“Dumb Pipes” No More: How Telcos Can Capture More Value from AI Services

For those of us who have served the industry for the last few decades, the challenges presented by artificial intelligence might remind us a lot of the broadband wars of 20 years ago, when telecommunications companies invested billions making it possible for customers to download more data at home and on the road yet somehow ended up no more profitable than before they began. Today telcos are still collecting their standard broadband rental every month even as over-the-top services take a disproportionately large share of overall revenue.

But just because the first few minutes of this AI movie look like a sequel doesn't mean it has to end the same way. This time around, telcos could be the hero: the spectrum, fiber, edge nodes, towers, and other assets they have spent decades building could give them real leverage in the AI era, provided telcos deploy them strategically in a way that enables them to become the intelligence layer of the digital economy.

A New Movie

At first glance, the challenge might seem similar from the telco point of view—how will they bring even more bytes from A to B and back?—but in fact, the special nature of AI presents telcos with a huge opportunity.

Amazon Web Services, Azure and the other hyperscalers scaled in a global market, and their business models still depend on global reach. However, they have a problem: not all AI can live in a cross-border cloud. Deploying AI for healthcare, government, and defense requires sovereign data infrastructure, not because hyperscalers are untrustworthy, but because regulated clients require it as a contractual condition. Beyond geography, there is also a performance gap. General-purpose large language models are not experts in everything. In telcos, for instance, they don't even get a passing grade: GPT-4 answered only 64% of 3GPP standards questions correctly in independent benchmark testing. This is below the threshold most standardized tests treat as passing, and the gap persists even in newer general-purpose models.¹

But even if they were better performers, generic solutions are likely to generate generic returns. If every operator runs an identical, off-the-shelf hyperscaler AI stack with no proprietary layer, the ability to differentiate on AI-driven products and services converges to zero. Sovereign AI capability—even a partial, domain-specific layer—creates a level of differentiation that prevents this convergence.

Building an AI Foundry

Fortunately, there is a solution that addresses the shortcomings of standard AI platforms to meet telcos' business needs and allows telcos to pursue a range of opportunities that generic hyperscaler AI, on its own, cannot unlock: a sovereign AI foundry. Telcos need to build a domestic AI service platform that can give consumer, business, and government customers in their own country the secure AI services they need. This is not a choice between hyperscaler partnerships and sovereignty. Instead, it is a hybrid model in which sovereign capability makes the operator a more strategic participant in those partnerships, not a passive consumer of them.

The success of a telco's AI foundry will depend on how well the organization can internalize four mantras:

• Data gravity is a telco's friend. A telco's network edge can process rich network telemetry locally, right where the data is created, effectively establishing a sovereign perimeter. This combination of low latency and strict data residency is exactly what high-sensitivity government and defense applications require. Instead of merely helping hyperscalers transfer data to a global cloud, telcos can anchor a sovereign cloud within their own borders. As Axiata Group's Dr. Rainer Deutschmann has put it in our recent podcast, compute is becoming a commodity while data is becoming king. This is exactly why the model itself is not the moat, but a telco's data, customer relationships, and trust are.²
• The telco AI journey should start with the client. A domain-specific language model (DSLM) can provide better service at a lower cost than a generic LLM. Unlike generic AI, a DSLM is trained on proprietary, carrier-grade data: network telemetry, OSS/BSS logs, and 3GPP standards. By keeping this data within sovereign boundaries, telcos build Long-Term Memory into their architecture. Unlike stateless chatbots, Long-Term Memory creates contextually aware systems that remember interaction histories and network states across sessions, transforming the AI into an anticipatory network manager. Importantly, building domain-specific models and building sovereign infrastructure are parallel workstreams, not sequential ones. Telcos can begin fine-tuning models on carrier data today, regardless of where their sovereign infrastructure investment stands.
• One model doesn't fit all. Some telcos will be best served by adjusting an open-source model. Fine-tuning models like Meta's Llama on operator data and then porting it to their network is a smart way to start, because it will deliver faster inference and lower costs. Other telcos will want to build a custom solution. For markets prioritizing national AI, a model built from the ground up like Tech Mahindra's 8-billion parameter Project Indus may also prove highly scalable.
• Owning the pipe is good. Owning the well is better. As trusted infrastructure, telcos become the essential conduit between the hyperscaler and the national client. Their decades of close relationships with national regulators can make them the preferred partner for sensitive infrastructure projects. Software relationships alone cannot replicate their hard-earned trust. This trusted position opens concrete revenue architectures: Government AI-as-a-Service for citizen platforms, sovereign enterprise cloud for regulated verticals such as healthcare and finance, and GPU-as-a-Service on carrier-grade infrastructure. All of these are best delivered by an operator that retains sovereign control of the intelligence layer, regardless of which compute partnerships underpin it. To further capitalize on this trusted infrastructure, telcos should transform their traditional NOCs into intelligent, closed-loop operations. This transformation is itself a sovereign AI workload: network operations data is among the most sensitive and performance-critical an operator holds, making the NOC a natural first proving ground for the foundry model before extending it to enterprise and government clients. This is a natural starting point for upgrading the "pipe" into an autonomous network because they concentrate high-volume, repeatable tasks in a place where AI can work with them.

These four mantras describe the stack a telco needs to build. But sovereignty is not only an architecture question, it is also a time question. As Dr. Deutschmann frames it in our conversation, an operator can control its compute environment today and still drift into dependency within a few years if it cannot secure its own software updates, licensing terms, and supply continuity. A foundry built on someone else's roadmap is not sovereign merely because it sits on domestic soil; telcos need to plan for sovereignty across both dimensions, the stack and the years, not just the one that is easiest to announce.

Happily AI-After

A domestic AI foundry could be a huge opportunity for telcos. As Morningstar puts it, telcos that move early to capitalize on sovereign AI initiatives will be "best positioned to capture the enterprise/government market share as this part of the industry is set to expand meaningfully over the next five years."³

However, it is important to be clear on what kind of opportunity this is. Dr. Deutschmann has pushed back on the notion that every operator without a sovereign AI play will be left behind in five years, calling that framing overstated: this is a chance for the strongest operator in each market to lead, not a baseline requirement for all of them. The foundry model rewards the telcos that move deliberately, not every telco that makes a move.

Many operators already know this. NVIDIA counts at least 18 operators across five continents (including Telenor, Fastweb, SoftBank, and Swisscom) that had launched NVIDIA-powered sovereign AI factories as of mid-2025, a number that has kept climbing since.⁴ In Germany, Deutsche Telekom has launched a €1 billion industrial AI cloud in Munich, which went live in early 2026, to boost digital sovereignty.⁵ One recent report tallied 53 telcos shifting from connectivity to cognitive infrastructure by buying GPU-dense facilities, training proprietary models, and packaging the result as sovereign enterprise platforms. That broader 53-telco figure tracks the industry-wide shift rather than NVIDIA-partnered deployments specifically, so it isn't directly adding to the 18-operator count above.

Ultimately, this foundry architecture will run on Open Telco principles, including TM Forum’s Open Digital Architecture (ODA) and the GSMA Open Gateway. It will leverage LLMs as a foundation layer, harness the power of lean DSLMs for specialized operational intelligence, and activate LTM to make agentic AI useful and flexible at carrier scale. The result will be a portable, auditable network that allows telcos to scale sovereign AI services across national infrastructures without writing custom code for every new government or enterprise client. By building their own AI foundry, telcos can solve the data sovereignty challenges of governments and enterprises and finally work as true partners with the hyperscalers.

A version of this article was originally published in The AI Journal.