Key Takeaways
- Lumentum’s 2026 roadmap shows hyperscale AI driving 800G and 1.6T optical module volumes past 10 million units.
- The move to co-packaged optics (CPO) is picking up speed, with first deployments in AI accelerators expected by late 2026 to deliver major power savings.
- Silicon photonics is set to become the standard for high-speed optics, making up over 70% of new AI cluster deployments because it scales so well.
- Even with big optical advances, managing heat in dense AI server racks remains a top-tier engineering challenge that demands new cooling solutions.
- The market for optical transceivers in AI data centers is on track to hit $4 billion by 2028, growing at a 35% compound annual rate from 2023.
The amount of data moving inside AI hardware is exploding, with some projections showing that intra-data center traffic will swamp traditional internet traffic by 3x within two years. This firehose of data absolutely requires a new class of optical interconnects. Lumentum’s 2026 roadmap gives us a glimpse of how this tech will prop up the next generation of AI, but it also exposes the immense engineering challenges we have to solve to make it a reality.
800G and 1.6T Modules: The New Baseline for Hyperscale
Lumentum’s internal analysis predicts that deployments of 800G and 1.6T optical interconnect modules will blow past 10 million units in hyperscale AI data centers by the end of 2026. This is a foundational reset of how data moves between GPUs and other accelerators in an AI cluster. While data center interconnects used to evolve on a predictable schedule, the raw compute hunger of large language models has thrown that timeline out the window. We’re now designing purpose-built AI engines that require instantaneous, fat pipes for communication.
Data center architects I talk to all confirm this path. When you’re training models with billions of parameters, the latency budget is so tight that every nanosecond you can shave off inter-processor communication goes straight to the bottom line with faster training and deployment. The benefits of moving to 800G and 1.6T are also about reducing the number of fibers and ports, which simplifies the network topology and cuts the power-per-bit. This consolidation is absolutely critical as data centers hit physical walls on space and strain local power grids.
Co-Packaged Optics (CPO): Breaking the Electrical Barrier
The Lumentum roadmap is calling for the first real deployments of co-packaged optics (CPO) in specialized AI accelerators around late 2026, with a serious ramp-up in 2027. For a long time, the industry has been stuck fighting the physics of electrical traces on circuit boards. As you push data rates higher, the signal degradation, power draw, and sheer physical space of copper routing just become untenable. CPO gets around this by moving the optical engine right next to the processor, sometimes on the very same package.
Reports from analysts at LightCounting Market Research consistently point to CPO as the way to get past this “electrical wall.” By shrinking the distance an electrical signal travels before it becomes light, CPO promises to cut interconnect power by as much as 50% while also slashing latency. This power efficiency is non-negotiable. AI data centers already draw staggering amounts of electricity, and without technologies like CPO, their growth will be choked off by grid capacity and opex. The manufacturing challenges are real (especially with heat and yield), but the performance gains for AI are too big to pass up. For the most intense AI jobs, this signals the end of the line for discrete pluggable optics.
Silicon Photonics Dominance: Scalability and Integration
Lumentum’s strategy puts silicon photonics front and center, projecting it will make up over 70% of new deployments in AI clusters. It leverages existing silicon manufacturing know-how (with some specialized steps, of course) to mass produce optics with a cost structure and volume that traditional materials like indium phosphide simply can’t match for these AI deployments.
Countless papers from the Institute of Electrical and Electronics Engineers (IEEE) have chronicled the march of silicon photonics, detailing how you can etch incredibly complex optical circuits, lasers, modulators, detectors, right onto a silicon wafer. This integration lets you build dense optical I/O at a scale that’s needed for AI hardware, where a single server rack might have thousands of optical links. Packing those components tightly and cheaply is the whole game. Any vendor that isn’t running full-speed-ahead with a silicon photonics strategy for their high-speed portfolio is going to get lapped in the AI market.
Thermal Management: The Silent Killer of Performance
For all the exciting work in optical interconnects, the biggest and most boring roadblock remains the same: thermal management. As you pack more optical functions into smaller modules, the heat density goes through the roof. Lumentum’s own engineers know that getting heat out of the package is a primary bottleneck to hitting performance and reliability targets in dense AI systems. Just think about it, a rack full of GPUs each burning hundreds of watts, and now you’re adding dozens of power-hungry optical modules right next to them. The total heat load is just enormous.
While everyone obsesses over electrical power consumption, the thermal consequences are just as critical for long-term system stability and the lifespan of the components. That’s why advanced cooling, from direct-to-chip liquid loops inside the rack to new heat sink designs for CPO, is now integral to the system architecture. The industry has to get past air cooling for these applications. We need to see serious, widespread adoption of direct liquid and even immersion cooling to handle the heat from next-gen AI hardware and its optical links. Ignoring the heat is just designing systems that are guaranteed to throttle and fail early.
Market Growth: A $4 Billion Opportunity
The total market for optical transceivers in AI data centers is forecast to hit $4 billion by 2028, a 35% compound annual growth rate from 2023. This projection, which firms like Dell’Oro Group generally agree with, shows just how much money is on the table. This growth is the fuel for an entire AI hardware stack. Every hyperscaler is pouring billions into new infrastructure, and these optical links are the arteries.
That $4 billion figure doesn’t even tell the whole story, because it hides the rising complexity and value inside these modules. Unit volumes are climbing, but the average selling price (ASP) is also jumping because of higher speeds, on-board digital signal processors (DSPs), and the simple fact that they enable billion-dollar AI workloads. This market explosion forces companies like Lumentum to innovate at a punishing pace just to secure their position. In this field, delaying the next generation of optical solutions by even one product cycle means forfeiting a huge slice of market share.
So, the path forward for optics in AI is all about higher speeds, tighter integration, and a ruthless focus on power and heat. Lumentum’s 2026 roadmap isn’t just a slide deck. It’s a set of engineering targets that will reshape AI infrastructure. For anyone involved in AI hardware or data center design, the message is blunt: get your optical interconnect strategy right and start planning for co-packaged optics now, or you’re already behind.
What are optical interconnects?
Using light to transmit data through fiber optic cables, optical interconnects replace electrical signals on copper wire. This approach delivers much higher bandwidth and lower latency with less power, especially over distance, making it a requirement for moving massive datasets in AI data centers.
Why are 800G and 1.6T modules important for AI?
AI models like LLMs need to shuffle huge amounts of data between processors like GPUs with minimal delay. 800G and 1.6T modules deliver the raw bandwidth to keep these processors fed with data, which prevents bottlenecks, shortens model training times, and makes the entire AI cluster run more efficiently.
What is co-packaged optics (CPO) and how does it benefit AI?
Co-packaged optics, or CPO, puts the optical I/O right on the same package as the main processor (like a GPU). By drastically shortening the distance electrical signals have to run, it cuts power consumption and latency, which are two of the biggest constraints in designing dense, powerful AI hardware.
Why is thermal management a significant challenge for optical interconnects in AI?
Faster, denser optical interconnects generate a lot of heat in a very small space. Inside an AI server that’s already packed with hot, power-hungry components, this extra heat degrades performance and threatens the reliability of the whole system. Proper thermal management, which increasingly means liquid cooling, is essential to keep these systems from cooking themselves.
What role does silicon photonics play in next-generation AI interconnects?
Silicon photonics uses standard silicon manufacturing techniques to build integrated optical components. It’s the key to mass-producing the small, power-efficient, and high-bandwidth optical transceivers that AI clusters require, offering a path to scale production and lower costs that isn’t possible with older optical technologies.
