory does not have to be packed as tightly around a compute die. That could give accelerator designers more freedom to scale memory capacity while maintaining high access bandwidth.
It is also a potentially important strategy as AI inference becomes more demanding. Agentic systems can generate many sequential model calls, and large-context applications can require enormous amounts of model and working memory. In such workloads, latency and memory movement can become just as important as raw arithmetic throughput.
India should watch this hardware race closely
India’s AI infrastructure ambitions are increasingly tied to the availability of advanced compute, data-centre capacity and semiconductor supply chains. The country is building out capabilities across chip manufacturing, packaging and AI infrastructure, but those systems ultimately face the same physical constraints as AI infrastructure elsewhere.
That makes developments such as Volantis’ photonic approach relevant even before the technology reaches India. If optical compute-to-memory links eventually become commercially viable at scale, they could influence the cost, power requirements and physical design of future AI data centres — including infrastructure serving Indian enterprises and public-sector workloads.
A promising bet, not yet a proven replacement for HBM
Volantis is not eliminating HBM or replacing Nvidia and AMD overnight. Its proposition is to change how processors connect to memory and make much larger memory pools behave more like high-bandwidth on-chip resources.
The company has raised significant capital and assembled a team with experience from Nvidia, AMD, Broadcom, Ayar Labs and Micron. But the harder test will come when A-1 moves from architecture claims to production hardware, customer workloads, manufacturing yields and real-world power efficiency.
For the AI industry, however, the direction is revealing: as models become larger and inference becomes more continuous, the next major breakthrough may come not from making the processor itself faster, but from making the data feeding that processor move faster and farther.
Read the Reuters report and Volantis’ technical overview for additional detail.
