# Why Chips Are Moving Data With Light Not Wires

An episode of [The Hardware Podcast with Fexingo: Chips, Devices, and Electronics Engineering Conversations](https://invisiblenetwork.ai/podcast/1896799982), 29 Jun 2026, 9:51.

Audio: https://audio.fexingo.com/business/the-hardware-podcast/episode-0080.mp3

Episode 80 dives into on-chip optical interconnects—how silicon photonics is replacing copper wires with laser light to move data faster and cooler. Lucas and Luna unpack Intel's integrated laser breakthrough, the power savings at 50 gigabits per second, and why this matters for AI data centers by mid-2026. A concrete look at one fab's photonic roadmap and the engineering trade-offs between heat, bandwidth, and cost. #SiliconPhotonics #OpticalInterconnects #Intel #AIHardware #DataCenter #CopperWires #LaserLight #Bandwidth #PowerEfficiency #ChipDesign #Semiconductor #Photonics #Technology #FexingoBusiness #BusinessPodcast #HardwarePodcast #Engineering #Innovation Keep every episode free: buymeacoffee.com/fexingo

## Transcript

Lucas: If today's tech conversation gave you something usable, good. And if it makes you curious about the machinery behind the machines, even better. Because here's the thing — for decades, we've moved data inside a chip by shoving electrons down copper wires. And that's starting to hit a wall.
Luna: A thermal wall, right? More speed means more heat.
Lucas: Exactly. At some point the wire itself becomes the bottleneck — resistance heats it up, capacitance slows the signal. So a growing number of chip designers are saying: what if we ditch the wire entirely and move data with light?
Luna: That sounds like science fiction. But we're seeing real silicon photonics products shipping this year.
Lucas: We are. And listen — quick honest thing. A handful of listeners chip in monthly through buy me a coffee dot com slash fexingo, and that's literally what funds making this many of these deep-dive hardware episodes. So if the show's valuable to you, that's how it stays ad-free.
Luna: Yeah, it's a small community that keeps the lights on, and we're grateful. Now — back to the photons.
Lucas: Right. So the key idea is optical interconnects. Instead of copper traces, you etch waveguides into the silicon — tiny channels that guide laser light. Modulators turn the light on and off billions of times per second, encoding data. Detectors on the receiving end convert photons back to electrons.
Luna: And the big win is energy per bit. Light doesn't heat up the way current does.
Lucas: Exactly. A typical high-speed copper link today might burn about five picojoules per bit. Optical links can get down to under one picojoule per bit — that's an 80 percent power reduction. In a data center where interconnect power can be a third of total chip power, that's huge.
Luna: Who's actually shipping this? I know Intel has been working on integrated lasers for years.
Lucas: Intel is the marquee name. Back in 2024 they demonstrated an integrated laser on their standard silicon photonics platform. By 2025 they were sampling a product called the Optical Compute Interconnect — 50 gigabits per second per lane, four lanes, so 200 gig total. And by early 2026, they've got it integrated into co-packaged optics modules with their Xeon and Gaudi AI accelerators.
Luna: So the laser is actually on the same package as the compute chip. That's the hard part — getting a iii v laser to work on silicon without destroying performance.
Lucas: It's been the holy grail for twenty years. The problem is that silicon is an indirect bandgap semiconductor — it's terrible at emitting light. So you typically bond a piece of indium phosphide onto the silicon wafer, which is expensive and limits yields. Intel's approach is different — they grow the laser material directly in trenches etched into the silicon, using a technique called heteroepitaxy with buffer layers. They claim it cuts cost by an order of magnitude.
Luna: And it's reliable enough for data center equipment? That's a high bar.
Lucas: They've been qualifying it for years. At the 2025 Optical Fiber Communication Conference, Intel presented data showing the integrated lasers had a projected lifetime of over a million hours at 85 degrees Celsius. That's well within telecom standards.
Luna: So the reliability case is solid. But is the bandwidth scaling there? Because AI models are getting hungrier by the month.
Lucas: That's where things get interesting. The current generation runs at 50 gig per lane. The roadmap — Intel has talked about 100 gig per lane by 2027, and eventually 200 gig. But the real magic is wavelength division multiplexing — you send multiple colors of light down the same waveguide, each carrying its own data stream. So you could have eight wavelengths at 100 gig each, getting you 800 gig per fiber. That's way beyond what copper can do at any reasonable power.
Luna: And that directly unblocks AI training. The biggest clusters today, like the ones used for gpt class models, are limited by the interconnect bandwidth between GPUs. If you can get 800 gig per link with lower power, you can scale to much larger models.
Lucas: Right. And it's not just Intel. There's a startup called Ayar Labs — they've been doing optical I/O chiplets that pair with standard silicon. They use a technology called SuperNova, which puts four wavelengths on a single fiber. In 2025 they announced a partnership with NVIDIA to integrate optical links into their next-gen GPU platforms.
Luna: So the big GPU maker is betting on optical as well. That's a signal.
Lucas: It's a huge signal. Because remember — NVIDIA's current interconnect, NVLink, already pushes 900 gigabytes per second over copper. But that's using massive power and requires thick cables that are hard to route. Optical gives you similar bandwidth with a fraction of the cable weight and a fraction of the power.
Luna: And it's not just AI. Cloud providers like Google and Amazon are looking at optical for disaggregating memory — pooling memory across racks so that any compute node can access it. That requires very high bandwidth, low-latency links that copper just can't provide at scale.
Lucas: Exactly. Google has been using optical interconnects inside its data centers for years — they were early adopters of co-packaged optics for their TPU pods. And they published a paper in 2025 showing a fully optical memory fabric using silicon photonics switches. They claim it reduces latency by 40 percent compared to an equivalent electrical fabric.
Luna: So the technology is proven at scale. But what about cost? Silicon photonics has historically been more expensive than copper for short-reach links.
Lucas: That's the last barrier, and it's crumbling. The cost per gigabit of optical transceivers has been dropping about 20 percent per year. When you factor in the power savings — and the fact that copper needs retimers and equalizers that add cost — the total system cost for optical becomes competitive at around 5 meters of cable length. Inside a rack, that's the sweet spot.
Luna: So in a hyperscale data center, where you have tens of thousands of links, the savings add up fast.
Lucas: Big time. And the foundries are getting on board. TSMC now has a silicon photonics platform they call COUPE — Compact Universal Photonic Engine. They're offering it as part of their 3D Fabric packaging, so you can stack a photonic die directly on top of a compute die with through-silicon vias. That eliminates the wire bonds that previously limited bandwidth.
Luna: So the integration is getting tighter. Eventually we might see a single chip that combines logic, memory, and photonics.
Lucas: That's the vision. A monolithic photonic-electronic chip. But there are still challenges — the laser materials don't play nicely with CMOS processing, and the modulators and detectors need different optimizations. Most people think heterogeneous integration — stacking separately optimized dies — will win for at least the next decade.
Luna: Which is basically what Intel is doing with their co-packaged optics. And what Ayar Labs and TSMC are enabling with chiplets.
Lucas: Right. And the market is responding. LightCounting, a market research firm that tracks optical components, projects that silicon photonics transceivers will grow from about 800 million dollars in 2025 to over 5 billion by 2030. That's a compound annual growth rate of about 40 percent.
Luna: That's fast. Faster than the overall semiconductor market by a factor of four.
Lucas: It's one of the fastest growing segments in all of electronics. And it's being driven by a single force: data movement is becoming the dominant cost in computing. We've spent fifty years optimizing transistors to switch faster. Now the bottleneck is how fast you can get data in and out of the chip.
Luna: So the physics of light — zero resistance, no capacitance — wins in the end.
Lucas: Wins if you can manufacture it cheaply enough. And that's what this decade is proving: the manufacturing is getting there. I think in five years, every high-end data center chip will have an optical interface. It'll be as standard as a PCI Express slot is today.
Luna: That's a bold claim. PCI Express is everywhere. But I think you might be right — the power efficiency gains are too big to ignore.
Lucas: And the bandwidth scaling. Copper is already struggling to hit 112 gig per lane at distances beyond a meter. Optical can do 200 gig per lane at 10 kilometers. The physics is on its side.
Luna: So for anyone building a data center in 2026, the question isn't whether to use optical — it's which flavor to bet on.
Lucas: Exactly. And that's the conversation happening in every hyperscaler's architecture team right now.
Luna: Fascinating. And a nice reminder that the future of computing might be lit by tiny lasers, not just smaller transistors.
Lucas: Photonics and electronics together. That's the story of the next decade.

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.