How the AI Ecosystem is Reshaping European Connectivity
Rhoderick van der Wyck, Managing Director, Eurofiber International, answers a few questions about the fast-growing AI ecosystem and its impact on European network expansion, purchasing and design.
Rhoderick van der Wyck, Managing Director Eurofiber International
Does international reach matter more in the era of AI?
Definitely. Today, Eurofiber operates almost 80,000 km of network in Benelux, Germany and France, and our network could not be better placed for emerging cloud and AI requirements.
Power is the key challenge for AI. Today data centers in the Netherlands alone, most of them in the Amsterdam area, consume the equivalent of two million households, and the grid cannot keep up. These power limitations are pushing data centers out from traditional hubs like Amsterdam, Frankfurt and Paris to new “spillover” second tier hubs - places on our network like Rotterdam, Brussels, Lille, Munich and Dusseldorf.
The other factor that is driving this is AI inference; AI and ML production environments. In the coming year inference infrastructure globally will overtake training infrastructure, and by 2030 it will account for 80% of all AI infrastructure worldwide. Inference needs to get huge amounts of data to users quickly and reliably, and requires a new, more distributed, cloud architecture.
Are new connectivity purchasing patterns emerging?
Yes. Hyperscalers and Tier 2 cloud providers appear to be gradually streamlining their supplier base, moving away from intermediaries to dealing directly with fiber owners, which is good news for Eurofiber. There is also a shift from lit network services to dark fiber as data volumes rise.
The move towards dealing direct with fiber owners is mainly to do with complexity, resilience and responsiveness. Many of our larger cloud customers have well over 1,000 connectivity suppliers worldwide, and they are looking to rationalise their connectivity partnerships. Also, by engaging directly with the fiber infrastructure owners, they gain better quality, higher reliability, lower latency and improved SLA's, especially when compared to complex chains involving two or more intermediaries.
Then there is the move from lit to dark fiber. We have observed a strategic pivot by hyperscalers from leasing lit bandwidth capacity to acquiring their own dark fiber and lighting it up themselves. As data levels rise we are offering an increasing number of AI customers a smooth path to scale, moving from, say, 100 GB lit capacity to dark fiber paths that can scale up to 400 or 800 Gigabits on the same route.
Is AI driving network innovation?
Yes. Not only is there heightened demand for brand new fiber, we are also seeing new technologies like Hollow Core Fiber being rolled out.
This is primarily to improve latency. The rise of second tier data hubs requires the extension of customer networks over greater distances. New fiber and the latest transmission equipment are critical to performance in these locations. As we have abundant of spare ducts and subducts on our routes, we are constantly pulling in brand new fibers which provides superior quality of service and supports the latest equipment. Our end-to-end routes are designed for high-power lasers like Raman Type 4 on which customers can run their own DWDM equipment optimized for creating scalable wavelengths and specific traffic patterns.
We have also seen an increasing number of customer requests for test routes for Hollow Core Fiber. This promises a latency improvement of up to 30%, which is equivalent to increasing the effective distance of a standard fiber by 30% while maintaining the same performance.
What impact is AI having on network design?
A huge impact, and this is the area where AI is having the biggest impact on our business. The two key shifts we are seeing are multi-cloud densification at the data center, and rising demand for new uncongested long haul routes.
Data centers are the heart of the AI Ecosystem, which is a collaboration between Tier 1 hyperscalers like Google and Amazon, Tier 2 cloud providers like OVH Cloud and IBM Cloud, and emerging Neoclouds providing GPU as a Service. In order to enable multi-cloud AI, every operator needs the capability to switch between all of the various hyperscaler AI models, creating unprecedented interconnection density. We interconnect hundreds of data centers, as well as owning and operating 11 of our own data centers. Our DCspine product, which connects 95 of those data centers, is proving to be a particularly useful strategic asset for enabling multi-cloud AI. It allows customers to program connectivity to various cloud and service providers on an interchangeable basis and to do that in minutes rather than weeks/months.
At the broader network level, AI anchor customers are helping us to overcome congestion and unnecessary diversion on European networks, by pioneering new "blue ocean" long-haul routes which avoid the highly competitive, "red ocean" congested paths. Our new northern route from Amsterdam to Hamburg bypasses traditional chokepoints. Brussels to Düsseldorf opens up new possibilities for customers looking to avoid congested corridors between Belgium and the Ruhr area. Amsterdam to Brussels and Frankfurt to Vienna are now complete. New links taking data from cables landing in Eemshaven to Mons and from Rotterdam to Frankfurt are underway, and we are constantly exploring possible new routes with key customers, driven by diversity needs.
We are in the fortunate position of having a significant portion of the physical fiber or ducts for these routes already in place, which means that they will mostly require connectivity and strategic partnership rather than extensive new digging. This means we can bring them to market faster, enabling the next phase of AI in Europe more rapidly.