Two Companies, One Bet on the AI Infrastructure Stack

You probably have not heard of Lumilens yet. You will.

In San Jose, California, roughly forty miles south of San Francisco, a two-year-old photonic interconnect company has been doing the kind of quiet work that tends to precede a loud arrival. It has drawn capital from two of Silicon Valley’s most disciplined infrastructure investors, Mayfield and Spark Capital. It has assembled an engineering team out of Cisco, Juniper, Meta, Lumentum, and Coherent — the institutions whose own transitions defined the last three decades of data center networking. And this week, after most of that work happening behind closed doors, Lumilens chose its first publicly announced partner: POET Technologies $POET

That choice is the story.

A Company Built by People Who Have Done This Before

Every major networking era reaches a point when the old architecture starts to strain and the next one becomes inevitable. Lumilens founder and CEO Ankur Singla has built a distinguished career out of recognizing those moments early.

At Aruba, later acquired by HPE, he helped shape the enterprise Wi-Fi wave before wireless became the default way businesses connected people and devices. At Juniper Networks, he worked on Ethernet fabrics as data centers became the backbone of the cloud. With Contrail, acquired by Juniper, he founded one of the defining software-defined networking companies of the cloud era. With Volterra, acquired by F5, he moved into distributed cloud services before enterprises had fully absorbed what the edge would become.

The pattern is hard to miss: wired to wireless, hardware to software, core to edge. Each was a structural change in how networks were built. Singla was early to each.

Now, with Lumilens, he is making the same kind of bet again — this time on the connectivity architecture required for large-scale GPU clusters and AI data centers. Lumilens is building a connectivity platform for AI infrastructure that accelerates the transition from electrical to optical connectivity inside the data center. AI infrastructure is forcing that transition into focus, pushing bandwidth, power, and distance requirements beyond what copper interconnects can comfortably support.

Lumilens may be a young company. But the insight behind it is not young-company intuition. It is the accumulated pattern recognition of someone who has helped drive several networking waves before they broke.

Smart Money, Moving Quietly

The capital structure tells the same story in a different language. Mayfield and Spark Capital are not transactional investors. They are infrastructure-focused venture firms with multi-decade track records of underwriting long, hard development cycles in semiconductors, networking, and systems software. Their presence on Lumilens' cap table signals conviction in a five-to-ten-year arc, not a quarterly narrative.

That conviction has translated into one of the fastest unicorn ascents the infrastructure sector has seen in recent memory. Lumilens crossed the $1 billion valuation mark in under a year of founding — an unusual pace in any environment, and a particularly rare one in deep-tech hardware, where long development cycles, capital intensity, and manufacturing risk typically slow the path to mature valuations. The trajectory says something about both the team and the moment: investors are not waiting for the photonics transition to become obvious. They are pricing it now.

Young company, seasoned operators, patient capital, and a valuation curve that suggests the market is already adjusting to the architectural shift underway. The combination tells you how some of the people closest to the photonics transition are thinking about its impact and opportunity.

The Physics That Forces the Shift

Copper interconnects move data by pushing electrons through metal. For short links, copper remains efficient, mature, and cost-effective. But at the speeds AI infrastructure now demands — 800 gigabits per second today, 1.6 terabits and beyond on the roadmap — copper’s useful reach begins to shrink. The issue is not simply bandwidth. It is distance, density, and scale.

Large AI clusters are no longer confined to a single rack. As model sizes grow, hyperscalers need GPU clusters to scale horizontally across multiple racks while still behaving like one tightly coupled system. That requires moving enormous volumes of data between GPUs, switches, servers, and racks with low latency and high reliability. Copper can support short, high-speed links, but as speeds rise, it becomes harder to extend those links across racks without signal integrity challenges, bulkier cable plants, and tighter physical constraints.

Photonics addresses the problem by changing the reach equation. Light moving through optical fiber can carry high-bandwidth signals farther across the data center without the same distance penalties that limit copper at higher speeds. Cable plants become lighter, thinner, and easier to route. Most importantly, optics allow AI clusters to scale beyond copper’s physical reach envelope, connecting multiple racks into larger GPU fabrics without forcing every critical link to stay within short copper distances.

This is not a marginal upgrade. It is an architectural requirement. The next generation of AI infrastructure depends on larger, denser, more distributed GPU clusters — and those clusters need interconnects that can span racks without compromising bandwidth or system design. The industry is not moving toward photonics because copper disappears. It is moving toward photonics because copper alone cannot support the reach and scale hyperscalers now require.

That is the demand-side pressure pulling capital, talent, and partnerships into the photonics stack — and the pressure Lumilens and POET were founded to address.

The Solution: A Rare Dual-Architecture Play

Lumilens describes itself as building “a connectivity platform for AI infrastructure.” That phrase does more work than it first appears.

The company designs, develops, and manufactures near-package optics, co-packaged optics, and pluggable optical transceivers, supported by its own silicon photonics, mixed-signal ICs, electrical-optical interposers, and optical systems. The breadth is the point. Lumilens is not betting on a single optical form factor. It is building across the interconnect stack as AI data centers move through multiple architectural phases at once.

Today, the dominant optical interconnect is the pluggable transceiver — a self-contained module that plugs into a switch faceplate or GPU tray. Pluggables remain essential to scale-out networks, where clusters expand across rows, racks, and data halls. They have served the industry well for more than a decade and will continue to matter. But as bandwidth climbs from 400G to 800G, 1.6T, and beyond, the limits of faceplate-based optics become harder to ignore.

That is pushing optics closer to the compute silicon. The first step is near-package optics, which brings optical engines closer to the GPU or switch ASIC while preserving some modularity. The next step is co-packaged optics, where optical engines sit directly alongside the ASIC package itself. Together, NPO and CPO represent the scale-up transition: the move toward tighter, denser, lower-latency connectivity inside the AI cluster.

Few companies in photonic interconnects can credibly play both sides of that transition — serving the scale-out architectures AI data centers rely on today while engineering the scale-up architectures they will require tomorrow. Lumilens is one of them.

That positioning is precisely why the $POET partnership makes strategic sense.

Why POET Is the Right Counterparty

POET Technologies has spent more than a decade developing its Optical Interposer and refining its manufacturability — a wafer-scale integration platform designed for the dense, co-packaged architectures the AI buildout now requires.

As Lumilens was advancing its own electrical-optical interposer, or EOI, it found in POET a partner that could accelerate its roadmap and strengthen execution against a technically demanding problem. Over the past year, that partnership has evolved into a highly manufacturable EOI solution that fast tracks Lumilens’ ability to deliver its unique silicon, mixed-signal ICs, and optical system IP into a tightly integrated platform for AI connectivity.

The result is not a simple component-level collaboration. It is a deeper effort to solve the packaging, system-design, and manufacturing challenges that emerge as optics move closer to GPUs and switch ASICs.

The pairing is strategically clean. POET extends its proven optical interposer and wafer-scale manufacturing, while Lumilens contributes the AI infrastructure connectivity roadmap, custom silicon, and systems architecture needed to turn that integration into products for scale-out and scale-up AI data centers.

This is the structural story underneath the headline. The integration challenge in photonics — combining silicon photonics, advanced packaging, electrical-optical interfaces, and systems-level engineering — is too broad for narrow point solutions in a market moving this quickly. Hyperscalers are evaluating combinations of platform technologies and systems specialists that can move faster together than either could alone.

The Market Beneath the Deal

The photonic interconnect market is valued at roughly $30 billion in 2026 and projected to exceed $100 billion by 2030, driven almost entirely by the AI data center buildout. Inside that market, the strategically important shift is the move from pluggable transceivers toward near-package and co-packaged optics that reside much closer to the GPU and switch ASIC. It is, in plain terms, the single most consequential transition in data center networking in a generation.

A market growing this fast is not an opportunity for any single winner. It is an opportunity for an ecosystem — and ecosystems are built through exactly the kind of complementary pairings that POET and Lumilens represent.

The question for the rest of the industry is no longer whether photonics will replace copper in the AI data center. That answer is settled. The question is which companies will build the platforms, partnerships, and manufacturing capacity to meet hyperscaler demand on the timeline the market requires. In San Jose, on a two-year-old unicorn backed by Mayfield, Spark Capital, and other venture investors, and now publicly aligned with a decade-old photonic platform on the Nasdaq, that question is starting to answer itself.