10 August 2026

^{Images: Lumilens}

 

For much of the artificial intelligence boom, access to graphics processing units has determined how quickly companies could expand their computing capacity. That constraint has not disappeared, but another is becoming increasingly important: moving data between all those processors.

Thousands of GPUs cannot operate efficiently as one computing system unless they exchange information quickly and continuously. As AI clusters grow, conventional electrical connections increasingly encounter physical limitations involving distance, bandwidth, heat and power consumption.

Silicon Valley startup Lumilens believes the answer is to use optical connections much closer to processors, while increasing the capacity of the optical networks already linking equipment across data centres. The company emerged from stealth on 6 August 2026 after raising more than $700 million in new financing, bringing its total funding to over $900 million and giving it a valuation of $5.51 billion.

“The constraint on AI has shifted from how many GPUs you can buy to how many you can connect,” founder and chief executive Ankur Singla said in the company’s announcement.

It is a familiar diagnosis in the semiconductor and networking industries. What distinguishes Lumilens’ strategy is the breadth of its proposed solution. Rather than concentrating exclusively on connections between data-centre racks or on optical links placed close to individual processors, it is building products for both.

 

Who founded Lumilens?

 

Founded in 2024 and headquartered in San Jose, California, Lumilens is the fourth infrastructure company led by Singla.

He previously founded software-defined networking company Contrail Systems, which Juniper Networks acquired in 2012 in a transaction reported at $176 million. Singla subsequently established distributed-cloud startup Volterra, acquired by F5 in 2021 for approximately $500 million. He also founded AI cybersecurity company Exaforce.

The wider Lumilens founding team combines networking, product development, semiconductor engineering and optical-manufacturing experience.

Chief technology officer and co-founder Ted Schmidt previously worked as a distinguished engineer at Juniper, where he helped develop its silicon-photonics and optical-integration capabilities following the company’s acquisition of Aurrion in 2016. He later held senior photonics positions at Lumentum and Effect Photonics.

The other founders are Ritesh Kapahi, Lumilens’ vice-president and general manager in India; product chief Samuel Liu; and operations chief Dave Friedman. Kapahi previously held engineering positions at companies including Cisco, Aruba and Juniper, while Liu’s career includes product-management roles in optical and networking technologies.

Lumilens says its broader team includes people recruited from Cisco, Juniper, Meta, Marvell, Lumentum and Coherent, reflecting that optical networking is not simply a chip-design challenge: products must combine photonic components, electronic control circuits, packaging, software, testing and high-volume manufacturing.

 

How Lumilens’ optical networking technology works

 

Inside an AI data centre, processors repeatedly send large quantities of information to one another. Copper connections carry this information as electrical signals and remain economical and effective over short distances. Their limitations become more pronounced, however, as data rates and distances increase.

Higher-speed electrical links can consume more power, generate more heat and suffer greater signal loss. Engineers can compensate with signal-conditioning technologies, but these add cost, complexity and energy consumption.

Optical connections work differently. An optical transceiver receives an electrical signal from a processor or network switch, converts it into pulses of light and sends those pulses through optical fibre. At the other end, another component converts the light back into an electrical signal that the receiving equipment can process.

Light can carry large volumes of data over longer distances with lower transmission losses than electrical signals. Silicon photonics makes it possible to produce many of the components used to generate, guide, modulate and detect that light using techniques derived from semiconductor manufacturing.

Lumilens is developing optical interconnect technology for two distinct parts of the AI data-centre network: scale-out and scale-up connectivity.

 

Scale-out networking between AI servers and racks

 

Scale-out networking connects servers, racks and computing clusters across a data centre. Optical fibre is already widely used for these longer connections, but expanding AI clusters require considerably more capacity at lower power per bit.

Lumilens is developing pluggable optical transceivers supporting data rates of 800 gigabits and 1.6 terabits per second, with faster generations planned. These modules can be inserted into compatible network equipment, making them relatively straightforward to install or replace.

The objective is not simply to replace every existing copper cable with fibre. It is to provide the bandwidth required for increasingly large AI computing clusters without an unsustainable rise in power consumption, heat and networking complexity.

Lumilens says its first scale-out product is already being shipped into the production data centres of an unnamed hyperscale cloud provider.

 

Scale-up networking between GPUs

 

Scale-up networking connects GPUs within a tightly coupled computing domain, allowing them to work together on the same AI workload. These connections may sit within a server or rack, although increasingly large scale-up systems can extend beyond a single rack.

Lumilens’ roadmap includes near-package optics, where optical components sit close to a processor, and co-packaged optics, where optical connectivity is integrated alongside computing or networking silicon.

Moving the optical interface closer to the GPU reduces the distance that data must travel electrically. In principle, this can increase bandwidth, reduce power consumption and allow larger numbers of processors to operate within one closely coordinated system.

Lumilens says its technology could initially connect thousands of processors and eventually tens of thousands. That larger target remains part of its development roadmap rather than a demonstrated commercial deployment.

 

What makes Lumilens different?

 

Several silicon-photonics companies concentrate on a particular section of the data pipeline. Some develop pluggable transceivers for connections between racks. Others are pursuing optical chiplets, near-package optics or co-packaged systems for links closer to GPUs and switches.

Lumilens is positioning itself differently by addressing both scale-out and scale-up AI networking through one technology and software platform. Its LumiCore architecture spans silicon photonics, mixed-signal integrated circuits, electrical-optical interposers and complete optical systems.

This does not mean one identical component performs every networking task. Scale-out transceivers and co-packaged scale-up systems have different technical and commercial requirements. The claimed advantage is that their underlying technologies are developed as parts of one integrated platform rather than as entirely separate product families.

Lumilens says this approach allows it to reuse designs, production processes, test systems and software across its portfolio. It could also help hyperscale customers move from today’s pluggable optical networks towards architectures in which optical connections sit much closer to processors.

The company develops technology across custom silicon, photonic components, mixed-signal electronics, interposers, systems and manufacturing automation. It also works with outside suppliers and production partners, meaning it does not manufacture every component independently.

This combination of scale-up and scale-out products with greater control over design, packaging and production is central to Lumilens’ pitch. Competitors may address some of the same technologies, but Lumilens is attempting to make the whole AI connectivity stack function as a coordinated platform.

 

Automating silicon-photonics manufacturing

 

Lumilens says it has designed proprietary robotics, production processes and automated testing equipment intended to make optical assembly faster and more repeatable.

Photonics manufacturing has historically been difficult to automate because optical components must be aligned with extreme precision. Small positioning errors can reduce performance or render a module unusable, making assembly, testing and production yield almost as important as the underlying photonic design. Lumilens is addressing this manufacturing bottleneck partly through a partnership with POET Technologies.

Announced in May 2026, the supply, joint-development and commercial agreement focuses on wafer-level optical integration. The companies’ planned electrical-optical interposer is intended to reduce or eliminate costly active-alignment steps during optical-engine production.

POET subsequently disclosed that the agreement included an initial $50 million purchase order and a framework targeting more than $500 million in cumulative revenue over five years. It cautioned that revenue remains dependent on successful development, qualification and manufacturing scale-up.

Lumilens’ attempt to address networking, packaging and production together may therefore prove to be its most significant differentiator. Designing a faster optical component is valuable, but hyperscale data centres need large numbers of components that can be manufactured reliably, tested quickly and deployed with acceptably low failure rates.

 

From stealth to production

 

Lumilens has reported unusually rapid commercial progress for a two-year-old photonics company. It says its first scale-out product is already being shipped into the production data centres of a hyperscale cloud provider.

The company describes the associated agreement as being worth multiple billions of dollars over several years. Neither the customer’s identity nor detailed shipment volumes have been disclosed, making it impossible to determine how much of that value represents firm orders, future options or deployments conditional on technical milestones.

Nonetheless, beginning production shipments and securing approval from a hyperscale customer would represent a significant milestone in an industry where product qualification can take years.

Its other milestones include the development of its LumiCore platform, the aforementioned POET Technologies partnership, and the establishment of manufacturing capabilities that Lumilens says can support volume production.

 

Lumilens raises more than $700 million

 

The latest Lumilens funding round was co-led by Atreides Management, Bain Capital Ventures, Meritech, Seligman Ventures and Spark Capital.

Other participants included Addition, Aiconic, Alkeon, EDBI, HarbourVest, J.P. Morgan Private Capital, Lingotto, Mayfield, MVP Ventures, Peak XV, Qualcomm Ventures, Redpoint Ventures, Seifdune and Thomvest Ventures.

The Series C brought the company’s total funding to more than $900 million and established a $5.51 billion valuation only two years after its founding.

Lumilens plans to use the capital to expand its engineering, software and manufacturing operations. This includes increasing production through manufacturing partners and its own facilities, developing its scale-up optical networking portfolio and accelerating future generations of scale-out transceivers.

 

A crowded race to connect AI data centres

 

Lumilens is entering a highly competitive market. Lightmatter, Ayar Labs and other startups are developing optical interconnect technologies for AI systems, while established semiconductor and networking companies are investing heavily in silicon photonics.

Marvell’s $3.25 billion acquisition of Celestial AI demonstrated both the value attached to the field and the likelihood of further consolidation. Nvidia, Broadcom and other major AI infrastructure suppliers are also expanding their optical networking capabilities.

Lumilens has therefore emerged at an opportune moment, but with much still to prove. Its valuation assumes that the company can convert an early hyperscaler deployment into sustained high-volume production while delivering its more ambitious scale-up roadmap.

The AI industry has spent enormous sums acquiring computational power. Its next infrastructure contest will help determine whether all that compute can communicate quickly enough to be used efficiently.

Lumilens is betting that solving both sides of that problem—and developing the manufacturing systems required to produce the solution at scale—will prove more defensible than supplying another isolated optical component.

 

Further reading on MoveTheNeedle.news:

Pilot Photonics secures €10.4 million EIC backing as Europe bets on photonic infrastructure

Oriole Networks wants to replace electronic switching with light. AMD is helping test it at AI scale

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