How to Plan Networks for AI and 6G

AI infrastructure planning often starts with a technology label: private 5G, edge computing, satellite connectivity, or 6G. That reverses the useful order. A network should be designed around the work it must carry, the failures the organization can tolerate, and the evidence needed before a pilot becomes operating infrastructure.

This matters because AI expands the network beyond moving requests between people and servers. Training systems exchange data among accelerators, storage, and data centers. Inference connects applications to models under latency, privacy, power, and cost constraints. Factories may add cameras, robots, sensors, and control systems whose timing requirements differ sharply from ordinary internet traffic. Future networks may also combine communications with sensing or allow software agents to discover and interact with one another.

China's 2026–2030 information and communications plan provides a useful planning case because it joins 5G-Advanced, prospective 6G, fiber, satellite links, computing infrastructure, and industrial networks in one program. Its value for other organizations is not the scale of its targets. It is the reminder that no single radio upgrade delivers an AI-ready network. Planning has to connect workload requirements, compute placement, standards maturity, security, and operational proof.

Define workloads before selecting network technology

Begin with an inventory of actual workloads, not a forecast of generic AI traffic. Separate at least four patterns. Large training jobs favor high-capacity links among accelerators, storage, and centralized computing clusters. Interactive inference favors responsive paths between users or machines and available compute. Machine vision may generate heavy upstream traffic and require decisions near a production line. Control workloads may send less data but demand predictable availability and latency.

For every workload, record the endpoints, direction and volume of data, peak concurrency, acceptable response time, operating hours, data sensitivity, mobility, and consequences of interruption. Also note whether the workload can queue, degrade gracefully, or continue locally when a remote connection fails. A video application can buffer small performance changes; a remotely controlled machine may not have that freedom.

Translate the inventory into service classes. A training transfer may prioritize aggregate throughput and utilization. An inspection system may prioritize bounded response time and local continuity. Emergency connectivity may trade capacity for reach. Keeping these classes distinct prevents a fast average connection from hiding poor performance for the workloads that matter most.

Include physical and organizational constraints. Industrial assets can remain in service for years, use proprietary protocols, and tolerate little migration downtime. Remote sites may lack fiber or dependable power. A service that crosses carrier, cloud, satellite, and enterprise boundaries also needs a named owner for each handoff and failure mode.

Convert requirements into measurable targets

A useful target describes the service users receive, the conditions under which it must hold, and how it will be measured. Peak laboratory throughput is rarely enough. Track latency distribution rather than a single best result, availability over the required operating window, packet loss, recovery time, geographic and indoor coverage, workload completion time, and the capacity available during congestion.

National construction metrics can help frame capacity, but they should not be mistaken for outcomes. The 2026–2030 plan reported by Xinhua calls for 95 percent 5G and 5G-Advanced penetration, 50 compatible base stations per 10,000 people, 500,000 new 5G-Advanced base stations, and 9,800 EFLOPS of intelligent computing capacity. It also sets 3.8 trillion yuan in cumulative information-infrastructure investment. These are measurable inputs and adoption goals, but they do not by themselves show whether an industrial workload is reliable or whether installed computing capacity is productively used.

Apply that distinction to enterprise plans. Base-station count should be paired with coverage maps, traffic, service quality, and utilization. Computing capacity should specify numerical precision, availability, energy use, network accessibility, and actual workload utilization. Satellite coverage should be paired with connection success, terminal compatibility, capacity, and handoff performance. Investment should be connected to operating cost and recurring use.

Security targets belong in the same scorecard. A network that identifies AI agents, senses objects, or controls machinery expands both the value of its data and the potential impact of compromised credentials. Measure identity coverage, authorization failures, time to isolate a device, recovery performance, and the ability to continue essential operations during a network or provider failure.

Place computing according to the workload

The edge-versus-cloud decision is not a permanent choice for an entire organization. It is a placement decision for each workload. Centralized facilities provide accelerator scale and can improve utilization for large training jobs. Regional capacity can shorten paths while pooling more resources than a single site. Edge systems can keep sensitive data close to its source, reduce dependency on a wide-area link, and serve time-critical inference near factories, applications, or users.

Each advantage carries a cost. Central clusters need high-capacity data movement and may be too distant for tight response requirements. Edge facilities distribute hardware, maintenance, security, and power demands across more locations. Regional systems add another orchestration layer. Poor scheduling can leave one facility congested while another sits underused.

Use a placement matrix rather than intuition. For each service, compare maximum acceptable latency, data-egress constraints, model and accelerator requirements, local resilience, expected utilization, energy availability, maintenance access, and total operating cost. Then define where the service runs normally, where it fails over, and what reduced mode remains available if both locations are unreachable.

The network path must be assessed end to end. Faster radio access cannot compensate for congested backhaul, slow storage, an overloaded inference service, or a fragile gateway. Fiber remains fundamental because base stations, satellite gateways, compute clusters, and enterprise sites all need dependable backhaul. Satellite can extend emergency or remote coverage, but early direct-to-device service should be treated as a complementary layer with different capacity characteristics, not as a substitute for dense terrestrial networks.

Sequence deployment around standards maturity

A 2030 planning horizon does not make every 6G capability available today. The International Telecommunication Union uses IMT-2030 for the standards family commonly called 6G. In March 2026, experts agreed on draft technical requirements for IMT-2030 spanning 20 performance measures and six usage scenarios. Formal approval of those requirements was scheduled for December 2026. Candidate technologies, evaluation, detailed industry specifications, equipment tests, spectrum decisions, and device integration still follow that requirements stage.

Treat those milestones as decision gates. Requirements define what candidate technologies will be judged against; they are not a commercial product specification or a guarantee of field performance. Trials reduce uncertainty but do not prove that equipment will remain interoperable, affordable, and reliable at scale. China's May 2026 approval of selected 6 GHz IMT-2030 trials is evidence of research and validation activity, not evidence of a finished 6G network.

Build a layered roadmap so useful work does not depend on one standards date. Improve fiber and backhaul first where they constrain workloads. Use 5G-Advanced when its capacity, positioning, reliability, or device support solves a defined problem. Add regional or edge compute where measurements justify placement. Use satellite for specific coverage or resilience gaps. Preserve upgrade paths for later 6G equipment without assuming that early designs will remain unchanged. MIIT's earlier next-generation network strategy similarly groups 6G with optical, satellite, industrial, and computing systems rather than treating it as a standalone replacement.

Contracts should reflect uncertainty. Ask vendors which functions depend on draft specifications, which interfaces are interoperable, what can be upgraded in software, which components require replacement, and who carries migration costs if standards change. Avoid tying a production dependency to a demonstration feature without a fallback.

Demand deployment evidence, not demonstration optics

A successful demonstration answers whether a system can work under selected conditions. Production acceptance asks whether it keeps working under ordinary constraints. Test walls, weather, interference, mobility, crowded cells, power limits, maintenance windows, and failures at provider boundaries. For satellite paths, include terminals, gateways, handoffs, and spectrum coordination in the test scope.

Run pilots against a baseline and predefined acceptance thresholds. Evidence can include workload completion time, latency percentiles, service availability, incident frequency, recovery time, energy use, operator effort, and total cost per useful task. For industrial deployments, add output, safety, maintenance, and migration downtime. A pilot should identify who collected each metric and whether the result can be reproduced outside a showcase site.

Scale only when recurring operational use supports the case. Paid adoption, renewals, sustained workload volume, documented reliability, and utilization are stronger evidence than announced coverage or installed capacity. Also record negative evidence: low utilization, repeated manual intervention, incompatible devices, security exceptions, or benefits that disappear once integration and maintenance costs are included.

Network planning checklist

Before approving a pilot or infrastructure commitment, confirm the following:

• Every proposed capability maps to a named workload, endpoint set, owner, and failure consequence.
• Service targets cover latency distribution, availability, loss, recovery, coverage, congestion, and workload completion—not just peak speed.
• The edge, regional, or central placement decision documents privacy, resilience, energy, utilization, maintenance, and cost tradeoffs.
• Radio, backhaul, compute, storage, gateways, and application services are tested as one path.
• Identity, authorization, isolation, data handling, and recovery controls cover devices, software agents, and provider handoffs.
• Standards-dependent features are labeled as requirements, trial functions, candidate specifications, or commercially supported capabilities.
• Vendor commitments state interoperability, upgrade limits, replacement obligations, and migration costs.
• The pilot has a baseline, acceptance thresholds, adverse-condition tests, and reproducible measurements.
• Scaling depends on sustained use, reliability, utilization, and economic benefit rather than construction totals or demonstration success.
• A fallback preserves essential operations if the edge, cloud, terrestrial, or satellite layer fails.

The durable planning principle is simple: make technology follow the workload and make investment follow evidence. 5G-Advanced, fiber, edge computing, satellites, and eventual 6G can all contribute to AI infrastructure, but their value appears only when the complete system meets measurable operational needs.