China is projected to account for about 5,400 MW of modeled potential AI grid capacity, highlighting how telcos can turn existing infrastructure into new AI revenue streams
Global technology intelligence firm ABI Research says China is set to dominate modeled potential AI grid capacity with about 5,400 MW, compared with 789 MW in the United States, as telcos seek to turn existing network infrastructure into new AI revenue streams. The finding underscores how infrastructure scale, power availability, and national policy are beginning to shape the next phase of distributed AI deployment.
“The near-term AI grid opportunity is not about putting GPUs everywhere in the network,” said Michael Moreno, Research Analyst at ABI Research. “It is about identifying the workloads where lower latency, data sovereignty, and tighter network integration create measurable business value, then building outward from centralized and regional sites that can support those economics.”
ABI Research’s model shows that early deployments will be concentrated in existing high-capacity sites. Core data centers account for approximately 5,000 MW of modeled potential AI grid capacity and central offices and exchanges represent about 1,500 MW, together making up roughly 98% of activatable capacity across the operators studied. This means the first commercial wave will center on facilities that already have the power, cooling, and connectivity needed for accelerated compute, while cost-per-token and GPU utilization will become critical benchmarks for determining whether distributed AI can compete with centralized cloud infrastructure.
The report also finds that operators have multiple paths to participate in the AI grid value chain, from leasing sites and power to AI providers, to operating their own compute, bundling AI infrastructure with connectivity and network services, and eventually offering broader AI platforms. Vendor strategies are already emerging across the ecosystem: Akamai stands out as an early revenue-generating example with an orchestration-led model spanning roughly 20 centrally located GPU sites, while NVIDIA is exerting outsized influence over AI grid architecture through its control of GPUs, networking, and software, alongside partnerships with companies including Nokia, Cisco, and AT&T.
“Telcos will capture more value only if they move beyond being landlords for AI infrastructure,” Moreno said. “The higher-value opportunity is to combine compute with connectivity, security, traffic management, workload placement, and SLAs, but operators should scale distributed deployments only where customers are willing to pay a premium over centralized alternatives.”
