The Photonics Revolution:
How Nvidia Is Betting
Light Powers the Next
AI Leap

Let us briefly review how we arrived at this point. For years, the general consensus in tech was that software was king and hardware was merely a commodity. Then ChatGPT happened. Suddenly, training and running large language models became one of the highest priorities for every major technology company.

Not only did this flip the narrative on hardware, but it also exposed a fundamental physical constraint. The smarter the AI model, the more data it needs to process, and the faster it needs to move that data around. Hardware is no longer a commodity; it is the bottleneck. And at the center of this bottleneck is the communication between devices.

As AI clusters grow larger, GPUs spread across multiple servers and racks must constantly exchange enormous amounts of data. The distances are longer, the number of communicating chips is vastly larger, and the required speeds are much higher. Copper simply was not designed for this regime.

To understand why copper is failing, we first need to look at the physics of electrical signals. At high frequencies, electrical current does not travel through the entire cross-section of a copper wire but increasingly concentrates near the surface. This phenomenon, known as the skin effect, causes resistance to rise and signal power to be absorbed as heat.

To put numbers on it: at 56 GHz, a standard copper trace loses around 10% of its signal per centimeter. By the time a signal travels even a few tens of centimeters at these frequencies, it has weakened so severely that you need expensive, power-hungry amplification just to recover it. Furthermore, all that signal energy absorbed by the copper does not disappear but becomes heat. Modern data centers already consume gigawatts of power, a significant portion of which goes into just moving data over copper rather than computing with it.

At the speeds modern AI clusters demand, copper is becoming the weakest link in the entire architecture. This manifests across three distinct domains:

Interconnects in data centers can be broadly divided into three domains: scale-up, scale-out, and scale-across. Scale-up connects GPUs directly to one another within a single server. Scale-out links servers through switch ASICs within the same data center. And scale-across connects one geographically separated data center to another.

While scale-up is currently still dominated by electrical interconnects like NVIDIA's NVLink, the thermal burden is exploding as the number of GPUs per server increases. Scale-out already requires optical transceivers, but the "last mile" copper stub between the transceiver and the switch ASIC is the new bottleneck. Scale-across has relied on optical cables for decades, but the demand for bandwidth is pushing the limits of coherent optical communication.

The answer to the copper bottleneck is the optical transceiver. This device converts electrical signals into optical signals for transmission, and converts received optical signals back into electrical signals. The core optical engine inside this transceiver is the Photonic Integrated Circuit (PIC).

However, there is a problem. The switch ASIC, which routes data between GPU servers, still operates on electrical signals. This means that optical signals arriving through optical fiber must be converted into electrical signals before reaching the switch ASIC. The length of this copper segment, which is the distance between the transceiver and the switch ASIC, has a significant impact on scale-out link performance. Light has solved the long-distance segment, but this last electrical segment remains a fundamental constraint.

To address this, the industry is fiercely competing to reduce the physical distance between the switch ASIC and the transceiver. We are currently witnessing a transition through four distinct approaches:

A natural question arises: if light is so superior, why are we still converting electrical signals to light and back again? Why not simply build an all-light GPU? The answer lies in the fundamental band structure of semiconductors.

Silicon, the bedrock of all modern compute, is an indirect bandgap material. When electrons lose energy in silicon, most of that energy is dissipated as heat rather than emitted as light. In contrast, the III-V compound semiconductors used in lasers, such as Indium Phosphide (InP) and Gallium Arsenide (GaAs), are direct bandgap materials that convert electricity into light far more efficiently.

Ultimately, a silicon-based PIC must receive light from a laser built with III-V semiconductors. This creates a massive integration challenge. The industry has developed several approaches to connect these fundamentally different materials into a single system. Hybrid integration, where a separately fabricated III-V laser is aligned and coupled to a silicon PIC, remains the current mainstream approach powering today's AI data centers. Heterogeneous integration, which bonds a III-V wafer directly onto a silicon PIC and forms the laser through lithography, is rapidly gaining attention as the next step to eliminate alignment costs and optical losses.

Companies like Ayar Labs are pushing optical I/O (TeraPHY) directly into GPU servers to replace copper NVLink for Scale-Up, but the core compute remains electrical. The job of photonics in the short-to-medium term is to accelerate electronic compute and connect it more efficiently, not to replace it.

According to Goldman Sachs, the total addressable market for photonics in AI infrastructure could reach $81 billion by 2027. But to understand the immediate urgency, we must look at the supply chain.

As investment in AI data centers has exploded, the supply of the core light source used in optical transceivers, specifically Electro-absorption Modulated Lasers (EMLs), has struggled to keep up with demand. Manufacturing 200G/lane EMLs is extraordinarily difficult, and only a handful of companies worldwide can produce them at scale. Due to this severe shortage, some AI data center EML allocations are already booked well beyond 2027.

No matter how many GPUs NVIDIA produces, without the optical modules needed to connect them, a data center cannot be completed. And at the beginning of every optical module is a laser. This is precisely why NVIDIA announced strategic investments of $2 billion each in Lumentum and Coherent in March 2026. These were not simple equity investments. Both deals included commitments to build new manufacturing facilities and long-term purchase agreements, effectively locking in future supply chains.

The photonics supply chain for AI compute is a layered stack. Understanding where each company sits in that stack is essential to understanding both their opportunity and their risk.

Every single company listed above is genuinely necessary for what Nvidia is making a reality. The market is large enough for all of them to generate meaningful revenue. But there are significant risks that keep us up at night.

The technology is real, the market is validated, and the demand driver is enormous. That should be enough. It rarely is. Let us look more closely at what keeps us up at night.

All of the above leads us to a view that sits, deliberately, in the uncomfortable middle ground. We are not going all-in. We are not staying out. And we think both of those positions, held without nuance, would be mistakes.

We align with this framing entirely. These technological developments will happen sooner rather than later. The money to be made is real. But we cannot responsibly ignore the consequences of an oil shock and a data center power wall.