The Next Tech Era Starts With Optics

By: Aliza Goldberg and Isaac Sigron

If you invest in the stock market or have followed the news this year, you’ve probably witnessed the debate over whether AI is in a financial bubble. Public market investors are constantly asking where AI demand is heading. At KDT, we think that question is only half the story. Every major computing shift from PC to mobile, and even AI itself, has been driven by hardware breakthroughs that expanded what was possible, and in turn, unlocked entirely new demand. So, the more pressing question isn’t whether AI demand will grow, but whether hardware can advance fast enough to spark the next demand wave. Just as better compute fueled the first AI wave, the next wave will depend on advances in optical hardware1. Existing copper cables simply can’t move data fast enough anymore, making fiber-optics the only viable path to supporting increasingly large AI models. 

To understand where those optical improvements will come from, it helps to understand how these systems work today. At the simplest level, data is carried as light through optical fibers (thin glass strands that guide light over long distances). How fast that data can move is determined by four key levers: the number of fibers available, the number of wavelengths (“colors”) you can use at once, the speed at which the light pulses are sent, and the amount of data encoded in each pulse. These make up the equation: Data Rate = Fiber Count × Wavelength (Color) Count × Symbol Rate × Bits Per Symbol, and define the limits of data movement in data centers. When any one of them plateaus, model size, training speed, inference cost, and overall AI capability hit a wall. Understanding these levers is key for anyone betting on the future of AI. 

One way to visualize these levers is to imagine ground transportation between two cities. At a high level, there are four practical ways to move more people: (1) build more highways, (2) add more lanes per highway, (3) increase vehicles’ speeds, or (4) increase the number of passengers per vehicle. Each lever increases throughput but comes with tradeoffs. Just as better roads and vehicles lead to more travel, infrastructure itself creates demand. As these four optical dimensions advance, they unlock entirely new waves of technological growth. 

Lever #1: Fiber Count (More Highways): The first lever is fiber count. In our transportation analogy, this is like building more highways between two cities. Each fiber is another parallel path for data, making this the most intuitive way for hyperscalers to increase capacity. But just like highways, adding more fibers isn’t free. Longdistance highways are expensive, space is finite, and maintenance scales with size. Adding fibers increases complexity and cost. Thick fiber bundles are hard to bend, and route and space on chips is limited; long cables are expensive (a single 3-5m cable can cost $100-$300+2), and reliability often declines (since a single fiber failure can take down an entire XPU and require cable replacement).  

Another simple but major physical challenge is connecting fibers to the chip. Optical fibers are relatively thick (~10 µm), while photonic waveguides are much thinner (~1 µm). Precisely aligning multiple fibers to these silicon waveguides is both a mechanical and manufacturing feat, especially when those connections must be detachable, repeatable, and serviceable. Without new technology, AI clusters are essentially stuck with backed-up roads and no equipment to build more highways. One of the leading companies tackling this challenge is Teramount, which builds detachable fiber-array units (FAUs) with self-aligning optical interfaces. Teramount makes it easier to attach the fibers to chips accurately and at scale. In highway terms, Teramount enables you to seamlessly add more highways, freeing up traffic jams across data center racks.  

Lever #2: Wavelengths (More Lanes per Highway): The second lever is wavelength count, or how many colors you use as carrier signals to transmit information. In our analogy, this is like adding more lanes to an existing highway. More lanes free up traffic but also increase complexity, introducing risks of more lane changes, accidents, and the need for better road design and tighter policing. Optical systems behave the same way. More wavelengths raise the risk of interference and distortion unless lasers become more accurate and signals are more tightly controlled. You can see this trade-off in the way wavelengths are deployed today. CWDM (Coarse WDM3) is the common way of using multiple wavelengths today, analogous to using a few wide, well-separated lanes for transportation. It is simpler, cheaper, and more forgiving, but limited in its total capacity. DWDM (Dense WDM) is an emerging technology analogous to squeezing many narrow lanes into the same highway. It enables more throughput but only works if the lanes are both precisely spaced and constantly monitored for accidents (or, in technical terms, if the system includes highly stable, low-drift lasers and better dispersion management).  

To date, the industry has largely standardized on CWDM, favoring lower cost and higher tolerance over maximum bandwidth. This choice made sense in the era of pluggable optics4, where laser costs are spread across only a small number of fibers. However, as architectures evolve toward higher fiber counts, enabled by approaches like CPO5, the economics are shifting. Tighter wavelength spacing is becoming more viable, opening the door for DWDM to scale to 8, 16, or even 32 wavelengths per fiber. Denser wavelengths are becoming essential for scaling GPU-to-GPU bandwidth and keeping pace with exponential compute growth. At KDT, we have yet to identify the right investment here. If you are building new laser technologies, reach out to us. 

Lever #3: Symbol Rate (Vehicle Speed): The third lever is symbol rate, the speed at which data is transmitted on each wavelength (lane). More simply, this refers to the speed at which light can change to carry information, or how fast each car can travel. Historically, data capacity per lane has effectively doubled each generation. The industry moved from 25G to 100G to 200G per lane today, targeting 400G next. However, reaching 400G per lane is difficult because the material used in the modulators (silicon) is hitting a physical “speed limit”.  

One way forward is to build a faster engine. This is where the company Polaris Electro-Optics innovates. Instead of using silicon to build modulators, Polaris uses an advanced liquid crystal material called “FinGlass”. Polaris is not the first to explore new modulator materials. Many are working with alternatives such as TFLN6, InPh7, and others. However, the novelty of Polaris’ approach is that FinGlass integrates seamlessly with silicon. Other materials typically require either an entirely new chip manufacturing process or complex and expensive integration. Since FinGlass can be simply added on top of the silicon chip (later in the manufacturing process), it allows you to replace the engine without changing the entire car. Without a new modulation material like Polaris, symbol rate stops scaling, killing network capacity altogether.  

Lever #4: Bits per Symbol (Passengers per Vehicle): The final lever is bits per symbol, the amount of data each pulse of light carries or the number of passengers per vehicle. Today, optical systems mainly use “NRZ8” and “PAM49” standards (cars that can carry one or two passengers, respectively). These standards encode data by varying amplitude (in a method known as IMDD10). In theory, to transmit more data, you should encode more bits per symbol (adding more passengers per car). However, higher bit levels demand a much cleaner signal, and the system can’t increase signal power enough to maintain the required signal-to-noise ratio. In other words, with more people, each car needs far more fuel, and because the fuel supply is fixed, the car can only travel short distances. 

To reduce congestion, the industry will need to leverage Coherent communication, a more advanced signaling approach that uses phase and polarization characteristics, as well as amplitude. Moving to Coherent is like switching from cars to buses. Buses (Coherent) can carry many more passengers efficiently, but require a more trained driver (complex DSPs11) and wider, better-maintained roads (high-quality or cooled lasers12). Historically, those trade-offs meant that buses were only worth using for long-distance routes (between data centers). As traffic piles up, though, buses are becoming necessary for shorter distances (intra-data center) at a price point comparable to cars. This is what the company Lucidean unlocks. Lucidean is building a simpler Coherent architecture that offers the efficiency benefits of better data throughput, while dramatically reducing the power, cost, and complexity of the system. Lucidean is the first solution to make buses practical for short-distance intra-data-center environments. 

The next wave of AI won’t be held back by compute, but by bandwidth. Copper cables are hitting their limits, and shifting to optical links is the only way to move data fast enough to keep GPUs busy. As such, optical progress is the real driver for what comes next. Unless these limits are solved, model training will slow, cluster costs will spike, and the next demand wave will fail to materialize. Over the past year, we’ve backed Teramount, Polaris, and Lucidean, each addressing a different piece of this bottleneck. As we begin 2026, we are looking for founders pushing these levers further or opening entirely new ones. If you’re building in this space, we’d love to connect. Please reach out via our website or LinkedIn, and together we can help scale the future of AI.  

About the Authors:

• Isaac Sigron leads Koch’s investments in data infrastructure globally with over 10 years of investment experience. Isaac focuses on picks and shovels solutions for the gold rush of our time – the mass adoption of Artificial Intelligence (AI). He is a Managing Director at Koch Disruptive Technologies (Israel), Ltd, an affiliate of KDT.
• Aliza Goldberg is an Investment Professional at Koch Disruptive Technologies (Israel), Ltd., covering data infrastructure – particularly technology in interconnect, power delivery, and hardware systems that support AI workloads.

[1]   For a deeper explanation, see our earlier paper: “The Silent Bottleneck in Data Infrastructure” (KDT Website): https://www.kochdisruptivetechnologies.com/Media/Celestial-AI-Article

[2]   Mouser Electronics, “MTP/MPO Cable Assemblies,” accessed December 21, 2025, https://www.mouser.co.il/c/connectors/cable-assemblies/?series=MTP%2FMPO&utm. 

[3]   Wavelength Division Multiplexing (WDM) is a fiber-optic technology that modulates data over multiple separate carrier signal wavelengths (data signals) over a single fiber. 

[4]   Pluggable transceivers are compact, removable transceiver modules that convert electrical signals to optical and vice versa. 

[5]   Co-Packaged Optics (CPO) is a technology which places optical engines (which convert electrical signals to light and vice versa) in the same package or on the same substrate as high-performance processing chips (ASICs, XPUs). 

[6]   Thin Film Lithium Niobate (TFLN) is a photonic material used to make very high-performance optical modulators. 

[7]   Indium Phosphide (InPh) is a III-V compound semiconductor widely used in photonics. 

[8]   Non Return to Zero (NRZ) is a simple binary signaling scheme where each bit is sent as a constant high or low level for the entire bit period, without returning to zero between bits. 

[9]   Pulse Amplitude Modulation (4-level) (PAM4) is a signaling format that uses four distinct amplitude levels instead of two, allowing each symbol to carry 2 bits of data. 

[10]   Intensity Modulation / Direct Detection (IMDD) is a signaling method where data is encoded in the amplitude of the light and decoded by directly measuring its intensity, without needing complex coherent optics. 

[11]   Digital Signal Processors (DSPs) are specialized processors used in coherent optical systems to correct impairments in the signal (like noise, dispersion, and phase shifts). 

[12]   Cooled Lasers use temperature control (typically via a thermoelectric cooler) to keep their wavelength stable.