The top data center news of today generally centers on artificial intelligence accelerating in capability. What isn’t so widely discussed are the high-performance digital highways that connect and fuel processors – a fundamental part of AI operation beyond just compute. As processing power expands to meet AI’s needs, networking demands grow right alongside it.

Heidi Adams, head of data center networks marketing at Nokia, describes networking as the critical foundation for the emerging era of AI ‘factories’ and even ‘gigafactories’ capable of producing the next generation of intelligent applications. In her view, the ability to train and deploy AI at scale hinges on networks that are fast, reliable, secure, and engineered to manage congestion while keeping latency to a minimum.

From interconnecting clusters of GPUs inside a single facility to linking vast, distributed AI workloads across multiple data centers, networking is the hidden engine driving AI’s progress – and Adams is here to shed light on the cloud’s best-kept secret.

Networking: The master key to AI factories

When building an ‘AI factory,’ the checklist is vast. Everything – from the physical building, power supply, and cooling systems to the AI chips, servers, storage, and application stack – needs to be accounted for. Each component is of course essential, but networking has a unique role as the connective medium that binds every element together.

Inside the data center, it links GPUs, servers, and storage into a unified system. Outside the data center, it takes facilities from isolated islands to parts of a global AI ecosystem, connecting other sites and delivering capabilities to end users.

With tech giants pouring billions into new infrastructure in the AI era, market dynamics for cloud and data center networking are shifting. Adams notes that by 2030, 70 percent of data center capacity will be dedicated to AI workloads, underscoring the massive demand for both compute and the networks that enable it.

Yet, networking remains a hidden foundation – often overshadowed by the headlines around compute power, renewable energy, or cooling innovation. As Adams puts it:

“If you were to take the cost base, networking is a small percentage of the total build, but without it, nothing else works.”

In the race to build the most powerful AI factories, that percentage could make all the difference.

The three pillars of networking for AI

According to Adams, meeting the demands of today’s AI landscape requires a layered network infrastructure that spans from the data center core, all the way to the farthest reaches of end-user access. She describes three essential pillars of networking for AI:

1. Inside the data center – AI training and compute fabric

Within the AI factory, networks must be high-speed, reliable, and lossless to interconnect GPUs and AI accelerators – critical when training large language models. Adams emphasizes that advanced networking technologies are needed to minimize training time and maximize GPU usage, directly impacting the cost of building AI models.

2. Data center interconnect (DCI)

As AI workloads increasingly span multiple facilities, ultra-low latency, high-throughput connections are essential to link them. These interconnects need to support distributed workloads and factor in the availability of sustainable power, which often influences where facilities are built and how much capacity they can support.

“The network is what turns isolated data centers into a cohesive AI infrastructure,” says Adams.

3. Access and distribution networks

Once AI models are trained and workloads are ready, they must be delivered to enterprises and end users. This requires fast, reliable access networks that provide great connectivity into AI workloads for consumption. Adams points out that without robust last-mile delivery, even the most advanced AI capabilities remain out of reach for those who need them most.

From the compute fabric to inter-facility links and end-user access, these three pillars form the backbone of AI at scale – quietly powering the next generation of intelligence.

Different AI workloads, different networks

AI workloads fall into two main categories, known as training and inference. Each places very different demands on the network. Training is the process of building the large language models that power generative AI, processing tasks involving massive amounts of data that run in parallel, often over many hours or days. This calls for lossless, highly reliable networks to ensure that expensive compute cycles are never wasted.

Inference, by contrast, takes a trained model and uses it to generate answers or predictions in real time. In mission-critical applications, the network must enable rapid, responsive processing to deliver results instantly. Here, latency may not just be inconvenient, but unacceptable.

Beyond the data center, these workloads may need to travel across cities, countries, and even continents, requiring high-speed optical transport technologies to maintain efficiency and performance over long distances.

Locating AI workloads for efficient data centers

The choice of data center location is critical when it comes to operational efficiency. Building facilities in cooler climates can reduce the demand on cooling systems, while proximity to sustainable power sources ensures that the energy-hungry GPU servers that drive AI workloads are supported responsibly.

But location alone isn’t enough – networking becomes a crucial factor in connecting remote facilities to other data centers and to end users, ensuring that sustainable infrastructure doesn’t come at the cost of performance. Adams explains:

“If I look at strategies that Nokia recommends on the networking side, we have a concept that we call the network cloud continuum. It spans from what happens inside the data center to how that facility connects with the wider world. We believe these networks need to be operationally agile to meet evolving demands.”

Given the considerable differences in network access, reliability, and regulations worldwide, the quality of networking should take a more prominent place in discussions of AI equity.

When networks fail, AI training pays the price

As established, inside the data center, lossless networking is a necessity. With the vast quantities of data moving in parallel across lanes and between compute servers and GPUs in AI training – executed in sync for the cycle to complete – if a packet is dropped or a data flow is interrupted, the system resets to its last checkpoint, forcing the process to start again. For training runs that may take days or even weeks, these setbacks have a direct and costly impact.

“Any loss in the network can derail the entire job and send you back hours, even days,” explains Adams.

Congestion is another risk: when network capacity is insufficient, dropped data becomes inevitable. To prevent this, AI networks increasingly use advanced congestion management techniques to ensure information flows smoothly and no GPU cycle is wasted.

Nokia’s push for sustainable networking

Both inside and outside the data center, Nokia is advancing network technologies to meet the extreme demands of AI workloads while improving efficiency.

Inside the data center, this manifests as an evolution of platforms to deliver advanced congestion control and the creation of an essentially lossless environment to prevent dropping data.

Outside, Nokia’s investment shifts to coherent optical engines that enable long-distance links, ensuring high-speed, reliable, and energy-efficient interconnections between data centers.

Leveraging its partnerships with industry groups, such as the Ultra Ethernet Consortium, Nokia is also actively working to evolve Ethernet (which is already the world’s most widely deployed networking technology) to meet AI’s stringent data center requirements.

“Networks have to be fast, and with advances in Ethernet standards, we’re already pushing connectivity from 800 gigabits toward 1.2 terabits and beyond,” says Adams.

Together, these innovations form a multi-pronged strategy to tackle the networking challenges of AI.

Designing AI-ready networks

For Nokia, building networks that can power the AI era comes down to a core set of design principles.

Privacy and cybersecurity are paramount.. With AI models ingesting sensitive data, these principles become foundational requirements, ensuring traffic cannot be tampered with and that connections are fully protected.

Reliability is equally critical in what Adams calls the emerging “AI economy,” where downtime can disrupt communications, emergency services, and commerce. She notes that past outages, such as a nationwide network failure in Canada, illustrate just how crucial it is to ensure cloud networks remain accessible at all times.

To achieve that reliability, Nokia builds automation into its networks to detect and resolve issues before they cause problems. It also embraces the digital twin concept – creating a virtual replica of the network to safely test changes before implementing them in production, reducing the risk of outages.

Emerging tools like AIOps help operators manage networks more efficiently, while secure architecture ensures sensitive data and privacy are protected.

That said, while the industry jokes that networks are “most reliable around the holidays” because inactivity prevents disruption, Adams emphasizes that innovation can’t stop. The goal, she says, is to design and operate networks that can evolve without fear of downtime, ensuring AI infrastructure can keep pace with the technology it supports.

Preparing networks for the future of AI

As AI continues to transform industries, data center operators, cloud providers, and enterprises must recognize that the network is the foundation of the cloud and the emerging AI economy.

Often operating behind the scenes, networks are the unsung heroes that enable high-speed, reliable, and secure AI applications. Adams emphasizes:

“We see the need for these networks to evolve, to make sure they’re really fast, really reliable, and really secure – everything you need to support new and emerging AI applications. We strive to create networks that are adaptable and able to thrive in a rapidly changing environment, no matter what the future holds.”

Evolve your data center network for the AI era. Learn more here.