Overview
Figures released by the Zhejiang communications administration and the provincial department of economy and information technology sketch the infrastructure base of a regional AI industry. Provincial backbone networks now have full 200G OTN optical transport coverage, with 162 collaborative AI projects under way. Revenue from the core AI sector above designated size reached 683.9 billion yuan in the first three quarters of 2026, up 27.2% year on year. The province has built 113 computing centres with combined capacity above 198 EFLOPS, 95% of it intelligent computing, plus five ten-thousand-GPU clusters.
How that capacity is consumed matters more than the total. In traffic management, robots have logged over 10,000 operating hours, cutting peak-hour police workload by 30-50% and issuing 310,000 warnings for traffic violations; the approach has spread to eight cities within Zhejiang and more than 30 locations nationwide. Operators say it has moved beyond demonstration into a replicable, routine human-machine coordination model. In rural last-mile logistics, a general-purpose unmanned delivery platform navigates narrow lanes shared by pedestrians and vehicles, lowers its deck to ground level on arrival, and dispatches an onboard quadruped robot for the final leg from vehicle to door. That shared-capacity design of cloud compute, unmanned vehicle and robot dog relies on cloud scheduling rather than full onboard computation.
Tourism offers a different commercial validation metric. AI glasses provide plant commentary, route guidance, multilingual explanation and real-time Q&A; since launching in July, weekend rental rates at one park have exceeded 90%, and similar deployments now cover more than 200 projects nationwide. The experience depends on full 5G-Advanced coverage. In healthcare, a national AI application pilot-test base has been set up around compute, data, models, evaluation and commercialisation, addressing three long-standing bottlenecks - factor availability, scenario validation and rollout - by carrying the verification stage between lab prototype and clinical deployment.
Background & Interpretation
1. Technical Background and Evolution
For the past two years investment concentrated on model training and cluster construction, while applications largely remained isolated pilots. The change is that compute, networks, models and scenarios are now planned as one chain: 200G OTN and 5G-Advanced handle data backhaul and real-time interaction, ten-thousand-GPU clusters supply centralised training and inference, and pilot-test bases fill the missing verification stage between laboratory and scale deployment. Output from compute investment is therefore no longer measured only in EFLOPS but in how much capacity is continuously consumed by real workloads.
2. Core Drivers and Underlying Mechanism
The economics come from task layering. Perception, recognition and path planning sit in cloud or edge centres while the machine retains only essential safety control and actuation, cutting per-unit hardware cost and power draw. High-frequency, fixed scenarios such as traffic, logistics and guiding also amortise development cost by reusing one model and scheduling platform. Conversely, real-time interaction is latency-sensitive, so guided tours and remote dispatch depend heavily on 5G-Advanced coverage quality - which is why communications infrastructure is counted within AI project statistics.
3. Market Structure and Industry Chain Impact
Value is migrating from selling compute to selling usable capability. Computing centres need higher utilisation rather than pure expansion; model vendors compete on industry adaptation and inference cost; integrators take on scenario decomposition, data governance and operational response. Public platforms such as pilot-test bases lower the barrier for smaller teams to obtain evaluation and compliant data, but divergent regional standards, datasets and test protocols may force repeated validation when solutions cross provincial boundaries.
Implications & Outlook
Near term, watch three indicators: actual utilisation of intelligent computing centres and the share of inference load, human intervention frequency for unmanned systems in open environments, and the volume of verification outputs the pilot base delivers to clinical or government settings. Medium term, replicability matters - whether coverage across eight cities and 30-plus locations comes with comparable operational stability, and whether shared-capacity and human-machine models can be written as priced service contracts. Long term, if compute, networks and validation platforms are planned together, regional advantage will depend more on scenario density and data governance than on cluster size.
Takeaways for Industry Participants
Compute operators should disclose utilisation, inference latency and cost per task so downstream buyers can estimate real outlay, rather than pitching peak capacity alone. Application developers should split cloud and edge responsibilities by scenario, reserve network redundancy for latency-sensitive functions, and use intervention rate rather than demo success rate as the acceptance metric. Public platforms and standards bodies should prioritise mutual recognition of evaluation protocols, data formats and liability boundaries to cut repeated cross-region validation. Procurement teams should budget network upgrades, model fine-tuning and long-term operations together, and write routine operability into contracts instead of leaving it at the pilot-description stage.
This column compiles industry information and shares technical perspectives; it does not constitute investment advice.