Let's talk about CPO again!

GPU shortage? That's just half the story. The real AI bottleneck is optical interconnect. CPO market to explode from $160M to $91B by 2028.

Over the past three years, everyone has been scrambling to buy GPUs.

NVIDIA’s stock price has skyrocketed 15-fold in three years. H100 GPUs are nearly impossible to obtain. HBM memory has been in a severe supply-demand imbalance, driving Micron and SK Hynix stock prices sharply higher amid the capital influx. The entire AI revolution’s success or failure has been attributed to one word: compute.

But you might be missing one crucial point — no matter how powerful a single GB300 GPU is, if it cannot communicate at high speed with tens of thousands of other GPUs, most of its computing power will be wasted.

Think of it this way: Imagine an AI training cluster as a massive building where tens of thousands of people work together. The GPUs are the employees. HBM is the notepad on each desk. Cloud storage is the company's archive. Everyone is debating “how many employees we should hire” and “whether the notepads are big enough,” but no one is asking the key question: Are the staircases and hallways wide enough?

A large language model with trillions of parameters splits its training task into thousands of parts, distributes them across thousands of GPUs for parallel computing, and then synchronizes the intermediate results to all GPUs. If the “channels” between GPUs are not wide enough or fast enough, simply adding more GPUs only creates more isolated computing islands.

“GPUs are the brains of AI, but the neural network connecting those brains determines the speed of the entire system. Without optical interconnects, more GPUs are just isolated islands.”

That neural network is the optical interconnect network.