● Training, inference, and resilience: Orange is deploying hybrid solutions and collaborating with datacenter and manufacturer ecosystems to adapt its networks to the demands of AI.
● In response to the rapid obsolescence of GPUs, the operator is prioritizing investment in modular infrastructure and focusing on standards such as those proposed by Mplify to remain competitive in a rapidly evolving market.
● Hello Future talks to Senior Strategic Marketing Manager for AI at Orange Wholesale Jamy Rousseau.
The boom in AI and the emergence of neocloud providers has led to soaring demand for faster connectivity. How is Orange taking up this challenge?
Orange is a key player in datacenter interconnectivity. With regard to AI, we are focused on meeting the needs of hyperscalers (Amazon, Microsoft, Google, etc.), neocloud providers (Coreweave, Nscale, Nebius, etc.) and Chinese players such as Alibaba and BytePlus, the subsidiary of ByteDance (TikTok), who are investing massively in Europe and Africa. Our core business is international backbone services: providing datacenter interconnectivity, with particular emphasis on high-speed bandwidth and network resilience.
Our core business is international backbone services: providing datacenter interconnectivity, with particular emphasis on high-speed bandwidth and network resilience
What are the main technical challenges linked to AI?
Training AI models that rely on data aggregated from various clusters demands massive bandwidth, typically between 400 and 800 Gbps. So we are partnering with leading hardware providers like Nokia and Ciena to deploy high-capacity solutions using WDM (Wavelength-Division Multiplexing) technology. Minimizing cost and latency is critical, as data transfer speeds between datacenters have a direct effect on GPU performance during large language model (LLM) training.
What about innovation in inference?
Inference involves a different set of challenges. We are working with Orange Innovation and OINIS to anticipate future architectures, notably on the possibility of using a service profile for AI such as those defined by Mplify, which aims to standardize the transmission of huge data packets or jumbo frames. The challenge is to be able to guarantee minimal latency and high availability. For example, when a master request is distributed across several AI agents, the overall response time is dictated by the slowest sub-request’s execution. If it is divided into 10 tasks, and one of them takes ten seconds, the overall response will be delayed accordingly. Our mission is to streamline these exchanges.
Certain Chinese models (DeepSeek, Moonshot Kimi K2), which are both open source and high performance but also more energy efficient, have effectively demonstrated that it is possible to reconcile technical sophistication with reduced power use.
How does Orange aim to stand out from the competition, notably in emerging markets such as Africa?
In Africa, we are focusing on partnerships with Chinese and European manufacturers with the goal of providing solutions that are tailored to local constraints, such as access to affordable electricity. The African market is moving towards decentralization. With Edge-based infrastructure, inference can now be localized or routed from Marseille, the central point of convergence for IP flows into North and West Africa, effectively driving down costs and minimizing latency.
How do you expect datacenters to develop between now and 2030?
Between now and 2030, more than 600 datacenters will be built in Europe. The major challenge will be to provide them with access to cheap and sustainable electricity. Orange is deploying fibre with redundant routing to connect this infrastructure and anticipating exponential growth in inference traffic. Today, 50% of GPU workloads are already dedicated to inference, a figure that is set to rise. Players like Mistral and CoreWeave recognize that GPUs which are currently used for training will quickly become obsolete, and they and other neoclouds are eager to repurpose this hardware for cost-efficient inference.
What are the risks faced by operators who are slow to adapt?
The risk is they won’t be able to keep up with the pace of hardware innovation. GPUs that you buy today can lose 80% of their value in 18 months with the arrival of new generation equipment that is 4-5 times more powerful. It follows that investment needs to be focused on infrastructure that can be redeployed quickly. Orange, which according to Kentik is now ranked 10th in the world for IP transit visibility, is prioritizing hybrid solutions: Dark Fiber and WDM for training, and products such as Ethernet and IP Transit for inference.
Jamy Rousseau
