Ray Le Maistre, TelecomTV (00:07):
So we're in Barcelona for MWC26. I'm here with Louis Powell, Director of AI Industry at GSMA. Louis, thanks so much for taking the time to join us today. TelecomTV is very focused on AI-native telco development, and the GSMA has announced an initiative here at MWC that is right in our wheelhouse, and that is the Open Telco AI. Can you tell us first what Open Telco AI is and why the GSMA has decided to launch it?

Louis Powell, GSMA (00:36):
The Open Telco AI is trying to provide the industry with models, data, and access to compute and a community so that they can build AI that deeply understands telecoms. At the moment, frontier models are very advanced in many capabilities, mass reasoning. Every day there's a new groundbreaking achievement they have in different parts of the industry and in the wider environment. But with telecoms, they don't understand things like standards. They don't understand root cause analysis. They struggle with documentation and multimodality of telecoms. So what we are doing is calling the industry together to try to build AI models and systems that really work at what we call telco-grade AI. The levels of performance, safety, accuracy, and costs that can really allow operators to adopt AI across their whole spectrum of needs.

Ray Le Maistre, TelecomTV (01:32):
So AI is not new to the telecom sector. There's been a really heavy focus on AI and how AI can be used within the telecom operator community in the past couple of years. And the telcos know, and the operators know how their networks work, what kind of information is needed. So what have been the barriers to these kind of telco-specific models having been developed, say, in the past 12, 18 months? What has kind of stopped that happening do you think so far?

Louis Powell, GSMA (02:02):
Yeah, sure. So a lot of the models are trained on data from the internet. They go out, they use a thing called Common Crawl, they use Wikipedia, they use academic journals to understand different domains. Now for many applications, that's very useful and you've seen the development of those models and how powerful they can be. The telco is a different beast. It's hugely complicated. It's very driven by different vendors, equipment manufacturers, it's fragmented, that data is in different formats and also it's not readily available. So whilst these models have pioneered and moved ahead on openly available data, they've struggled to have access to what telcos really provide, which is that deep network knowledge and expertise. What we are trying to do is break down some of those barriers so we can get the best of AI for the best of telco.

Ray Le Maistre, TelecomTV (02:48):
So there's a lot of partners on board for this initiative, which is great, but AT&T and AMD seem to be playing key roles. What are they bringing to the table here?

Louis Powell, GSMA (03:00):
Yeah, it's really great. As I said, so we've got benchmarks. So we work with multiple partners across the industry to decide what is it we want to achieve. What does good look like? And that is where our partners come in. When it comes to AT&T and AMD, AT&T have worked with us to launch 30 different models, all designed for telecoms. So we collaborated with about 14 different research academic partners across the globe and worked within AT&T to then design a family of models that operators and the industry can use, open source in their systems to improve how their models understand telecoms. So these are complementary to their models. They allow them to deeply understand telco, and that's a whole suite of models they've released. These are called Open Telco Hotel models. They also have released the code, so allowing you to run this on your own infrastructure, whether that be AMD or NVIDIA, so GPU agnostic.

(03:54):
And we've also released the dataset that we used to train that model as well. So end to end, anybody can now go improve their own models. They can use the dataset to build their own models and they can run that on their infrastructure that's required. When it comes to AMD, they are providing the compute to support other operators or other industry players wanting to build open source models. So they can apply through the GSMA and then we can actually introduce them to AMD and they will provide the compute, allowing us to build more models, more data, more capability for this Open Telco initiative.

Ray Le Maistre, TelecomTV (04:26):
That's where TensorWave comes in then, is it? I think it's an AMD GPU, so the largest AMD-focused GPU deployment around at the moment. Is that right?

Louis Powell, GSMA (04:38):
Exactly. So they're the cloud provider that will provide the access, the portal to the GPUs, allowing those people across the globe to have access to those GPUs for open source research in telecoms.

Ray Le Maistre, TelecomTV (04:51):
Okay, excellent. So you're not starting from scratch. You've got these models that you've helped to develop with AT&T. So it sounds like this has been something you've been developing for a while then. I mean, how long have you been working with AT&T on this to get to that 30 model point?

Louis Powell, GSMA (05:10):
So we've been working on the benchmarking for over a year and AT&T have been heavily involved in that piece. They've been working on root cause analysis where they fine-tuned models that outperformed Google Gemini and other frontier models. And that gave us the insight that if we fine-tune small models on telco-specific problems, then we can get significant performance gains as well as reduce the cost whilst maintaining the same accuracy. So that was the initial insight that gave us that if we can do this in an industry, we can really shift the needle and have that sense where open small models can really complement the great work that our hyperscalers and frontier labs are doing. So that's been going on for a year, but really the work kicked off in earnest in August, summer holidays, so really September. So it's been an amazing achievement in the last six months.

(05:55):
I'm very proud of the team.

Ray Le Maistre, TelecomTV (05:57):
I mean, that is not a timescale that the telco industry really is used to. I mean, this is the kind of thing that might have taken four or five years before. So that is pretty impressive. So you've got this starting point of models to kickstart the whole process, but when are you hoping to have more tangible outcomes from other partners and companies coming in and working with AMD to develop themselves? Do you expect this to be rolling out almost imminently and just building over time?

Louis Powell, GSMA (06:31):
Yeah. We also have some launch partners. So NVIDIA have also launched a model with Adapt Key as well, so trained on the Newmont model. We also have models from Khalifa University who have trained a model on radio frequency. So we are seeing an ever-increasing number of open models across the telco domain, so that's very encouraging. So this isn't just the GSMA alone initiative. This is a call to action for everybody to do great work. And we use the GSMA platform to highlight, showcase and make sure that we're not wasting too many resources by overlapping.

Ray Le Maistre, TelecomTV (07:01):
Yeah.

Louis Powell, GSMA (07:02):
I think when it comes to the AMD activity, I think we'll really start to see some activity in sort of April, May time. Collecting the data is some activity and then actually getting things going. Training takes time, but we are very excited to have some tangible results for later in the year with Doha being a big opportunity for MWC Doha being a big opportunity for us to showcase that work. So looking forward to a version two of this conversation then.

Ray Le Maistre, TelecomTV (07:25):
So again, that's a pretty short timescale, but this I think is probably what the industry wants to hear. So applause for that. This is a great initiative. Well, Louis, I'm sure there's a million people wanting to talk to you about this and to find out because this is going to help an awful lot of companies, I think, accelerate progress with what they're trying to achieve. So thanks so much for taking the time to come and talk to us today. Good luck and hope to catch up with you later in the year and find out how things are developing.

Louis Powell, GSMA (07:55):
Pleasure. Thank you for the time.

Ray Le Maistre, TelecomTV (07:57):
Thank you.