Orange’s Laurent Leboucher explains ‘AI for network’ vs. ‘network for AI’

As communications service providers (CSPs) speed their adoption of AI, they are grappling with two distinct challenges: using the technology to optimize their own networks and ensuring that their networks can support the dynamic requirements of emerging (and potentially revenue-generating) AI-powered applications. In an interview with Inform, Laurent Leboucher, Orange Group CTO, discusses this dual focus, which he describes as “AI for network” versus “network for AI”. He also stresses the need for cooperation among standards bodies and open-source groups to support end-to-end provisioning, management and assurance of services.

“Often we are very focused on efficiency, and obviously with AI for network the main goal is to become much more efficient on the network side – to automate,” says Leboucher. “But it’s not only about efficiency… It’s also about the new kinds of workloads, meaning services, that we will have to deliver very soon and that will need to be maintained – assured – at scale and on demand.”

Leboucher points to use cases like mobile private networks, quality-on-demand services and real-time video interactions powered by generative AI (GenAI) models like ChatGPT-4, which engages in human-like conversations.

"GPT 4 was really new for two reasons. First, it was multimodal. So, it’s not just about text; it’s about image, voice, and it foreshadows video. And that has a lot of consequences,” says Leboucher. “If you think of video, it’s not just the video that we know, which is streaming video to users. It’s also video uplink…and it needs to be inferred in real time.”

Seeing the opportunity

While AI applications like ChatGPT and Google Gemini operate with latency of around 250-300 milliseconds today, interactive video will require much lower latency of just a few tenths of a millisecond.

“This makes a big difference because in that case the network may become a bottleneck, but it is not yet the case.” Leboucher explains. However, he believes that an access network bottleneck could also become an opportunity: “It could be a solution if we do it properly and if we monetize it.”

As a potential use case, Leboucher gives the example of an AI-enabled virtual guide in a museum. If a patron wants to understand an impressionist painting in-depth, they might use their AI glasses to request a 3D virtual guide to explain the painting.

“So, the virtual guide appears in front of me, and he is able to show different parts of the painting with his finger in real time. At the same time, he is able also to look into my eyes face to face,” Leboucher explains. “I can look at the painting and I can show him something immediately. He can reply and say, ‘Okay, this is what the painter wanted to say here’.”

Leboucher thinks telcos will have an opportunity to provide these types of multimodal AI inferencing capabilities first to B2B customers – providing smart cameras and connectivity to monitor a production line or automated guided vehicles (AGVs) in a warehouse, for example.

“But I think on the consumer side, this is where the volume will come, so we need to be prepared for that,” he says, adding that such services will require “a very dynamic” network cloud factory that is able to spin up, provision, charge and bill for a new network slice that can be assured with a service level agreement (SLA).

AI for network

To address the demands of network for AI, telcos must adopt AI for networks, shifting from static, rules-based network automation to AI-enabled, dynamic orchestration across multiple network domains – from 5G radio access to IP transport and optical core.

“The only way to address this is through the use of AI to detect patterns, to detect when something starts to shift from the normal way of working,” Leboucher says. “AI is the only way to go.”

However, he acknowledges that the business case for telcos to provide edge computing resources to support AI services is not yet clear.

“On the consumer side…it’s still a question of how we as operators can take a key position in the ecosystem,” Leboucher explains. “For the time being, there is a competition of power between the chipset on the device and the capacity in the cloud. There is no clear business case yet for inference at the edge of the network.”

Collaboration on ‘concrete use cases’

To unlock the full potential of AI-driven networks, Leboucher emphasizes the need for extensive cross-industry collaboration on standardizing network APIs, defining common information and data models, and aligning automation efforts across different standards bodies. This must happen in both AI-for-network and network-for-AI scenarios.

On the network-for-AI side, “we need collaboration between telcos, because telcos need to understand how they can unleash the capacity of the network, most probably through APIs,” says Leboucher. Orange is collaborating in the CAMARA Project on such APIs, and in September the company along with 11 other telcos announced a global joint venture with Ericsson to sell network APIs.

Standardizing the right APIs requires “working backwards from concrete use cases,” according to Leboucher. “That means that we cannot just do it between telcos in a closed room. It has to be together with AI players like Google, like Meta, like OpenAI, Mistral and so on – they have to be in the room,” he explains. “Maybe the first definition will be a kind of quality observation or quality-on-demand API; maybe it will be different from the one that we thought about at the beginning, but…it’s really an industry move.”

Collaboration on the AI-for-network side is equally important. Leboucher points to TM Forum’s business-driven, use-case approach as strengths but adds that the organization could do more to align its work on autonomous networks (AN) with global IT automation trends.

“Cloud native is really at the core of the transformation that we are doing," Leboucher explains, noting that he believes TM Forum's AN definition could benefit from more integration with best practices from the IT industry, such as GitOps and other cloud-native methodologies.

Reimagining the Information Framework

Leboucher has advocated for industry-wide collaboration on a “horizontal operating model” to enable automation end to end across domains. “This model needs to leverage a common telco cloud stack, on top of which we implement GitOps at scale and pipelines for network functions and network services chains aggregating different network functions,” he explained in a previous interview.

TM Forum’s Information Framework (SID), which provides an information/data reference model along with a common vocabulary for implementing business processes, might be able to assume this role. However, Leboucher believes the framework needs to evolve to better support the needs of data scientists and AI use cases.

“I think there is now a discussion about: is the SID the right level of abstraction, or is it a bit too restrictive?” Leboucher says. He explains that data scientists at Orange often don’t use the framework, instead relying on their own, non-standardized data models.

This reliance on proprietary models presents a problem because it makes it difficult for Orange to deploy repeatable AI use cases across different regions and operating companies. And it certainly impedes working with partners.

“So, I think there is a need to do something here,” Leboucher says, suggesting that TM Forum could facilitate debate within the industry about how the Information Framework should evolve to better support the needs of AI and data-driven use cases.

Ultimately, Leboucher wants to see more alignment across standards organizations, and he is calling for recognition and coordination of overlapping work across groups.

“I think we should simplify at some point. We don't need everybody to do the same,” he says, pointing to duplication of effort between groups like TM Forum and ETSI on zero-touch and network automation. “Within Orange when we work in forums or standards, we try to keep a very consistent position in the different bodies. But it’s hard because there are too many bodies,” Leboucher concludes.