Copper has hit a hard wall inside the rack, and optical transceivers are already in short supply. The next phase of AI depends on scale-up networks that couple thousands of GPUs into a single machine — and the lowest-risk way to build them is near-package optics (NPO) first, co-packaged optics (CPO) next, manufactured at a volume the optics industry has never delivered. Here's why we built Lumilens™ to be first down that path, and how we put a product into production AI data centers just two years after founding.
The AI problem nobody talked about…until now
For the last several years, the AI infrastructure conversation has been almost entirely about GPUs — how many you can get, how fast you can deploy them, how you cluster them…whether into one giant training run or across the fleets of clusters that inference and agentic workloads now demand. GPUs are still scarce — that hasn't changed. What has changed is that compute is no longer the only constraint that matters.
Modern large-scale AI workloads — frontier training and high-volume inference alike — require many thousands of processors to behave as a single, coherent machine. The performance of that machine is now set as much by the network stitching the compute chips together as by the chips themselves. And that network is running into two walls at once.
The first wall is outside the rack, in the scale-out fabric that ties racks and rows together — a supply problem, and a serious one, which we'll come back to. The second is inside the rack, in the scale-up fabric that lets GPUs share memory and exchange gradients (the partial results of learning that every chip must merge with every other after each training step, billions of values at a time) at full bandwidth, and this one is physics. And the two walls share a trait the industry keeps underestimating: both are manufacturing problems as much as photonics problems — won on manufacturing floors, not just in labs. This post is mostly about the second wall, because it is where the next phase of AI will be won or lost, and it is where Lumilens intends to lead - in the photonics and in manufacturing.
Copper's last meter
Today's scale-up fabrics run over copper. At the signaling rates AI now demands, electrical signals survive roughly a meter and a half before attenuation makes them unusable. That puts a hard ceiling on how many GPUs can be tightly coupled into one domain — a single rack's worth, a few hundred GPUs at most. When a hyperscaler's roadmap calls for thousands of GPUs behaving as one logical machine, copper-based scale-up fabrics simply cannot get there. Reach, power, and bandwidth density all give out at once.
The answer has to be photonics. Light doesn't care about a meter in distance; it barely notices a hundred. But "the answer is photonics" is where the real question begins — because scale-up optics sit in the most valuable and least forgiving location in the entire data center, millimeters from the most expensive silicon ever mass-produced. How you bring light to the GPU matters as much as whether you do.
The scale-out squeeze
Meanwhile, the first wall keeps closing in. Every processor added to a cluster multiplies the optical transceivers required to scale it out: a 400,000-GPU data center needs more than 2.4 million transceivers and over five million individual fiber strands. The optics industry cannot keep pace with that volume — independent analysis projects that 800G transceiver production will fall 40–60% short of demand through 2027, with 1.6T shortfalls of 30–40% persisting into 2029.¹
Now consider what scale-up does to that math. Moving GPU-to-GPU traffic onto light multiplies the optical content attached to every accelerator well beyond today's scale-out ratios. The industry is about to ask far more of a supply chain that is already failing to deliver. Whoever solves scale-up optics has to solve manufacturing at the same time — a point we'll return to, because it shaped Lumilens as much as any photonics decision did.
NPO first, CPO next: the lowest-risk path to optical scale-up
Why can't pluggable transceivers carry scale-up too? At today's 100G-per-lane, the trace from ASIC to front-panel pluggable crosses hundreds of mm of host board — lossy enough to require a retiming DSP that burns roughly half the module's power and adds more than 15% to its cost. At 200G-per-lane and beyond, that channel doesn't close at all. The optics must move toward the silicon.
Near-package optics (NPO) places the optical engine in a socket beside the processor, cutting the channel to tens of mm of substrate — short enough to eliminate the retiming DSPs, recovering most of that power at lower cost. It keeps what operators won't give up: a failed engine is a socketed module you reseat, not a package you scrap; lasers stay internal to the module, with no external-laser (EL-SFP) cost; and socket-level compatibility keeps the supply chain multi-vendor.
Co-packaged optics (CPO) integrates the engine into or beside the processor package — the end state, the best power and density physics allows. But every picojoule saved carries concentrated risk: optics yield couples to the most expensive silicon in the system, a failure can strand a whole package, serviceability shrinks, and the architecture tends toward single-vendor lock-in — all before fleet-scale reliability has been proven.
We did not arrive at this conviction secondhand. We learned this lesson during our founding team's decade-long journey on silicon photonics and CPO at Juniper — close enough to the silicon to know what optical integration truly demands, and long enough to learn that packaging and manufacturing decisions are risk decisions. That experience is the origin of the Lumilens roadmap.
The industry will quietly reach the same conclusion. The leading next-generation accelerator programs will pair copper inside the rack with NPO across racks, and operators will run NPO programs alongside their CPO efforts precisely to keep sourcing multi-vendor while the risk retires. Industry analysts have eventually reached the same conclusion: the near-term center of gravity for next-generation optical interconnects will shift to NPO, with CPO projected to take the majority of optical-engine shipments by the end of the decade.
NPO now, CPO next is not a hedge — it is the lowest-risk and also the lowest-cost path to optical scale-up. You qualify light beside the package, at fleet scale, with serviceability and second sources intact; then you move it inside the package once the reliability data has earned it. It is the path operators will choose, and it is the path Lumilens was built to lead.
Built to be first
Lumilens intends to be first to market with production NPO for GPU scale-up at hyperscalers’ level of deployment, not lab demonstrations.
First, the platform was designed for this from day one. LumiCore™ spans pluggables, NPO, and CPO on one photonic and electronic foundation — which means that for us, CPO is a packaging evolution of an already-qualified platform, not a new platform. The NPO-then-CPO sequence that de-risks the transition for operators is on LumiCore, a single continuous roadmap.
Second — and this is the part that can't be shortcut — NPO (and CPO) readiness requires manufacturing readiness. The companies racing toward scale-up optics today are mostly silicon vendors bundling optics they've never built at volume, or photonics startups that have never shipped through a hyperscaler qualification.
LumiCore is a generational overlap: the same platform headed into NPO sockets is already qualified and shipping into production AI data centers today with pluggables, in high volume, through the industry's hardest qualification gauntlet with one of the largest hyperscalers in the world. A customer who has qualified LumiCore for a pluggable use case doesn't start over for NPO — and neither do we.
Meet Lumilens and its LumiCore photonics platform
Most optical interconnect vendors build for either scale-out or scale-up. Lumilens is innovating across both using a single, unified technology platform called LumiCore, spanning four integrated layers:
- Silicon photonics: The foundational layer converting between electrical and optical signals, built in-house to control integration density and performance at scale.
- Mixed-signal ICs: The analog and digital interface circuitry driving the photonic layer — custom-designed for the signal-integrity constraints of high-density AI interconnects.
- Electrical-optical interposers: Lumilens' proprietary interposer technology dramatically simplifies photonic chip assembly — one of the historically hard manufacturing problems in silicon photonics — and is the foundation of the company's ability to scale production rapidly.
- Optical systems: Full system integration and qualification, from chip to module to customer-facing product.
.png)
As a unified platform, LumiCore compounds. Design iterations propagate – a process improvement in the silicon photonics layer simultaneously benefits the pluggable transceiver shipping today and the NPO engine that will follow it. And when a customer initially qualifies the platform, expanding into adjacent products doesn't restart the clock. The practical result: Lumilens moved from company founding to a customer-qualified pluggable product shipping into production AI data centers to an NPO tapeout, all within two years.
.png)
At Lumilens, manufacturing is a first-class citizen
Having a highly integrated technology platform is one thing — leveraging it to build millions of units annually against massively growing hyperscaler demand is another thing entirely. That's why Lumilens treats manufacturing as a product discipline in its own right.
.png)
The company's proprietary electrical-optical interposer simplifies the most failure-prone step in photonic assembly. Lumilens has developed its own process recipes, custom robotics for assembly steps, and test automation designed to meet the stringent requirements of the AI data centers.
Lumilens' production model combines this in-house design, robotics, and test capability with top-tier contract manufacturing partners for established processes, plus Lumilens-owned facilities for the steps where proprietary process control matters most. The goal is to decouple innovation velocity from production scaling — so a new generation of silicon photonics enters production without displacing capacity committed to current products.
For hyperscaler customers, this matters as much as the specs. A transceiver with better power efficiency at 1.6T means nothing if you can't get a few million of them delivered on a predictable schedule — and an NPO engine that halves the power means nothing if it can't be built, tested, and serviced at fleet scale. Manufacturing discipline is not adjacent to the NPO-first thesis; it is the thesis.
What's shipping and what's next
Lumilens' initial product — a scale-out pluggable optical transceiver — completed qualification and began shipping into a top hyperscaler's production AI data centers earlier this year under a multi-billion-dollar supply agreement. In parallel, the scale-up program is in active development: NPO engines for direct GPU-to-GPU and GPU-to-memory optical interconnect, targeting clusters that aggregate many thousands of processors, well beyond what copper fabrics can support — with CPO following on the same LumiCore platform as the reliability case matures. The roadmap extends to tens of thousands of GPUs operating as a single logical domain.
The Lumilens team
The team building this portfolio brings hundreds of years of combined experience in silicon design, photonic integration, mixed-signal IC design, systems engineering, process engineering, and high-volume manufacturing. It is drawn from Cisco, Juniper Networks, Meta, Marvell, Lumentum, Coherent, and other top-tier vendors.
Leading Lumilens is CEO Ankur Singla — this is the fourth startup he has founded. His first two, Contrail Systems and Volterra, were acquired by Juniper Networks and F5, respectively. CTO and co-founder Ted Schmidt was Sr. Distinguished Engineer at Juniper and led its silicon photonics, CPO, and optical integration following the 2016 acquisition of Aurrion. VP of Products and co-founder Samuel Liu brings a deep optical and signal processing background, having helped Marvell build and scale key optical networking product lines.
This depth of experience across optical technologies, silicon design, physical systems, and high-volume manufacturing is essential for a company entering the photonics market.
The road ahead
The bottleneck in AI has moved from compute and memory chips to the network. The network, in turn, is moving to light — first beside the package, then inside it. That sequence — NPO now, CPO next — is the lowest-risk path for the operators betting hundreds of billions on their next generation of AI infrastructure. Lumilens has built the optical connectivity platform, the manufacturing excellence, and the innovation team to help them succeed in every step of this sequence.
Two years in, we're shipping optical interconnects where the AI demand is most acute, while building toward where the industry is headed. If you're an infrastructure team confronting the scale-up wall — or an engineer who wants to help tear it down — we'd like to talk.
