If you’re new here, start with this piece. Everything else we publish will build on what you learn in the next hour. Bookmark it. Come back to it. This is the foundation.
What Is Photonics?
Light vs. electricity
Most data inside a data center moves in one of two ways: as electrical signals through copper wires, or as light signals through fiber optic cables or silicon waveguides. Photonics is the science and industry focused on using light to transmit, process, and detect information. In modern systems, photonics and electronics work together where electricity moves data over very short distances, and light moves data efficiently over longer distances and higher bandwidth links.
Passive copper is like sending everyone by car down a crowded highway. It works, but as traffic increases, the roads get congested, movement slows, and you burn a lot more fuel. Photonics is like putting that traffic onto a high-speed rail line. The trains move faster, carry far more people at once, and do it with much less wasted energy.
That, in a nutshell, is why photonics is increasingly replacing copper where data needs to move at faster and faster speeds and in larger volumes.
What is a photon?
A photon is the smallest unit of light. When you flip a light switch, billions of photons are emitted every second. In optical communications, we generate photons very precisely using lasers, encode information into the light signal, send that light through a glass fiber thinner than a human hair, and detect it at the other end. That’s how data moves through much of the modern internet and data center infrastructure.
The faster we can change or modulate the light signal, the more data we can send. Over time, optical systems have gone from sending millions of bits per second (megabits), to billions (gigabits), to hundreds of billions per second. When you hear “400G” or “800G,” the G stands for gigabits per second, which is roughly how much data that connection can carry every second.
Why light is (usually) better than copper for moving data
Copper has been the backbone of communications for a long time. It is cheap, easy to work with, and still works very well for short distances. But as data rates rise, copper runs into physical limits that become much harder to manage.
First, copper links consume more power as speeds climb, and that extra power often shows up as heat. In a dense data center, heat is expensive because it has to be removed with cooling systems.
Second, electrical signals degrade over distance. The faster the signal, the harder it is to preserve cleanly over copper. At 800G, passive copper cables are generally limited to very short reaches, often around 1 to 3 meters depending on the design.
Third, copper is more vulnerable to signal integrity problems such as noise and interference, especially as link speeds and cable density increase.
Optical fiber behaves much better on all three fronts. Light traveling through fiber loses much less signal over distance, generates very little heat in the fiber itself, and can carry data over much longer reaches than copper. Fiber also supports wavelength division multiplexing, or WDM, which means multiple wavelengths of light can travel through the same strand of fiber at the same time, with each wavelength acting like its own lane of traffic.
That does not mean fiber is perfect. Light still weakens over long distances, which is why long-haul networks need amplifiers. Fiber is also more fragile than copper, harder to handle in the field, and more sensitive to dirt and damage at the connector. Optical transceivers are precision components and can be more expensive than simple copper cables. That is why copper DACs are still common for the very shortest links. But once you need to move more data, over longer distances, with better power efficiency, optics usually wins.
What the industry is doing to push copper further
Copper is not going away quietly. Because optical transceivers are more expensive, data center operators want to use copper anywhere the physics still works, and engineers have developed several ways to extend copper’s reach as speeds increase.
Passive DAC (Direct Attach Copper) is the simplest option where a copper cable with connectors is attached at both ends and no active electronics in the cable itself. At 800G, these links are generally limited to very short reaches, often around 1 to 3 meters.
ACC (Active Copper Cable) adds analog equalization in the cable assembly to improve signal integrity and extend reach. At 800G, ACC can often stretch copper into the roughly 3 to 5 meter range.
AEC (Active Electrical Cable) goes further by adding retimer-based electronics in the cable ends. These devices recover and retransmit a cleaner signal, which can extend copper reach to roughly 5 to 7 meters at 800G while improving reliability and compatibility. The tradeoff is added cost, some extra power, and a bit more complexity.
At the same time, switch and NIC chips are getting better at signal conditioning. More advanced SerDes and equalization on the host side allow systems to tolerate noisier electrical links than before. That is part of why newer linear architectures, including LPO, have become more practical.
The broader point is that copper and optics are not simply winner-take-all technologies. They coexist across a distance and power spectrum. Copper still dominates the shortest links because it is cheaper and simpler. But as speeds rise, the distance over which copper remains practical shrinks, and more links move into the optical category. That is one of the structural reasons demand for optical interconnects keeps rising.
What “optical” means
You’ll see the word “optical” constantly. It simply means “uses light.” An optical transceiver is a device that converts electrical signals from a chip into light signals for transmission, and then converts those light signals back into electrical signals at the other end. An optical cable carries data using light rather than electrical current through copper. An optical network moves data primarily using light across its links. So photonics is the broader field, while optical usually refers to the specific components and systems built from it.
Why Does This Matter Right Now?
Optical fiber has been used in long-distance communications for decades. The cables connecting continents are fiber. The backbone of the internet is fiber. None of that is new.
What changed is where the bottleneck now sits. For much of the internet’s history, the hardest communication problem was moving data over long distances like between cities, continents, and buildings. Inside a building or data center, shorter copper connections were often good enough.
That is changing. The amount of data moving inside a modern AI data center has become so large, and the required speeds so high, that short-reach electrical links are running into growing power, reach, and signal-integrity limits. The bottleneck has not disappeared outside the building, but it has increasingly moved inside the data center itself.
The data explosion
For decades, internet traffic kept growing as more of life moved online. Streaming video, cloud computing, social media, and mobile apps all pushed more and more data through networks every year.
But AI changes the problem again.
Traditional internet growth mostly meant more data moving across networks. AI also means far more data moving inside the data center itself, between chips, servers, racks, and clusters. And it has to move with much lower latency and much higher bandwidth than older workloads required.
That is why photonics matters so much right now. The challenge is no longer just sending data across the world. It is sending enormous amounts of data efficiently across the inside of a machine, a rack, and a data center.
The power problem is the other wall copper runs into
We just covered how copper struggles with distance at high speeds. But even where active copper cables and retimers can keep up, the range where AECs can still do the job reasonably well, a second problem remains in power.
Active copper does not extend reach for free. The electronics inside AECs draw real power on every link. That may not sound like much in isolation, but multiplied across a large AI cluster, it becomes meaningful. In effect, the industry is spending more power to keep copper viable for another generation.
Meanwhile, power has become one of the central constraints on AI infrastructure more broadly. Amazon, Microsoft, Alphabet, and Meta collectively spent more than $400 billion on capital expenditures in 2025, much of it tied directly or indirectly to AI infrastructure and the power, cooling, and facility buildout needed to support it.
Networking is not the largest line item on the power bill, but it does matter. At hyperscale, even relatively small increases in watts per link add up quickly when you are deploying thousands or hundreds of thousands of high-speed connections. That is one reason power efficiency matters so much in interconnect design as every extra watt has to be generated, delivered, and cooled.
How AI Is Turbocharging Photonics Demand
AI training is unlike most computing workloads that came before.
Many traditional computing tasks like serving web pages, handling database queries, or streaming video can be spread across large numbers of servers that work relatively independently. One server can handle one request without needing to stay in constant lockstep with thousands of others.
AI training is different. When you train a large model, one enormous calculation is split across thousands of chips working at the same time. Every few milliseconds, those chips need to synchronize, exchange updates, and stay aligned as they train the same model together.
A useful way to think about it is like 1,000 people working on the same shared document at once. Every so often, everyone has to send in their latest edits, combine them into one updated master copy, and then give that updated version back to every participant. In AI systems, one of the key communication steps that does this is called AllReduce.
This has a huge consequence where the network between chips becomes almost as important as the chips themselves. If the network is slow, congested, or power-inefficient, expensive processors sit idle waiting for data. In AI training, the system only moves as fast as its ability to keep all of those chips synchronized.