Fine-tuning provides telecom companies with a cheaper path to GenAI

• Fine-tuning offers telecom companies a cheaper alternative to building GenAI models from scratch
• Yet many telecom companies still build expensive models, despite the availability of off-the-shelf solutions
• Challenges and costs aside, telecom companies like Windstream expect long-term benefits from their GenAI investments

Since ChatGPT essentially caused the Big Bang of Generative AI (GenAI) in 2022, telecom operators have rushed to implement GenAI models into their operations. But the costs of training and running these models can be significant, and GenAI’s return on investment (ROI) remains a topic of debate across industries.

Refining GenAI models has emerged as a more affordable option for telecom companies. Refinement in the context of GenAI models refers to the process of taking a pre-trained model and further training it on a specific, smaller data set to optimize performance for a particular task or domain.

Ishwar Parulkar, chief technologist for Telecom and Edge Cloud at AWS, outlined the cost breakdown for different stages of GenAI model development, noting that pre-training – building a model from scratch – can cost millions. This task is usually performed by specialist providers such as Anthropic and Cohere.

On the other hand, refining existing models with proprietary data is a cheaper alternative, costing only thousands of dollars.

Windstream is an example of a telecom company that opts for fine-tuning. According to Windstream Chief Information Officer Stephen Farkouh, the company relies on Azure OpenAI and GPT-4O, which avoids the high costs of building models from the ground up. Instead, Windstream subscribes to Azure’s AI services, with costs depending on usage and the “complexity and volume of the data we customize,” Farkouh said. This approach saves the company “a significant amount of money and resources.”

New research from SAS shows that 70% of telecom companies are already using GenAI and 89% plan to invest in it within the next financial year.

Despite this widespread adoption, some telecom operators make avoidable mistakes.

A report from McKinsey & Company shows that many telecom companies are still building GenAI solutions from scratch, even though there are numerous off-the-shelf options. Only a third of telecom executives surveyed said they buy off-the-shelf products, which the report said suggests “many telecom companies continue to embrace a do-it-yourself model. This move will likely slow innovation and divert talent from more differentiating use cases, as has happened with other technologies in the past.”

Is GenAI worth the price tag?

GenAI comes with unavoidable costs, regardless of how it is deployed. Even fine-tuning is an iterative process, requiring repeated evaluations through LLMOps/FMOps procedures, which “contributes significantly to the overall cost structure,” Parulkar noted.

Adopting AI at scale also poses other unique challenges for telecom companies, especially when it comes to complex networks, massive data volumes, sensitive data and real-time processing needs.

Managing massive, heterogeneous data from sources such as network logs, customer interactions, and IoT devices is not only complicated but also expensive due to the high storage and computing power required. Real-time decision making, which relies on low-latency data processing and model inference, often requires large instance types or GPUs. In addition, telecom companies must implement strict security measures to protect sensitive customer data, which further increases the complexity and cost of these projects.

Finding the right specialized expertise within the company is also not cheap.

“Telecom and AI expertise is competitive and can be expensive,” Farkouh said.

The payout for telecom companies

Despite these hurdles, companies like Windstream remain confident that their GenAI investments will deliver long-term benefits. Farkouh said the company is “confident on several long-term benefits” of its AI investments, including customer experience and operational efficiency. “Automating routine tasks and predicting maintenance needs will help us save costs and improve service reliability,” he added.

And there are already promising results. According to McKinsey & Company, a European telecom company increased the conversion rates of its marketing campaigns by 40% while reducing costs by using GenAI to personalize content. Another telecom company in Latin America increased the productivity of call center agents by 25%. Furthermore, both companies deployed their GenAI models within weeks – the first in just two weeks, the second in five weeks.

“For an industry with a mixed track record of capitalizing on new technologies and legacy systems that slow innovation, these early results and implementation times illustrate the potentially transformative power of GenAI,” the McKinsey analysts wrote.