How Xscape Photonics plans to solve AI data centres’ networking bottleneck with silicon photonics
FalconX: the industry’s first fully redundant External Laser Small Form-factor Pluggable (ELSFP) device capable of emitting up to eight wavelengths or colors of light for ultra-fast, high-capacity and low-power optical data transmission.
Artificial intelligence companies are racing to build larger and more capable models, but a growing number of infrastructure engineers argue the real bottleneck is no longer the chip itself. Instead, the challenge is becoming how fast data can move between thousands of processors without networking delays and rising power consumption overwhelming the system. That problem sits at the centre of Xscape Photonics’ strategy. In an exclusive interview with MoveTheNeedle.news, Co-Founder and CEO Vivek Ragunathan discussed how the company plans to use multi-wavelength optical networking and silicon photonics to address what it describes as AI’s “escape bandwidth” problem.
The conversation with Ragunathan followed Xscape Photonics’ March 2026 announcement of a $37 million extension to its Series A funding round, and the launch of FalconX, a new optical networking module capable of generating up to eight wavelengths, or colours, of light from a single device. The company says the platform is designed to address what it calls the “escape bandwidth” problem increasingly affecting AI data centres as clusters grow larger and more power intensive.
AI infrastructure is becoming communication-limited
Much of the AI industry’s attention has focused on processors themselves, particularly the race to build faster GPUs and specialised accelerators. Increasingly, however, infrastructure providers are confronting another constraint: moving vast amounts of data between processors and memory quickly enough to keep systems fully utilised.
Modern large language models and reasoning-based AI systems require constant communication between GPUs and memory systems across increasingly large AI data centres. As those clusters scale, networking delays and bandwidth limitations increasingly affect overall performance.
“GPUs are data centres’ most expensive components, but often sit unused or underutilised while waiting for data,” Ragunathan said. “Running modern workload data on current physical hardware is like a traffic jam of high-performance sports cars on a dirt road.”
Xscape Photonics was founded specifically around that problem: on the realization that scaling high-performance computing and AI hardware depends on a high-speed communication fabric, as Ragunathan put it. “A lot of incumbent solutions continue to be limited in their ability to deliver high-speed data because of a lack of scalable laser solutions.”
Xscape Photonics is different: the company emerged from research breakthroughs achieved at Columbia University in comb-based laser technology, which can generate multiple wavelengths of light from a single chip. Ragunathan co-founded the company alongside fellow Columbia photonics researchers Alexander Gaeta, Yoshi Okawachi, Keren Bergman and Michal Lipson.
The broader shift in infrastructure priorities Ragunathan pointed out to us was also reinforced in a blog post published by Xscape Photonics earlier this week. In the post, the company argues that AI inference systems are increasingly becoming “communication-limited”, meaning networking bandwidth rather than compute performance determines overall efficiency.
According to the company, reasoning-capable AI agents require longer context windows and more intensive memory communication, dramatically increasing data traffic inside inference clusters. Xscape Photonics argues that this growing communication load causes performance-per-watt efficiency to deteriorate as systems scale.
The blog also introduces what the company describes as a new AI infrastructure metric: “tokens-per-second-per-megawatt” (TPS/MW), effectively measuring how efficiently an inference cluster generates AI output relative to energy consumption.
Xscape Photonics further argues that inefficient data movement contributes significantly to the widening energy gap between biological and artificial intelligence systems. The company points to estimates suggesting the human brain performs complex reasoning tasks using roughly 20 watts of power, while large AI inference systems handling comparable reasoning workloads can consume close to one megawatt.
Whether those comparisons remain scientifically precise across different workloads remains open to debate, but the broader point reflects growing industry concern about the energy intensity of large-scale AI deployment.
What “escape bandwidth” means for AI infrastructure
Xscape Photonics frequently refers to “escape bandwidth” as one of the defining bottlenecks facing AI infrastructure.
“When we talk about ‘escape bandwidth,’ we’re referring to how much data can physically get in and out of a GPU and memory in a data centre network,” Ragunathan elaborated. “It doesn’t matter how fast your processors are if only a limited amount of data can ‘escape’ them to communicate with the rest of the cluster.”
The consequences extend beyond slower training times. Bottlenecks in networking infrastructure can leave GPUs idle, increase power consumption and reduce throughput for AI inference systems.
The challenge has intensified rapidly. According to Xscape Photonics, AI clusters have grown more than tenfold in size over the past two years. Conventional copper-based interconnects increasingly struggle to keep pace with the bandwidth and power demands associated with those deployments.
The company’s blog expands on that argument further, warning that networking infrastructure itself is becoming a major contributor to data centre energy consumption. As clusters grow larger, the optical links, switching systems and networking hardware required to support them also consume increasing amounts of power.
In other words, the infrastructure challenge facing AI companies is no longer simply about building faster processors. It is also about connecting growing numbers of processors efficiently, reliably and sustainably.
Replacing copper interconnects with silicon photonics
The AI infrastructure industry has already been moving steadily towards optical interconnects. Companies such as Broadcom, Marvell Technology and Coherent have all expanded silicon photonics and optical networking efforts in response to AI demand.
Xscape Photonics is attempting to differentiate itself through multi-wavelength optical communication.
Instead of relying on electrical signals travelling through copper interconnects, the company uses multiple wavelengths of light transmitted across optical fibre. Its proprietary CombX technology generates several colours of light simultaneously from a single silicon photonics chip. FalconX currently supports up to eight wavelengths.
“Our approach is to present an alternative to these outdated interconnects, in the form of optical communication via silicon photonics,” Ragunathan said. “Using multi-wavelength lasers, data can be transmitted as light instead of electrons, which can dramatically increase the amount of data that can be rapidly moved through a single connection while also using less energy in the process.”
The underlying principle is wavelength-division multiplexing (WDM), a technology already widely used in telecommunications infrastructure. WDM allows multiple wavelengths of light to travel simultaneously through a single optical fibre, dramatically increasing bandwidth capacity.
Xscape Photonics is effectively attempting to bring those principles deeper into AI data centre architectures themselves.
The company says FalconX can deliver more than one watt of optical power and support multi-terabit-per-second bandwidth from a single pluggable module.
Reliability becomes central as AI clusters scale
As AI clusters become larger and more expensive, reliability is becoming as important as raw performance. Even relatively small hardware failures can create significant disruption when thousands of GPUs are linked together across large inference and training environments.
“In large AI clusters, a single optical transceiver failure due to laser degradation can disrupt links and interrupt workload execution, resulting in costly downtime,” Ragunathan said.
In the belief that reliability can no longer be treated as a secondary infrastructure concern, Xscape Photonics designed FalconX with built-in redundancy that continuously monitors laser degradation and can automatically switch to spare wavelengths without taking the connection offline. The company says the approach aligns with emerging industry initiatives such as the OCA-MSA standard focused on improving reliability for optical infrastructure.
The growing focus on resilience reflects a broader shift happening across AI infrastructure. As GPU clusters become larger and more costly to operate, networking reliability increasingly affects the economics of AI itself. Downtime, latency and underutilised processors now carry direct financial consequences for hyperscalers and cloud providers investing billions into AI infrastructure.
That shift also helps explain why companies such as NVIDIA, an investor in Xscape Photonics, have expanded investments beyond processors themselves into networking, memory, cooling and photonics technologies.
“Strategic investors and partners are critical in providing useful insights about our industry problem statement and product requirements, and valuable guidance on our ecosystem partnerships,” Ragunathan said.
The race towards 128 wavelengths
FalconX represents only one step in Xscape Photonics’ broader roadmap.
The company’s ChromX platform aims to scale beyond eight wavelengths towards systems capable of generating 16, 32 and eventually more than 128 colours of light. In August 2025, Xscape Photonics demonstrated a 16-colour CombX prototype in collaboration with Tower Semiconductor.
“Current performance scaling at cluster level is directly proportional to the bandwidth of communication between GPUs and memories,” Ragunathan said. “More colours translates to more bandwidth in a fibre, thereby translating to significantly increased token throughput at an application level.”
The broader industry trend is becoming increasingly clear: networking is no longer peripheral infrastructure in AI systems. It is becoming a determining factor in how efficiently those systems scale.
Photonics alone will not remove every infrastructure limitation. Optical systems remain technically complex and expensive to manufacture at scale, while integrating photonic technologies into existing data centre architectures presents its own engineering and operational challenges.
Still, investors increasingly appear willing to bet that photonics and optical networking will become foundational technologies for the next generation of AI infrastructure.
If Xscape Photonics succeeds in commercialising its roadmap, AI data centres could eventually operate less as rigid collections of discrete GPUs and more as flexible pools of compute and memory connected through high-bandwidth optical fabrics.
“Xscape Photonics will unlock a completely new fabric of connectivity, in which pools of GPUs will talk to pools of memories over a big, fast, multi-colour network,” Ragunathan stated.
Further reading on MoveTheNeedle.news:
Can Europe Build the AI Infrastructure It Needs? Inside Polarise’s High-Density Munich Data Centre