The Network
That Runs Itself

The analysts agree on where AI in telecom starts and disagree wildly on where it goes. Grand View Research puts the 2025 market at $4.6 billion, growing to $46.2 billion by 2033 at a 32.5% CAGR (Grand View). Fortune Business Insights starts almost identically at $4.73 billion for 2025 — and lands at $88.1 billion by 2034 (Fortune BI). GM Insights sees $45.1 billion by 2034; The Business Research Company sees $27.3 billion by 2030 (GMI · TBRC).

Same starting line, forecasts that differ by a factor of three. The reason is scope: “AI in telecom” can mean software an operator buys to optimise its radio network, or the GPU-as-a-service business an operator sells to everyone else. The first is a cost line. The second is a new industry. Analysts who count only the first land near $27–46 billion; those who count the second push past $80 billion.

That distinction — AI for the network versus AI on the network — is the thread that runs through everything below.

In February 2025 Deutsche Telekom and Google Cloud announced the RAN Guardian: an AI agent built on Gemini in Vertex AI that analyses radio-access-network behaviour, classifies degradations across monitoring, inventory, performance and coverage data, and executes corrective actions (Deutsche Telekom, Feb 2025). By November it was live in the production network, which DT describes as being “the first network operator to rely on a highly developed AI agent in network management” (Deutsche Telekom, Nov 2025).

The economics behind it are blunt. GSMA Intelligence reports that energy accounts for 15–20% of a typical operator’s operating spend, and that AI-driven RAN sleep states and cooling optimisation cutting energy use by 20% translates to roughly 4% opex savings (GSMA Foundry AI Use Case Library). For an operator with a $10 billion cost base, that is $400 million a year from one use case.

The more radical idea is AI-RAN: running the radio network and AI inference on the same accelerated hardware. SoftBank’s AITRAS hosts both workloads on NVIDIA Grace Hopper platforms, is being deployed across SoftBank’s own commercial network, and is planned for sale to other operators from 2026 (SoftBank, Nov 2024). The pitch: a base station that is idle 60% of the day becomes an inference node the other 60%.

Network AI gets the keynotes; contact-centre AI gets the receipts. The GSMA’s Global Mobile Trends 2025 cites Deutsche Telekom’s AI chatbots cutting customer resolution times by 40% (GSMA via TechInformed, Mar 2025). KDDI’s generative-AI chatbot reduced response time by 20% (GSMA, Nov 2025). Vodafone’s TOBi has handled over 13.5 million conversations with 65–80% fully automated, per industry reporting.

In the Gulf, e& UAE is the most transparent. Its joint white paper with Open Innovation reports results across more than 1,100 AI use cases, including fewer network incidents and shorter repair times, underpinned by sovereign multi-GPU infrastructure and an AI governance platform aligned to UAE PDPL and EU GDPR (e&, Oct 2025). Earlier disclosures put 160+ machine-learning models in production across customer segmentation, autonomous networking and fraud detection (TeckNexus).

McKinsey’s December 2025 survey of 49 telco executives finds that most operators already use AI in marketing (copy, promotion design, campaign execution) and are extending it into frontline sales and contact centres — and estimates that leaders can lift both ROIC and EBITDA margins by up to ten percentage points within five years (McKinsey, Feb 2026).

The Gulf operators are running the most aggressive version of the second strategy: not using AI to run the network, but using the network to host AI. In December 2025 stc group’s center3 and HUMAIN announced a joint venture to build AI data centres in Saudi Arabia with capacity of up to 1 GW, starting with a 250 MW first phase; HUMAIN holds 51%, stc 49% (Data Centre Magazine, Dec 2025 · DCD, Mar 2026). center3 already operates or is developing some 20 data centres in Saudi Arabia and Bahrain, targeting 300 MW by 2027.

HUMAIN itself — the PIF’s AI subsidiary launched in May 2025 — is building AI factories of up to 500 MW with several hundred thousand NVIDIA GPUs over five years, beginning with an 18,000-GPU GB300 supercomputer, and states a decade target of 6.6 GW (NVIDIA, May 2025). Ooredoo struck its NVIDIA deal in June 2024 to deploy GPUs across 26 data centres in Qatar, Kuwait, Oman and beyond (Fortune, Aug 2026). Mobily announced investments exceeding SAR 3.4 billion in data centres and submarine cables (e&, Jul 2025).

Is there revenue in it yet? The best public data point is Ooredoo’s Indonesian affiliate: its GPU-cloud business earned $33 million in the first half of 2026, more than all of 2025, with $1.2–1.5 billion contracted for the coming years (Fortune, Aug 2026). McKinsey sizes the telco-addressable GPU-as-a-service market at $35–70 billion by 2030 (McKinsey via IEEE ComSoc).

Two things temper the story. First, size. McKinsey’s estimate of generative-AI impact across technology, media and telecom puts telecom at $60–100 billion a year — against $80–130 billion for media and $240–460 billion for high tech (McKinsey, Beyond the Hype). Telecom is the smallest of the three. That is why the Gulf pivot to compute hosting matters: it is an attempt to move from the $60 billion column into the $460 billion one.

Second, energy. The same GSMA research that credits AI with cutting RAN energy use warns that AI workloads drive new demand in data centres, and its Q4 2025 report concludes operators must now treat energy as a strategic priority (GSMA Intelligence, Q4 2025). An operator that saves 4% of opex on its radio network while building a gigawatt of GPU halls has not reduced its energy exposure; it has multiplied it and changed who pays.

What is documented today, in dated corporate and analyst sources: RAN Guardian live in production, e&’s 1,100 use cases, DT’s 40% resolution-time cut, the stc–HUMAIN 1 GW JV, Ooredoo’s $33 million half-year of GPU revenue. What is not yet documented: a Gulf operator publishing the realised P&L value of its AI programme the way Aramco did for energy. That number, when it appears, is the one to watch.