For the past few years, the AI industry’s central obsession has been buying enough GPUs. That constraint is quietly shifting. Lumilens, an optical data center connectivity startup founded by former Aruba Networks CTO Ankur Singla, emerged from stealth this week with more than $900 million in total funding and a $5.51 billion valuation — built entirely around a bet that the real bottleneck in AI infrastructure isn’t how many chips a company can buy, but how many of them it can actually connect to each other. It’s a strange kind of story for a wiring problem to become one of the biggest funding events in tech this year, but Lumilens is far from alone in treating it that way.
This article breaks down what AI interconnectivity actually means, why copper wiring has become AI’s real physical limit, how Lumilens fits into a rapidly consolidating race to replace it with light, and what this shift means for the future of AI infrastructure.
What Is AI Interconnectivity, and Why Does It Suddenly Matter?
AI interconnectivity refers to the networking technology that moves data between GPUs, accelerator chips, and servers inside an AI data center — the wiring, effectively, that lets thousands of individual processors function as a single coordinated system rather than isolated machines. It sounds like plumbing rather than innovation, but as AI models have grown large enough to require training and running across thousands of GPUs simultaneously, the speed and capacity of that wiring has become just as important to overall performance as the processing power of the chips themselves.
That shift explains why Lumilens’ own framing of the problem has resonated so quickly with investors: the constraint on AI has moved from how many GPUs a company can buy to how many of them it can actually connect, whether the goal is training frontier models or serving them to hundreds of millions of users simultaneously.
The Copper Wall: Why Wiring Became AI’s Real Bottleneck
The physics behind this shift is straightforward, if underappreciated. At the data transmission speeds modern AI chips operate at, an electrical signal traveling over copper wiring only holds its integrity for about a meter and a half before it degrades. That physical limit caps how many chips can be tightly linked together using copper alone at a few hundred — nowhere near enough for training runs that increasingly require tens of thousands of GPUs acting as a single system. Replacing copper with fiber-optic light-based connections removes that distance constraint almost entirely, which is why “optical interconnect” has become one of the most heavily funded categories in AI infrastructure over the past two years.
The scale of the wiring problem is staggering once you actually count the components involved. A single AI data center site with 400,000 GPUs requires more than 2.4 million optical transceivers just to keep those chips connected — a volume the existing optical components supply chain has struggled to keep pace with, even before accounting for the even denser chip-to-chip connections newer architectures increasingly require. Understanding the sheer physical scale involved helps explain why the infrastructure behind massive hyperscale data centers has become one of tech’s most consequential — and expensive — battlegrounds in its own right, independent of whatever AI model happens to be running on top of it.
Meet Lumilens: The Optical Startup Betting $900M on Light
Lumilens was founded in 2024 by Ankur Singla, a Cisco and F5 alum whose two previous networking startups — Contrail Systems and Volterra — were acquired by Juniper Networks and F5, respectively, for a combined $676 million. Just two years after founding Lumilens, the company is already commercially shipping its first optical interconnect products into production AI data centers under a multi-billion-dollar customer agreement, and its August 2026 stealth exit included a fresh $700 million-plus Series C round co-led by Atreides Management, Bain Capital Ventures, Meritech, Seligman Ventures, and Spark Capital, with participation from Qualcomm Ventures, J.P. Morgan Private Capital, and more than a dozen other investors.
The company’s technology platform, called LumiCore, spans in-house developed silicon photonics, mixed-signal integrated circuits, electrical-optical interposers, and full optical systems — each element, Lumilens says, designed for high-volume manufacturing from the start rather than retrofitted from lab-scale prototypes later. Its product roadmap covers both near-package optics, which are ready for deployment now, and co-packaged optics — light-based connections integrated directly onto the chip package itself — expected to reach market in two to three years. Lumilens is led by CTO Ted Schmidt, formerly a Distinguished Engineer at Juniper Networks who helped pioneer the company’s silicon photonics efforts following Juniper’s 2016 acquisition of photonics innovator Aurrion, giving the founding team a technical pedigree in optical networking that predates the current AI boom by nearly a decade.
Lumilens Isn’t Alone: The Crowded Race to Replace Copper
What makes Lumilens’ funding round notable isn’t that it’s solving a novel problem — it’s how crowded and well-capitalized the race to solve this exact problem has become. Ayar Labs, an MIT spinout backed by Nvidia, AMD, Intel, and the U.S. Department of Defense, has already shipped optical I/O chiplets and raised a $500 million Series E at a $3.75 billion valuation, with deployments underway at AWS’s custom Trainium chips. Celestial AI, whose Photonic Fabric technology directly competes with Ayar Labs at the chip-to-chip layer, was acquired by Marvell for roughly $3.25 billion in late 2025 — giving Marvell a combined custom-ASIC-and-optical-interconnect platform aimed squarely at Broadcom’s dominant position in the same market.
Nvidia’s own involvement in this space goes well beyond simply backing startups. The company has committed roughly $6.5 billion in photonics purchase agreements, with commitments to suppliers like Coherent and Lumentum reportedly significant enough to lock up much of the global supply of high-end laser components through 2027 — a move that gives Nvidia a real structural advantage over chip rivals racing to secure the same scarce optical components. That kind of aggressive infrastructure positioning fits a broader pattern of Nvidia making outsized strategic bets to secure its position across the AI hardware stack, extending its influence well past GPU design and into the physical components that determine how effectively those GPUs can actually be deployed at scale.
Elsewhere in the field, Lightmatter has reached a $4.4 billion valuation developing a 3D-stacked silicon photonics engine capable of 6.4 terabits per second of bandwidth in each direction, while Broadcom has taken a different architectural approach, integrating co-packaged optics directly into its Ethernet switch products rather than building standalone chiplets. Even legacy networking players are finding renewed relevance in this shift — Nokia’s own billion-dollar bet from Nvidia reflects a similar recognition that networking infrastructure, once treated as a commodity layer beneath the more exciting parts of AI, has become genuinely strategic ground worth defending.
Why Hyperscalers Are Betting Big on Optical Now
The urgency behind all this investment traces back to a specific engineering reality: memory and interconnect bandwidth have improved far more slowly than raw compute power over the past decade, by some estimates lagging 10 to 100 times behind. Since large AI models must be split across multiple GPUs that need to constantly exchange data during both training and live inference, that widening gap means adding more GPUs to a cluster increasingly fails to deliver proportional performance gains once data movement, rather than computation, becomes the actual limiting factor. That’s the specific inefficiency optical interconnects are designed to close, and it’s part of why silicon photonics technology can deliver meaningfully better throughput per watt of power consumed compared to traditional copper connections — a distinction that matters enormously for hyperscalers already managing enormous power and cooling budgets, a challenge covered in more depth in analyses of how data centers are working to manage their environmental footprint even as computing demand keeps climbing.
Competition among optical component suppliers for hyperscaler business often centers on winning pilot programs with companies like AWS, Google, and Microsoft — deals that, once successful, tend to translate into multi-year, multi-hundred-million-dollar production agreements as deployments scale from pilot to full data center rollout. That dynamic explains why so much venture capital has piled into this specific layer of AI infrastructure over the past two years: the winners of these hyperscaler contracts stand to capture recurring revenue tied directly to the continued expansion of AI compute itself, rather than depending on any single company’s model or product succeeding.
What This Means for the Future of AI Infrastructure
The emergence of Lumilens, layered on top of Ayar Labs’ production ramp, Marvell’s Celestial AI acquisition, and Nvidia’s multi-billion-dollar photonics purchasing commitments, signals that the industry has moved well past treating optical interconnects as a future research curiosity. It’s now considered foundational infrastructure, on the same order of strategic importance as the GPUs themselves. That shift mirrors a broader lesson about how AI’s true bottlenecks tend to move: for years, the constraint was compute capacity, driving the intense competition covered in stories like Nvidia’s continued dominance over rivals like Intel in AI graphics processing; now, with compute scaling reasonably well, the constraint has shifted one layer down, to the physical wiring holding all that compute together.
For an industry accustomed to talking about AI progress almost entirely in terms of model parameters and GPU counts, the sudden prominence of companies solving what amounts to an advanced wiring problem is a useful reminder that the physical infrastructure underneath AI is just as consequential to its future trajectory as the algorithms running on top of it — and, increasingly, just as lucrative a place for investors to place their bets.
Frequently Asked Questions
What is AI interconnectivity?
AI interconnectivity refers to the networking technology — increasingly optical rather than copper-based — that connects GPUs and accelerator chips inside AI data centers, allowing thousands of processors to function together as a single coordinated system for training and running AI models.
Why is copper wiring a problem for AI data centers?
At the high data transmission speeds AI chips require, electrical signals over copper degrade after about a meter and a half, capping how many chips can be tightly connected. That limit falls far short of the tens of thousands of GPUs modern AI training clusters increasingly require.
What is Lumilens and how much funding has it raised?
Lumilens is an optical data center connectivity startup founded in 2024 by former Aruba Networks CTO Ankur Singla. It emerged from stealth in August 2026 with more than $900 million in total funding, including a $700 million-plus Series C round, at a $5.51 billion valuation.
Who are Lumilens’ main competitors in optical interconnects?
Key competitors include Ayar Labs (backed by Nvidia, AMD, and Intel), Celestial AI (acquired by Marvell for roughly $3.25 billion), Lightmatter, and Broadcom, all developing competing approaches to replacing copper interconnects with optical technology inside AI data centers.
How much has Nvidia invested in optical interconnect technology?
Nvidia has committed roughly $6.5 billion in photonics purchase agreements with suppliers like Coherent and Lumentum, alongside direct investments in startups like Ayar Labs, as part of a broader strategy to secure its position across the AI hardware infrastructure stack.
What’s the difference between near-package and co-packaged optics?
Near-package optics place optical components close to a chip but not directly integrated onto its package, and are largely deployment-ready today. Co-packaged optics integrate optical components directly onto the chip package itself for even higher bandwidth, with most vendors expecting commercial availability within two to three years.
