All-In: Telecom giants build national AI data centers instead of distributing GPUs across networks

Mobile operators worldwide are investing in AI infrastructure to create nationally-focused GPU-as-a-Service offerings, marking a strategic shift in how telecom companies approach artificial intelligence services. Rather than distributing GPUs throughout their networks for low-latency applications, carriers like SKT, DOCOMO, Verizon, and Swisscom are building centralized AI data centers designed to provide sovereign AI services within national borders. This approach represents both a departure from unsuccessful “walled garden” strategies of the past and a recognition of the realistic timeline for AI applications that genuinely require network-edge processing.

The big picture: Major telecom operators across the globe are establishing centralized GPU data centers to offer AI services, but they’re not distributing this computing power throughout their radio access networks (RAN).

• Companies including SKT, DOCOMO, Verizon, T-Mobile, Singtel, Swisscom, Telenor, and Softbank have all begun investing in GPU infrastructure.
• These investments are focused on creating nationally-centralized AI data centers rather than distributing GPU computing across cellular networks.

Key insight: Very little of the potential GPU-as-a-Service market requires the ultra-low latency that would necessitate placing GPUs throughout the cellular network.

• Market analysis reveals minimal revenue opportunities for AI services requiring network latency below 100 milliseconds.
• Promising applications in augmented reality, automotive, and remote machinery that could drive sub-100ms AI revenue are likely to mature in the 2030-2035 timeframe.

Behind the strategy: Telecom operators are primarily pursuing “sovereign AI services” that keep data processing within national boundaries rather than ultra-low latency applications.

• Carriers like Swisscom, Telenor, and Singtel are positioned to succeed by offering AI services that respect national data sovereignty requirements.
• This approach allows operators to leverage their natural advantage in providing services precisely within a country’s borders.

Economic reality: The return on investment for distributing GPUs throughout the radio access network remains challenging.

• Implementing distributed GPU computing would require significant investments in Grace Blackwell servers, fiber upgrades, and power/cooling infrastructure overhauls at central office sites.
• While there are potential energy savings from centralized RAN pooling, these benefits plus new revenue from localized services may not justify the massive infrastructure investment.

Reading between the lines: This approach echoes previous unsuccessful “walled garden” and “edge computing” initiatives, but with a more focused and potentially viable business model.

• Rather than attempting to create proprietary ecosystems or unnecessarily distributed computing, operators are focusing on the specific national-level services where they have natural advantages.
• The approach acknowledges market realities while carving out a potentially sustainable niche for telecom operators in the AI services landscape.