Demand for B2B fiber optics is skyrocketing due to AI, but according to Eurofiber, there is no cause for alarm. Those who anticipated capacity years ago can comfortably handle today’s exponential growth.
“In the fiber optics sector, things have effectively changed in the past two to three years,” observes Hans Witdouck, CEO of Eurofiber in Belgium. You can certainly call the boss of the fiber optics specialist an expert. Witdouck started his career more than 30 years ago with a preliminary study for the launch of Telenet. Today, he sees a number of important shifts, partly driven by AI, but certainly not exclusively.
The first themes are independent of it. Witdouck notices an increased interest in sovereignty. This is not just about data sovereignty, but also about the network itself. “Who owns the network, and can we maintain full control over it? These are questions that customers are asking today,” says the CEO.
(Late) sounding alarm bell
Following this, Witdouck sees the impact of the changed geopolitical context. “Events such as the severed undersea cable in the Baltic Sea have opened many parties’ eyes,” he notes. “Yet those eyes should have been open for longer: on land, cables are pulled apart almost daily during construction works.”
Nevertheless, the incident has rung a bell for companies that were previously resting on their laurels. They realize that critical business processes might be running in a data center that may still be insufficiently connected. “Where we used to have to draw people’s attention to the dangers of being connected with only one leg, those are much clearer today,” says Witdouck.
In other words, being well-connected is about more than just connection speed. In a context where connectivity to customers and internal applications is essential, a redundant connection ensures business continuity and digital resilience.
AI and exponential growth
In the context of the network, we also cannot ignore AI. It goes without saying that the hype surrounding generative AI is also being felt in the segment that links IT environments together.
Eurofiber first encountered this through the hyperscalers, with whom the company has long done business. “They used to come by, for example, to connect two important data centers over a long-haul route,” Witdouck illustrates. “Then it was usually a request for a single fiber: two pairs or at most four. Today, those same parties come with requests that are a factor of ten or even a hundred higher.”
Better too much than too little
The jump initially surprised Eurofiber a bit. “It could be a sign that they themselves don’t know exactly where it’s going, but that they want to be prepared when the capacity proves necessary,” says Witdouck.
That interpretation is in line with what we see happening worldwide. Hyperscalers and AI specialists are rapidly building AI data centers tailored to a demand that is yet to come.
Pipes within pipes
Despite this drastic increase in demand, Eurofiber’s network is ready for it. There are already plenty of fibers waiting for customers, but expanding is also no problem. When constructing the network, Eurofiber provided pipes through which engineers can immediately blow extra fiber capacity, without requiring major new excavation work.
Witdouck clarifies: “The backbone contains multiple pipes, or ducts, and within those ducts are subducts. Where until recently 72 fibers were blown per subduct, that is now 144 to 288. So the basic capacity is already there. The major work, such as digging, applying for permits, and coordinating plans with involved parties, is behind us. The actual blowing in of fibers happens over sections of about two kilometers and is a lot faster than building entire routes from scratch again.”
Virtual fibers
On top of that physical capacity sits a virtual layer. Eurofiber places different light frequencies on each fiber, where each ‘color’ can carry its own hundreds of gigabits. This way, a kind of virtual fiber is hidden within each fiber, and the theoretical upper limit is far from being reached. Witdouck is clear: ‘Demand may rise exponentially; what is already in the ground can handle that growth quite comfortably.”
Demand may rise exponentially; what is already in the ground can handle that growth quite comfortably.
Hans Witdouck, CEO of Eurofiber Belgium
According to Witdouck, the fact that Eurofiber is so well-prepared stems from a conscious philosophy. As a pure B2B player, the company offers direct connectivity, with one fiber per customer or per connection between two points. Anyone working this way must provide ample capacity from the start, and so there are ducts and subducts ready that have not yet been fully blown.
Training versus inference
Witdouck also notes that not all AI is the same. “A first question is what the network is being used for: is it for training large language models, or for using applications built around AI?”
With training, the traffic is completely different in nature. “It involves traffic of varying intensity, with occasional peaks,” he says. “At certain moments, enormous amounts of data are uploaded to the processing data centers. The capacity required for this is exceptionally large, but only temporary. Latency plays a less significant role in this.”
Training large models happens mainly in the United States, less so in Europe. Yet Eurofiber also sees direct demand here. “It’s not about the very largest models, but for applied LLMs,” Witdouck clarifies. “In the healthcare sector, for example, models are continuously trained. There, demand arises for high bandwidths between the location where those images are stored and the place where the model is built.”
Not so different from cloud
With inference, the actual use of the models, it’s different. There, latency suddenly does play an important role. Witdouck: “In itself, a single request to a large language model doesn’t send that much data back and forth, but the aggregated traffic from all users certainly adds up.”
A single request to a large language model does not send that much data back and forth.
Hans Witdouck, CEO of Eurofiber Belgium
Technically, that doesn’t really bring new challenges. Structurally, inference differs little from consuming other cloud applications. Witdouck makes the comparison with content networks: “Frequently used content, such as a popular series on Netflix or VRT Max, is placed in a cache close to the user to avoid latency.”
Closer to the customer
Something similar is now happening with AI-native applications that are widely used: they are moving closer to the user. The only question is what ‘close’ means in a small country. “In Belgium, Brussels with one data center is often already edge enough,” says Witdouck. In the US, you have to think per state for that.
More than the popularity of AI itself, the issue surrounding data centers will determine how the demand for network capacity continues to evolve. The question is how those data centers can continue to grow in a world where there are increasing restrictions on available power.
Witdouck expects that data centers will be built where power is still available. “That will no longer be in Brussels, where everything is concentrated,” he predicts, “but in more remote areas. The price for that is that those many smaller data centers must be hyper-connected so that they form a single whole together.” This is exactly where the network again plays a key role.
Meanwhile, Eurofiber continues to look ahead, beyond AI. “We are also already looking at quantum technology,” says Witdouck. For example, Eurofiber has been working on encryption with Quantum Key Distribution for several years. The intention is to again be ready with sufficient capacity when another technology creates additional demand on the fiber optic network tomorrow.