Lumentum’s big bet on die-to-die optical interconnects is set to completely re-architect how high-performance AI computing actually works. This is about more than a simple strategy change. By integrating photonics directly onto silicon dies, the company is aiming to blow past the bandwidth and power consumption ceilings that are throttling today’s AI accelerators. It’s a huge gamble, but will it lock in Lumentum’s position in the AI infrastructure market for the next decade?
Key Takeaways
- Lumentum’s optical interconnects are meant to fix the bandwidth and power problems of old-school electrical connections inside AI data centers.
- Putting photonics directly on the die cuts latency and power use, making AI model training and inference way more efficient.
- To keep scaling AI superclusters, the industry has to move to co-packaged optics and then die-to-die solutions because electrical I/O just can’t keep up.
- Getting the whole industry on board will require solid standards and massive manufacturing capacity, two areas where Lumentum is spending heavily to get the market ready.
- The future of AI hardware depends on a total overhaul of interconnects, and it looks like optical is going to be the default for any serious processing unit.
Opinion: Lumentum’s heavy investment in die-to-die optical interconnects is an absolutely necessary move, positioning them as a key supplier for the next generation of AI computing infrastructure.
The Shift from Electrical to Optical at the Chip Level
The insane amount of data that large AI models need to process has slammed traditional electrical interconnects right into a wall. We’re talking terabits per second flying between GPUs and TPUs in a single rack, never mind an entire datacenter. Over even short distances, electrical signals degrade, suffer from impedance issues, and burn a ton of power as heat. This is a hard physical barrier to scaling performance. In fact, a 2025 Omdia report projects that if we stick with electrical interconnects, they’ll be eating up over 30% of an AI datacenter’s total system power by 2028. That kind of power draw just isn’t sustainable.
Lumentum’s die-to-die optical interconnects tackle this problem head-on. By moving light signals as close as possible to the processors, they get rid of the bottlenecks you see in copper traces. Imagine data zipping between cores on a wafer at light speed with almost no energy loss. This is the future Lumentum is building right now. Their plan leans on deep expertise in indium phosphide (InP) photonics, a material that’s perfect for cramming high-speed optical components into tiny optical engines that can sit right next to an AI processor die. This tight integration is what separates die-to-die from earlier co-packaged optics (CPO), which still kept the optical engines on a separate piece of substrate, and it’s a huge jump in density and efficiency.
I’ve seen datacenter architects tear their hair out trying to get past the 1.6 terabit Ethernet barrier with copper. The cable bulk, the power draw, the cooling, it all becomes a nightmare. Die-to-die optics present an entirely new path forward by changing the physics of data transfer right where it matters most: at the chip interface. This sea change will redefine the physical limits of AI supercomputing.
Overcoming Bandwidth and Power Limits for AI Accelerators
Today’s AI accelerators, like NVIDIA’s H100 or Google’s TPUs, are incredibly fast, but they’re often starved for data because they can’t talk to each other or to memory quickly enough. High-bandwidth memory (HBM) has helped with the memory access part of the problem, but communication between different chips is still a major bottleneck. When you’re training a trillion-parameter model, every nanosecond and every milliwatt adds up. Electrical traces just can’t keep a clean signal at the multi-terabit speeds AI workloads require, which forces engineers to add complex equalization circuits that burn even more power and add latency. This is exactly the problem Lumentum’s optical interconnects are built to solve.
By replacing electrical traces with optical waveguides inside the chip package, Lumentum is promising massive improvements in bandwidth density and energy efficiency. A single optical fiber carries way more data than a bundle of copper wires, and light is immune to electromagnetic interference, which makes board design simpler and more reliable. The power needed to drive a light signal a few millimeters inside a package is a fraction of what it takes to push an electrical signal at the same speed. Projections from the 2025 Optical Fiber Communication Conference (OFC) suggest that optical interconnects could slash communication energy use by up to 70% in future AI clusters compared to today’s electrical links. That kind of power reduction directly cuts operating costs and allows for much denser server designs.
And think about the heat. Less power wasted as heat means you don’t need as much complex, expensive cooling hardware, which is a huge part of the bill for any hyperscale AI deployment. This lets you pack more compute into the same rack space, speeding up everything. Lumentum’s real competitive advantage here is its ability to deliver integrated laser sources and photodetectors in a tiny, reliable package, something they’ve gotten very good at after decades of making optical components.
Challenges and Lumentum’s Path to Dominance
The benefits of die-to-die optics are obvious, but getting to mass adoption won’t be easy. There are huge hurdles in manufacturing, ensuring reliability in a hot datacenter, and getting the cost down to compete with copper. Integrating photonic components with standard CMOS fabrication is notoriously difficult and requires specialized techniques and a ton of testing. On top of that, a lack of universal standards for on-chip optical interfaces could splinter the market and slow everything down. This whole effort requires serious capital and deep engineering talent.
Lumentum, however, is probably in the best position to pull this off. They’re already a leader in high-volume optical component manufacturing for the telecom and datacom markets, so they have the foundation for scaling up. They’re also actively involved in the industry groups working to standardize these new optical interfaces, and their partnerships with major AI chip designers and cloud providers show they’re building what the market actually needs. For example, their recent announcements on silicon photonics integration point to a clear plan for making this technology affordable. That kind of work on standardization is often ignored, but it’s absolutely necessary for a new technology to take hold.
You could argue that better electrical signaling, like new SerDes tech, might postpone the switch to optics. And while electrical links will get better, they are in the end running up against the laws of physics. There’s a hard physical ceiling. Photons just don’t have the same problems with signal degradation over distance. The power and bandwidth gains from die-to-die optics are so large that for high-end AI, they’ll become unavoidable. Lumentum’s early investment puts them in a position to actively shape this future, tackling the need for AI hardware optical solutions.
Lumentum is making it clear with its focus on die-to-die optical interconnects that it intends to be a foundational technology supplier for the AI era. By solving the bandwidth and power problems at the heart of AI computing, they’re enabling the next wave of AI innovation. For anyone building AI infrastructure, it’s time to accept that optical interconnects are fast becoming a requirement, not a choice, for future scale and performance.
What are die-to-die optical interconnects?
They’re a technology that uses light instead of electricity to send data between different processor dies, either on the same package or very close to each other. This method offers huge benefits in speed, power efficiency, and the amount of data you can move compared to old-school copper traces.
Why are optical interconnects important for AI computing?
AI, especially training huge models, needs to shuttle massive amounts of data between processors and memory at insane speeds. Electrical connections are hitting their physical limits on bandwidth and power efficiency. Optical interconnects break through those limits, allowing for faster data transfer with less power, which is essential for building the next generation of AI systems.
How do die-to-die optics differ from co-packaged optics (CPO)?
Co-packaged optics (CPO) puts optical transceivers on the same board as the main processor chip. Die-to-die optics takes that integration a step further, placing the optical components even closer to the processing unit, sometimes directly on the same silicon which minimizes electrical signal paths even more for maximum efficiency.
What challenges does Lumentum face in bringing die-to-die optics to market?
The main hurdles are scaling up manufacturing to high volumes while keeping quality high, getting the cost down to be competitive with electrical alternatives, and helping to create industry-wide standards so components from different vendors can work together. Lumentum is using its manufacturing experience and key partnerships to tackle these issues.
What role does Lumentum’s indium phosphide (InP) expertise play?
Lumentum’s deep knowledge of indium phosphide (InP) is a major advantage. InP is the perfect material for creating the tiny, high-performance lasers and detectors needed for these optical engines. This expertise allows them to design and build more efficient and compact interconnects than competitors who may not have the same materials science background.
