# How On-Chip Silicon Photonics Are Shrinking Data Centers

An episode of [The Hardware Podcast with Fexingo: Chips, Devices, and Electronics Engineering Conversations](https://invisiblenetwork.ai/podcast/1896799982), 1 Aug 2026, 9:50.

Audio: https://audio.fexingo.com/business/the-hardware-podcast/episode-0142.mp3

Episode 142 of The Hardware Podcast dives into silicon photonics—the technology replacing copper wires with light inside data centers. Lucas and Luna break down why we're hitting a bandwidth wall, how companies like Intel and Broadcom are shipping optical engines on chip, and what this means for power consumption and AI clusters. They explore the physics of waveguides, modulators, and photodetectors, and how manufacturing silicon photonics leverages existing CMOS fabs. The episode also touches on the economics: why optical interconnects are dropping in cost and where co-packaged optics will land next. If you're curious about the hardware that will make future AI models possible, this episode is for you. #SiliconPhotonics #DataCenters #OpticalInterconnects #CoPackagedOptics #CMOS #AIHardware #Bandwidth #LightOnChip #Intel #Broadcom #PhotonicsTechnology #SemiconductorIndustry #Technology #HardwarePodcast #FexingoBusiness #BusinessPodcast #EngineeringInnovation #FutureOfComputing Keep every episode free: buymeacoffee.com/fexingo

## Transcript

Lucas: So we've talked a lot about what chips can do inside the package, but today I want to zoom out a little and talk about the wires that connect them.
Luna: You mean like the copper cables in a data center? I feel like that's not the most glamorous topic.
Lucas: Exactly, but it's becoming the bottleneck. And the fix is light. Silicon photonics — putting optical components on the same chip as the electronics. It's not new in the lab, but it's finally hitting the mainstream in a big way.
Luna: I know we've covered optical interconnects before on a chip scale, but this sounds like it's about the network between chips, not inside them.
Lucas: Right, and that's the key distinction. We did an episode on on-chip optical interconnects for speed, which is about moving data within a chip. Today it's about chip to chip, rack to rack, and how data centers are migrating from electrical to optical links.
Luna: So why now? Copper has worked for decades.
Lucas: Because we're hitting a wall. As data rates push past 100 gigabits per second per lane, copper gets lossy, power-hungry, and the reach shrinks. You can't run a 100-gigabit electrical signal more than a meter or two without serious equalization. In a data center, racks can be tens of meters apart.
Luna: So you switch to fiber. But the cost of pluggable optical transceivers has always been high.
Lucas: Historically, yes. But silicon photonics changes the economics. By using standard CMOS manufacturing — the same fabs that make your phone's processor — you can integrate modulators, waveguides, photodetectors, and even the laser driver circuitry on one chip. That drives the cost per bit down dramatically.
Luna: So it's about leveraging the entire semiconductor ecosystem. That's a huge moat for companies like Intel and TSMC.
Lucas: Exactly. And Intel's been shipping silicon photonics for years, mainly in pluggable modules for cloud networking. But the next wave is co-packaged optics, where the optical engine sits right next to the switch chip or the accelerator, in the same package.
Luna: That's a big deal because the closer the optics are to the compute, the less power you burn on electrical signaling.
Lucas: And the latency drops too. We're talking about going from pluggable modules at the faceplate to optics on the substrate. That cuts the electrical trace length from tens of centimeters to maybe a few millimeters. For AI clusters, where you're shuttling terabytes between GPUs, that's a game changer.
Luna: I've read that NVIDIA's NVLink has been moving to optics for the next generation. Is that part of this trend?
Lucas: It's the same underlying physics. NVLink is a proprietary interconnect, but the industry is standardizing on optical I/O for the data center fabric. The Open Compute Project has a whole initiative on co-packaged optics, and companies like Broadcom, Marvell, and TSMC are all investing heavily.
Luna: So what does a silicon photonics chip actually look like? Is it a separate chiplet or is it integrated into the main die?
Lucas: Both, depending on the approach. Some designs use a separate photonic chiplet that's co-packaged with the switch or GPU. Others are more monolithic, integrating photonics directly onto the same die as the electronics. But the challenge is that photonics needs different process steps than logic — silicon waveguides, germanium photodetectors, and so on.
Luna: So they're not just shrinking transistors. They're adding new materials and structures.
Lucas: Right. And that's why TSMC has a dedicated photonics platform. They've been working on it for a decade. The idea is to provide a standard process that designers can plug into, just like they would a logic process.
Luna: But what about the laser? You can't make a laser on a standard CMOS process. Or can you?
Lucas: That's the one piece that's still not fully integrated. Most silicon photonics use an external laser or a laser co-packaged with the photonic chip. But there's a lot of research on integrating iii v materials like indium phosphide onto silicon. Intel's been working on hybrid lasers for a while.
Luna: So the laser is the last holdout. But the rest of the optical chain — modulators, waveguides, detectors — can be made in a regular fab.
Lucas: Exactly. And that's what makes it manufacturable at scale. You can produce millions of photonic chips a year using equipment that's already depreciated. The unit cost drops, and suddenly optical interconnects become cost-competitive with copper at shorter and shorter distances.
Luna: You mentioned power earlier. How much power are we actually saving by going optical?
Lucas: At the system level, it's significant. A typical electrical SerDes at 112 gigabits per second can consume 10 to 15 milliwatts per gigabit per second. That's over a watt per link. Optical links, especially co-packaged, can get that down to maybe 5 to 10 milliwatts per gigabit per second. In a data center with tens of thousands of links, that's a huge reduction in electricity and cooling.
Luna: And for AI clusters, since they're already power-hungry, every watt you save on networking is a watt you can use for compute.
Lucas: That's the whole point. The bottleneck isn't just bandwidth; it's power. And silicon photonics addresses both. That's why I think this is the decade where optical goes mainstream inside the data center.
Luna: So what's the timeline? When will we see co-packaged optics in production at scale?
Lucas: We're already seeing early deployments. Broadcom has been shipping co-packaged optics for their Tomahawk switch lines. And by 2027 or so, I expect most new high-end switches and accelerators to have optical I/O as an option. The next few years are going to be critical.
Luna: It's fascinating how a technology that was once relegated to long-haul telecom is now coming inside the machine room. It's a real convergence of materials science and semiconductor engineering.
Lucas: Absolutely. And you know, this episode is a perfect example of why we keep this show ad-free. We want to dive deep into topics like this without any sponsor telling us to keep it simple or to mention their product. If you value that, a small contribution at buy me a coffee dot com slash fexingo helps us keep doing it.
Luna: Yeah, it's the listeners who make this possible. We appreciate every coffee.
Lucas: So back to the technology — one of the more exciting developments is the move toward 200-gigabit per lane optical links. The industry is already shipping 800-gigabit modules, and co-packaged optics are expected to lead the way to 1.6 and even 3.2 terabits per switch.
Luna: And that's not just about more bandwidth. It's about reducing the number of fibers and cables, which simplifies the data center infrastructure.
Lucas: Exactly. Fewer cables mean less airflow blockage, easier maintenance, and lower cost. Plus, with co-packaged optics, you eliminate the faceplate pluggables, so you get a cleaner signal path and better signal integrity.
Luna: But doesn't co-packaged optics make it harder to replace a failed component? With pluggables, you just swap the module.
Lucas: That's a valid concern. If a co-packaged optical engine dies, you might have to replace the whole switch or GPU board. But the reliability of these components is improving, and the power and performance benefits outweigh the serviceability trade-off for many workloads.
Luna: I suppose for AI training, where you can't afford downtime, redundancy is built in anyway. So the risk is mitigated.
Lucas: And that's why the hyperscalers are pushing this. They're the ones with the most to gain from lower power and higher bandwidth. They can also absorb the cost of early adoption.
Luna: So if I'm a smaller company, I might wait a few years until the ecosystem matures and the prices drop further.
Lucas: That's the typical cycle. But as the volume ramps, the cost per link will fall, and eventually optical will be the default for most data center connections, even at short distances.
Luna: And what about the technology itself? Are there any remaining physics hurdles?
Lucas: The big one is the laser again. We need low-cost, high-reliability lasers that can be integrated at scale. Also, the packaging of these photonic chips with electronics is still tricky — you have to align fibers to the waveguides with sub-micron precision. But the industry has developed passive alignment techniques that are making it more manufacturable.
Luna: So it's more about engineering than fundamental science at this point.
Lucas: Exactly. We know how to do it; we just need to make it cheap and reliable enough. That's the classic semiconductor challenge. And given the track record of this industry, I'm optimistic we'll get there.
Luna: It's exciting to think that in a few years, the data centers powering our AI services will be optically interconnected at their core. It's a quiet revolution.
Lucas: And it's happening right now. The next time you use a cloud service or an AI chatbot, there's a good chance your request is being routed through a silicon photonics link somewhere. And that's only going to become more common.

Shows are listed from Apple Podcasts' charts and directory, and Podcast Index where it is used. Every show and episode belongs to its publisher and plays from the publisher's own feed; this site hosts none of them.