Summary

This piece explores why silicon photonics, CPO (Co-Packaged Optics), and optical I/O technologies are necessary, and specifically why light plays a central role in information transmission. Starting from the fundamental principles of how information travels, it compares copper and light, and traces the historical arc of light expanding from long-haul communication into ever-shorter distances. It pays particular attention to how the AI era — with its explosive growth in data transfer demands — has dramatically accelerated the shift to light-based technologies, and walks through the three generations of technology bringing optics closer to the chip: pluggable transceivers, CPO, and optical I/O. It closes with an introduction to silicon photonics and the pJ/bit metric, laying the groundwork for understanding where the key players in this market are positioned.

1. The Most Efficient Way to Transmit Information: Light 💡

If you follow semiconductor news these days, you've probably come across terms like silicon photonics, CPO (Co-Packaged Optics), and optical I/O. When NVIDIA declared it would "connect chips with light," what did that actually mean? This is the first installment of a three-part series that answers that question — explaining in detail why light is essential for information transmission.

To move cargo, you need a truck. To transmit power, you need a wire. To transfer force, you need gears or pulleys. What all of these have in common is that something with mass must physically move.

But information transmission is fundamentally different.

Imagine flashing a flashlight in Morse code to someone far away in the dark of night, or lighting a signal fire on a mountaintop like ancient beacon towers. Nothing with mass flew between those peaks — light carried the information. The key insight is that transmitting information requires no movement of physical matter. All you need is a "pattern" — a signal — agreed upon between sender and receiver. And the lighter, faster, and more energy-efficient the medium carrying that signal, the better.

Is there a medium with no mass, moving at extraordinary speed, that doesn't interfere with itself?

Yes: light. ✨

2. Copper vs. Light: Two Pipes for Data Transmission ⚔️

So how do modern data centers actually move information? Imagine two kinds of pipes.

2.1. The Copper Pipe and Electrons ⚡️

The first is a steel ball pipe. When you shoot a steel ball through a metal pipe, the ball has mass, so it collides with the pipe walls and loses energy — that energy becomes heat. The faster you shoot the ball, the worse the collisions and the more heat is generated. If there's another pipe nearby, vibrations in one affect the other, creating crosstalk. The longer the pipe and the faster the ball, the worse all of these problems become.

This is precisely what happens when electrons move through copper wire. Electrons have mass and charge; they constantly collide with atoms in the conductor and interact electromagnetically with neighboring conductors. 😥

2.2. The Laser Pipe and Light 🌟

The second is a laser pipe — imagine a glass tube lined on the inside with mirrors. When you shoot a laser through it, light has no mass, so bouncing off the mirrored walls costs it no energy. And even if there's another glass tube right next to it, light generates no electric field, so there is zero interference. Remarkable, isn't it?

Light has one more astonishing property. If ten kids are racing in a hallway and can't pass through each other, ten is about the limit before collisions. But what if the kids could pass straight through each other? Thirty, a hundred kids could run simultaneously without any collision at all. Light works exactly like that. Beams of different wavelengths can share the same space without interfering with one another. This is called Wavelength Division Multiplexing (WDM).

The performance difference is staggering. With today's technology, a single copper wire can carry roughly 100–200 Gbps.

But a single optical fiber using WDM can carry 64–96 Tbps — hundreds of times more capacity than copper. 🚀

Physically, light wins overwhelmingly. So the real question becomes:

Why are we still using copper?

3. Light's Victories and Copper's Resistance: An Expanding Battlefield 🗺️

Remarkably, light has already beaten copper — it just started with the longest distances first.

In 1988, when the first transoceanic fiber optic cable, TAT-8, was laid across the Atlantic, it delivered ten times the capacity of copper cables in one stroke. Since then, light has steadily expanded its territory.

• 1980s: Intercontinental (thousands of km) — transoceanic fiber optic cables
• 1990s: City to city (hundreds of km) — commercial WDM and optical amplifiers
• 2000s: Data center to data center (a few km) — 10 Gigabit Ethernet
• 2010s: Inside data centers, rack to rack (tens to hundreds of meters) — VCSELs and multimode fiber

Light has always won at long range first, then pressed inward — conquering concentric rings from the outside in.

But copper still holds its short-range fortress with tenacity. There's an old networking saying:

"If you can do it in copper, do it in copper." — The first rule of photonics

At distances of a few meters and speeds of hundreds of gigabits, copper is cheap and easy to deploy. Existing infrastructure is enormous. Even though light is physically superior in pure terms, at short range copper's advantages more than offset its weaknesses.

Optics, on the other hand, requires modules to convert electricity to light on the transmit side and light back to electricity on the receive side — lasers, conversion circuitry, transceivers. All of that cost has kept optics from beating copper at short range.

3.1. The Arrival of AI and the Shifting Battlefield 🤖

But that balance is being broken — by AI. 💥

What AI changed is not simply "more data." The requirement that tens of thousands of GPUs function as a single giant brain has brought about a fundamental shift.

In the early 2010s, four to eight GPUs per server was plenty. The physical distance between GPUs was just a few centimeters — easily handled by copper. But AI model scale has grown exponentially. GPT-4 used roughly 25,000 GPUs; xAI's Colossus uses 100,000. These GPUs are spread across hundreds of server racks, stretching physical distances from centimeters to tens of meters. And they must exchange computation results every few milliseconds. If even one GPU falls behind, everyone waits.

Every time the speed doubles, the distance copper can reliably handle shrinks by 30–50%. A standard server rack is about two meters tall, and at 800G speeds, two meters is already the limit. Copper is beginning to struggle even within the same rack.

Today, the front line between optics and copper is at server-to-server distance (a few meters). The next battlefield will be chip-to-chip distance (a few centimeters).

This is not a one-sided fight. The optics camp has the laws of physics as an absolute ally and silicon photonics as a new weapon. But the copper camp has fought back every generation with faster SerDes and more sophisticated equalizers. The cost advantage at short range and decades of entrenched ecosystem are not a fortress that falls easily.

Worth noting: even in copper's deepest stronghold — inside the chip — scouts from the optics camp are already operating behind enemy lines. That story will be revisited in the epilogue of Part 2.

Understanding this battlefield map has a practical purpose: it is the key to locating the companies covered in Part 2 within this market. Some are generating steady revenue from the already-conquered outer rings. Some are fighting on the current front line. And others are betting on the inner rings that haven't opened yet.

4. Silicon Photonics: A Smart Partnership for Silicon That Can't Make Light 🤝

Silicon is the king of semiconductors — excellent at making transistors, memory, processors, and virtually everything else. But there is exactly one thing it does terribly: making light. 😫

4.1. Bandgap: The Ball-Falling-Off-a-Shelf Analogy ⚾️

To understand why, you need to know about the bandgap. (This is the same concept covered in a previous piece on power semiconductors.)

Imagine a ball sitting on a high shelf. Drop it to the floor and it makes a loud "thud" — the energy corresponding to the shelf height is released as sound. Something similar happens in semiconductors. When an electron "falls" from a high-energy state (the conduction band) to a low-energy state (the valence band), that energy difference can be released as light (a photon). Just as a falling ball makes sound, a falling electron makes light. This is the principle behind LEDs and lasers.

But here is where the paths diverge.

• In direct bandgap materials (e.g., InP, GaAs), the ball falls cleanly off the edge of the shelf. It makes a loud, clean thud — bright, efficient light.
• In indirect bandgap materials (e.g., silicon), the ball catches on a lip at the shelf edge. A lattice vibration (phonon) has to nudge it sideways before it can fall. Three things must align simultaneously — a "triple coincidence." The odds are so low that making light from silicon is essentially impossible. 😔

4.2. Silicon Photonics: A Collaboration Between Silicon and Other Materials 💡

If silicon can't make light, the answer is to find a material that can. Just as silicon ceded ground to GaN and SiC for high-voltage handling in power semiconductors, the answer here lies in III-V compound semiconductors. InP is used for long-haul fiber lasers; GaAs is used for short-range VCSELs.

But silicon has a trick of its own. Silicon has a high refractive index (~3.5), which means it can confine light within extremely narrow channels. The "mirror-lined glass tube" from section 1 can be etched onto a chip at widths of a few hundred nanometers — this is a waveguide. On top of that, you can build devices to modulate, filter, and convert light back to electricity. Crucially, all of this can be fabricated in existing CMOS fabs. 🏭

Silicon photonics in one sentence:

Materials that make light well (InP, GaAs) generate the light; silicon guides and manipulates it. A partnership.

This combination is what makes optical interconnects industrially scalable.

5. pJ/bit: The Key Metric for Optical Technology Progress 📉

In section 2, I mentioned that the "cost of electrical-to-optical conversion" is why optics couldn't beat copper at short range. Let's look more closely.

Chips compute using electricity. Therefore every optical link requires an electrical → optical → electrical (E-O-E) conversion. The device that performs this is the transceiver.

On the transmit side, a driver circuit flickers a laser at tens of GHz. The receive side is more demanding: a photodetector converts light into current, but this current is as faint as a candle flickering in a dark room. The TIA (Transimpedance Amplifier) acts like a telescope, making that faint flicker visible from a distance.

Then comes the DSP (Digital Signal Processor). No matter how well you transmit, signals become distorted and degraded in transit. The DSP restores the damaged signal. Indispensable — but the problem is that this restoration process consumes roughly half of the transceiver's total power. 😱

5.1. The Most Important Metric: pJ/bit

Of all the metrics for comparing optical interconnect technologies, if you can only track one, it's pJ/bit (picojoules per bit) — the amount of energy required to transmit one bit of information.

The direction of decreasing pJ/bit is the direction of technological progress.

The technology for bringing light closer to the chip has evolved across generations.

One singular goal: reducing the electrical distance between light and chip.

Think of it like watering a flower bed. The closer the bed is to the tap, the easier and less wasteful the watering. 🌸

5.2. Generation 1: Pluggable Transceivers 🔌

Pluggable modules are exactly what the name says — easy to plug in and pull out. Think of them like USB dongles: devices that convert electricity to light and light back to electricity. They slot into the front panel of a server or switch, and when one fails, you just replace the module. They're compatible across manufacturers. 👍

The problem is that the electrical signal has to travel tens of centimeters from the front panel to the switch ASIC inside. It's like pumping water to a flower bed on the far side of a wide field. Over that distance, the signal attenuates to less than 1/100th of its original strength, and the DSP must reconstruct it — consuming enormous power in the process.

The numbers tell the story: a single 800G transceiver consumes about 15W. A 51.2 Tbps switch needs 64 such modules, so transceivers alone consume roughly 960W. 🤯 That's comparable to running a powerful hair dryer continuously — and it's roughly equal to the power draw of the switch ASIC itself, spent purely on converting electricity to light. The saying "if you can do it in copper, do it in copper" didn't come from nowhere.

5.3. Generation 2: CPO (Co-Packaged Optics) 📦

True to its name, co-packaged optics puts the switch ASIC and the optical engine responsible for E-O conversion inside the same package, fundamentally shrinking the physical distance.

Tens of centimeters become a few millimeters. Instead of pumping water across a field, you're watering a pot right at your front door. Signal loss drops dramatically, the DSP burden lightens, and power consumption can fall by up to 70% compared to pluggables. 👍👍

But there's a thermal paradox. Lasers are extremely sensitive to temperature — even a one-degree shift destabilizes the wavelength. Meanwhile, ASICs and GPUs consume hundreds of watts and run extremely hot. It's like putting a freezer and a furnace under the same roof. Serviceability is also a concern: with pluggables you swap just the module; with CPO you may need to replace the entire package.

How are companies solving this? Broadcom separated the laser into a replaceable External Laser Source (ELS) module; NVIDIA is pursuing liquid cooling.

If you want to know when the CPO market will materialize, watch how each company resolves the thermal problem and the serviceability problem.

5.4. Generation 3: Optical I/O 🚀

If CPO is the front door, optical I/O is the desktop. The photonic engine becomes a chiplet, sitting alongside GPU or CPU chiplets within a single package. The electrical distance between light and chip shrinks to a few hundred micrometers — the width of a few strands of hair.

The standard making this possible is UCIe (Universal Chiplet Interconnect Express). GPU chiplets, memory chiplets, and optical I/O chiplets communicate in a common language, collaborating within a single package. Just as different software tools in a "digital twin" setup need a shared language to build a common 3D world, different chiplets need standards like UCIe to work together.

If this succeeds, it would represent the ultimate integration — and could push copper out of chip-to-chip communication entirely. NVIDIA has expressed confidence that NVLink itself will transition to optical by around 2028. The day copper's front line gets pushed back may be closer than it seems. 🤩

Conclusion

We've traveled together through the journey toward the age of light.

Information can be transmitted without mass, and light — which has no mass, no charge, and does not interfere with itself — holds an overwhelming physical advantage.

Light began conquering from a distance, crossing oceans, spanning cities, bridging data centers, moving rack to rack, and now pressing into the server itself. And the arrival of the AI era has accelerated this advance enormously. 📈

Today's technology is evolving through three generations to bring light ever closer to the chip. Pluggable transceivers have hit a wall of power consumption. CPO faces the challenge of a thermal paradox. Optical I/O is now just crossing the threshold. And beyond all three lies an unexplored frontier: photonic computing, where light itself is used for computation.

The technological terrain has been mapped. The next question is this:

Who is standing where on that terrain?

Part 2 will unfold the company map of this market. Stay tuned! 😉