AI capabilities continue to grow exponentially – and industries are racing to keep up with an insatiable compute demand. According to Epoch AI, the training compute of frontier AI models is expected to grow by up to five-times per year as generative AI and specific use cases like agentic AI and inference AI become standard.

Particularly at the hyperscale level, AI is an arms race and this race does not come without its challenges. Not only is there a need to get networks upgraded and running as needs evolve, but hyperscalers are also faced with the challenge to build in the fiber infrastructure and processing capacity required to realize AI’s full potential. Here’s a look at a few key emerging trends in data center network management that will help operators stay ahead of the curve.

Managing scalability efforts

Increasing network density and capacity to prepare to enable larger high-bandwidth AI clusters (a grouping of GPUs connected by fiber interconnects) is a major factor driving data center construction. As models become more complex, more GPUs are being networked together to create one big cluster. Imagine hundreds of GPUs and fiber interconnects networked together to achieve gigawatt networks. That’s a lot of fiber to incorporate.

Recently, I’ve noticed some hyperscalers creating these networks with dense, distributed GPU clusters, essentially putting more “brains” on the server – a form of scalable growth the industry calls “scale out.” This expansion requires bigger switches, multiplane fabrics, and dense cabling architectures that interconnect GPUs across scalable units, buildings, and even campuses. Historically, an AI node has been within a single server or server rack, but with this model, AI nodes are shifting to stretch across multiple racks – creating another layer of network complexities.

This, of course, creates cabling challenges. Data centers supporting generative AI networks already need more than 10-times more fiber than traditional centers. As a result, the distance to link GPUs within the node increases, eventually causing the links to reach about 100 Gb/s per meter. While traditionally copper has been used in these architectures, with the stacking of scalable units, fiber connections become far more economical – both from a cost and a space perspective.

Distributed AI data centers spotlight long-haul fiber networks

As data center hubs face mounting obstacles around power, heat, and space constraints, hyperscalers are increasingly looking to pretrain large language models (LLMs) using long-haul interconnects (data center campuses connected to each other). As the computing demands for AI shift from building LLMs to applications, there’s a move to disperse AI training models amongst data center locations.

This is thought to network boost performance as each campus would share computations, memory, and power to train one high-bandwidth brain. This distributed model puts an emphasis on the need for low latency, higher bandwidth fiber-optic cabling to serve as the bridge of these data center campuses and facilitate the strenuous data processing that will be required in the future.

CPO technology to increase network processing speeds

Co-packaged optics (CPO) represents another major shift in network design, bringing optics and electronics together in a single package to deliver faster, more efficient data processing for next-generation AI and cloud applications.

By integrating optics directly with the switch, CPO reduces the distance signals must travel before converting to light, lowering power consumption and improving performance. It also addresses the escalating demand for bandwidth by enabling significantly higher port counts, allowing hyperscalers to build larger, more efficient switches without the cost and limitations of today’s pluggable transceivers.

The benefits are clear: lower cost per bit, reduced energy use, and the ability to scale networks far beyond current architectures. As next-generation switches are being designed around CPO, adoption is expected to accelerate in the coming years as hyperscalers push to meet the speed, efficiency, and scalability demands of massive AI workloads.

Rapid AI growth is driving unprecedented demands on data center networks, pushing hyperscalers to scale up GPU clusters, scale out across campuses and regions, and embrace new technologies. Meeting these challenges calls for bold innovation in how networks are designed and interconnected. The path forward is clear: continued innovation and adoption of cutting-edge fiber solutions will be essential to stay ahead in the AI era.