分布式推理：人工智能如何增强边缘

The GSMA is a global organisation unifying the mobileecosystem to discover, develop and deliver innovationfoundational to positive business environments and societalchange. Our vision is to unlock the full power of connectivityso that people, industry and society thrive. Representingmobile operators and organisations across the mobileecosystem and adjacent industries, the GSMA delivers forits members across three broad pillars: Connectivity for GSMA Intelligence is the definitive source of global mobileoperator data, analysis and forecasts, and publisher ofauthoritative industry reports and research. Our data coversevery operator group, network and MVNO in every countryworldwide – from Afghanistan to Zimbabwe. It is the most GSMA Intelligence is relied on by leading operators,vendors, regulators, financial institutions and third-partyindustry players, to support strategic decision-makingand long-term investment planning. The data is used as We invite you to find out more atgsma.com Our team of analysts and experts produce regular thought-leading research reports across a range of industry topics. gsmaintelligence.com info@gsmaintelligence.com Published March 2025 Authors Tim Hatt, Head of Research and ConsultingPeter Jarich, Head of GSMA Intelligence This report was produced with the support of NVIDIA and Dell Technologies Contents Executive summary 1Context for AI in telecoms 1.1An AI-first world1.2AI objectives: low-hanging fruit versus slower burns 2AI along the network edge: where and why 2.1Defining the edge as a continuum2.2Two phases of AI: training and inference2.3Edge advantages: why now? 3Use cases for AI inference at the edge 3.1Mapping the use cases for inference3.2Prioritising early wins 4Upstream versus downstream: network requirementsand impacts 5Implications and outlook 5.1IoT driving demand for edge185.2The operator view of priority industries205.3AI’s revenue moment21 Executive AI offers a new dimension at the edge Edge compute has been a key part of the 5G value proposition for telecomsoperators selling into industries for several years – even before AI. However, AIadds a new dimension to the value of edge through distributed inference, with •speeding up AI inference by reducing compute time•ensuring data sovereignty and security•improving energy efficiency•offering deterministic service and resiliency •providing financial benefits –cost savings (the total cost of ownershipversus cloud) AI is driving, and will continue to drive, an upwardinflection in mobile data traffic growth from consumerand enterprise customers. How big an impactremains to be seen. However, even before AI, GSMAIntelligence analysis indicates that the public cloud and across the network. The rise in AI-driven traffic andworkloads means networks must be AI-ready. Within telecom networks, the best location of AIinference depends on the nature of the service, thecomputing and bandwidth requirements, and thespecific benefits to customers. Figure 1 shows whereworkloads for top use cases could be processed. Having compute demand closer to end customers – on-premises or at the device edge – means operators willhave to move their own processing capabilities to adapt, Figure 1 AI edge use cases: priorities and potential The priority AI edge use cases that leverage inferenceinclude robotics, cameras, digital twins and other IoTapplications. This chimes with the 5G sales push toenterprise clients that is now well established. Operatorssee AI as a tool to allow them to expand their enterpriseservice offering. The use cases shown in Figure 1 are technologies and supporting ecosystems. Pre-emptivenetwork maintenance and engineering troubleshootinguse cases within an operator network are not revenue The choices between use cases are not mutuallyexclusive, and we expect all to be adopted to a greateror lesser extent. The differences will come down to the The AI edge story should be couched in termsof broader telecoms sector activity around AI.Momentum in AI deployments has continued in2025, building on the surge in 2024 as the utility ofthe technology became clear and the competitivenecessity of using it emerged. There is now aconsensus that AI is a must for companies across the the largest operators. However, the levels have notyet reached the point where they have significantlyboosted overall operator revenue growth. Much of theearly AI experimentation reflects a desire to redresssluggish 5G monetisation results. Selling GPU as a While edge was a niche topic 10 years ago, it is nowmainstream for enterprise IT and connectivity buyers.The addition of AI in the mix gives operators addedcapability to significantly improve the overall valueproposition. However, it also raises the urgency to act Revenue is the top goal, with nearly 50% of operatorsrating it a priority from AI. This reflects the sectorcontext of needing to restart growth. Enterprise wasthe original promise of 5G. 5G enterprise revenu