Mavenir tripled its spend on the use of AI when it moved its software developers from a licence to a usage model, according to CEO Pardeep Kohli.
“We have 1,600 people using LLMs for coding. We originally started with a seat-based model where we paid x dollars per person per month. Starting from the beginning of June we moved to a usage model and our expenses went up three times,” he says.
That sort of rising cost means it makes sense for larger companies, including telcos, to buy and own their GPU platforms, he said, using open source models . Vodafone’s CTIO Scott Petty, for example, has in the past said that the company would avoid buying GPUs, as it risks locking the company into the generation of product that you have bought, and also cuts you off from the leading model developments.
But Kohli says that GPUs have a useful working life of 7-8 years, well past the cost of ROI on a rent vs buy decision. And he claims that open source models are only 6-7 months behind the “frontier” models.
“If I start doing this for the next one year, I will make one-time investment of maybe four or five million and say I am spending a million a month on usage, I would save 12 million.”
Data integrity
Another reason for telcos to invest in and use their own platforms is to keep their data secure.
Kohli believes that public LLMs have been harvesting his developers’ own data for training – which is something that should not be happening according to the terms of their enterprise usage. He says that models seem to have deep and detailed knowledge of Mavenir’s products (such as the capacity of a specific node) that he thinks can only have been discovered by the model using data disclosed within developer’s prompts. Although he stresses he can’t prove it, in his view it could not have discovered detailed performance data any other way: “Nobody wrote it. I can’t find it anywhere,” he said.
“The reality is your in-house knowledge is getting trained into their brain,” he claimed.
So Kohli suggested that such privacy concerns mean first the operators should be wary of over-relying on public models, but also that they may be better positioned than hyperscalers to host other sensitive enterprise AI workloads.
“I would think they [enterprises] should trust operator more than hyperscalers,” he said.
The argument is particularly aimed at European markets where concerns around data residency and sovereignty continue to grow.
Monetise the AI platform
For Kohli, these are arguments for telcos to invest in their own AI platforms. Kohli outlined a vision in which operators build AI platforms to run their own autonomous networks and then monetise those same platforms by offering GPU capacity, AI agents and sovereign AI services to enterprises.
“If you already are doing it for yourself,” he argued, “why not then offer it to other people?”
To enable this, the company has developed its Integrated AI Platform in partnership with Red Hat. A key feature of the platform is its support for token-based AI consumption, allowing operators to package and bill AI services in a similar way to mobile data plans. The platform provides token metering, charging and billing integration, enabling operators to offer AI services to consumers and enterprises through subscription plans, usage-based charging and service-level agreements (SLAs).
The platform also combines AI model management, intelligent model routing and MLOps capabilities with service assurance, security and billing within a single architecture.
Kohli said that model routing would mean that operators could route only the most complex queries to the frontier models, keeping the bulk of operations on their own open source models.
“If they only do AI to reduce their own cost, they’re going to miss out on a bigger opportunity.”
Tokens as a means to AI monetisation?
While DTW Ignite, held this week in Copenhagen, was awash with Agentic AI-based and autonomous operations use cases and stories, some were asking if monetisation of AI capabilities deserved a higher priority.
Some telcos are beginning to monetise AI capabilities within the communications experience. T-Mobile has launched in-call live translation. Jio in India is offering in-call AI assistance as have the three main Chinese operators – including in-call AI avatars. The Chinese operators also introduced token-based AI plans in May, packaging AI access in a manner similar to traditional mobile data bundles. Deutsche Telekom, SoftBank, and SK Telecom are making significant investments in AI infrastructure but have not yet announced token models similar to the Chinese operators.
While some doubt telco’s ability to innovate and monetise, Kohli points to their customer relationships and billing capabilities. But he also admits that the model to start monetising AI is challenging to business operations.
“The people we talk to, they are the network people,” the executive admitted. “This belongs to the product group or marketing group.”
That reflects a broader challenge facing the telecom industry. Operators have spent years trying to move beyond connectivity into higher-value digital services, often with mixed results.
“I don’t know the answer,” Kohli said, but he warned operators against repeating what he sees as mistakes from previous technology cycles. “If they only do AI to reduce their own cost, they’re going to miss out on a bigger opportunity.”
