China Telecom Computing Power Network Push Links AI Capacity to the Grid

China Telecom has proposed three coordinated moves for its computing power network, despite major technical and commercial barriers between regional infrastructure operators.

President Liu Guiqing presented the plan during the China Computing Power Conference, held September 11 through September 13, 2026, in Langfang, Hebei. His proposals cover electricity coordination, computing resource interconnection, and better utilization of domestic AI hardware.

The speech was more than another call to build data centers. China already has substantial computing infrastructure, but those resources remain divided by geography, hardware architecture, interfaces, and operating rules.

China Telecom now wants to turn that collection into a schedulable national service. That goal pressures other carriers, cloud providers, power companies, and hardware vendors to support common mechanisms instead of isolated capacity pools.

The central contest is therefore not China Telecom against one cloud provider. It is coordinated infrastructure against the fragmented model that still defines much of the computing market.

China Telecom Computing Power Network Moves From Capacity to Coordination

Liu's proposal shifts attention from owning computing capacity to making different resources function as one accessible system.

At the conference, Liu called for deeper coordination between computing infrastructure and electricity systems. He also advocated faster interconnection among computing resources and higher efficiency from domestic AI hardware.

The three proposals address separate constraints, but they depend on one another. More accelerators provide little value when workloads cannot reach them, electricity cannot support them, or software cannot use them efficiently.

Computing and electricity coordination means planning data centers alongside power generation, transmission, storage, and demand management. It also includes operating facilities according to energy availability instead of treating electricity as a fixed input.

Liu highlighted direct renewable power supply and integrated source-grid-load-storage systems. The latter coordinate generation, electrical networks, consumption, and storage within a shared operating model.

This model matters because AI workloads create large, concentrated power requirements. Data center operators must secure sufficient electricity while limiting costs, grid pressure, and carbon emissions.

The second proposal addresses interconnection. Liu called for joint work on heterogeneous resource access, system-wide monitoring, cross-region scheduling, and computing-network integration.

He also proposed unified identifiers, measurement methods, and transaction systems. Those mechanisms would help customers discover, compare, purchase, and schedule capacity from different operators.

China's Ministry of Industry and Information Technology outlined a related structure in its February 2026 node construction plan. That system includes one national service node, multiple regional nodes, and industry-specific nodes.

The ministry said participating nodes should use unified identifiers, standards, and operating rules. Regional platforms would support resource registration, selection, monitoring, scheduling, and security management.

Liu's third proposal focuses on domestic hardware. He called for closer optimization between locally produced computing systems and AI models, covering the full process of token production and delivery.

A token is a unit that an AI model processes when interpreting input or producing output. Token efficiency measures how effectively infrastructure converts chips, energy, memory, and networking into usable AI responses.

This focus shows that headline chip counts are insufficient. Domestic accelerators need compatible software, optimized models, high-speed memory, reliable networking, and production-grade management tools.

China Telecom did not announce a new national platform or binding industry agreement during Liu's speech. The event instead clarified the operating model the carrier wants partners to adopt.

That distinction is important. The proposal establishes direction, while measurable implementation still depends on technical standards, commercial participation, and deployed services.

Why Electricity Has Become Part of AI Scheduling

The next infrastructure bottleneck is not simply access to chips, because usable AI capacity also depends on when and where electricity is available.

China has encouraged data center development in western regions with abundant land and renewable energy. Many AI customers, however, remain concentrated in eastern commercial centers.

That separation creates a matching problem. Operators must decide which workloads can move west, which require nearby capacity, and which tolerate delays created by distance.

Model training can sometimes run far from users because it processes large batches over extended periods. Interactive inference usually requires lower latency because users expect immediate responses.

A coordinated network therefore cannot send every job to the cheapest power source. It must match each workload with suitable latency, hardware, data access, security, and energy conditions.

China Telecom says its infrastructure already includes more than 590,000 data center racks. Its Xirang platform could schedule 87 EFLOPS when Liu spoke at Mobile World Congress in March 2026.

An EFLOPS represents one quintillion floating-point operations per second. However, capacities reported under that label can differ according to hardware, precision, workload, and accounting method.

The company also said it had built a 100G and 400G optical network with 12-millisecond round-trip latency between major hubs. Those figures describe the network foundation behind its scheduling ambitions.

China Telecom has offered one physical example of computing and electricity coordination in Shanghai. Its underwater data center receives direct electricity from an offshore wind farm.

According to the company's infrastructure account, the project uses more than 95 percent green electricity and reduces electricity costs by 50 percent. Those remain company-reported results.

The project illustrates the concept, but it does not establish nationwide economics. An underwater installation has different construction, maintenance, cooling, and connectivity conditions from an inland AI campus.

Renewable generation also varies over time. A system that follows low-cost electricity needs flexible workloads, storage capacity, reliable forecasts, and contractual arrangements with power providers.

Some jobs cannot move when electricity becomes cheaper elsewhere. They may depend on local datasets, regulated environments, specialized accelerators, or latency commitments.

This creates the article's main tradeoff. Central coordination can improve utilization, but only when operators accurately represent the requirements and limits of each resource.

China's national policy is moving in this direction. A May 2026 government statement called for joint planning between computing supply, demand, and electricity systems.

The statement also emphasized monitoring, precision matching, direct green power supply, and stronger security across infrastructure, models, data, and networks.

That policy support gives China Telecom a favorable environment. It does not remove the operational work required to translate energy availability into dependable cloud services.

Fragmented Computing Pools Are the Real Opponent

China Telecom's toughest opponent is infrastructure fragmentation, not a single carrier or cloud company.

China Mobile, China Unicom, Alibaba Cloud, regional governments, supercomputing centers, and specialist operators all control computing resources. Their systems do not automatically expose capacity through identical interfaces.

Hardware differences add another layer. A workload designed for one accelerator family may require substantial adaptation before it performs efficiently on another.

Network conditions also vary. Bandwidth, congestion, packet loss, and physical distance can determine whether remote capacity is useful for a specific AI application.

Commercial terms present a separate barrier. Providers measure availability, utilization, performance, and service quality differently, complicating comparisons across computing pools.

Unified resource identifiers could make capacity discoverable across platforms. Common monitoring interfaces could then report whether those resources are available and performing as advertised.

Scheduling systems would still need to interpret those reports. They must consider processor type, memory, interconnect performance, data location, energy conditions, and customer service commitments.

The proposed national system is designed around that challenge. Its regional and industry nodes are expected to connect different architectures through common interfaces and rules.

A draft technical document registered in January 2026 covers resource connection requirements. China Telecom, China Mobile, China Unicom, Huawei, ZTE, and several research institutions appear among its drafting organizations.

Related projects cover monitoring interfaces, resource management, scheduling, security, measurement, transaction services, and intelligent computing pools. Their scope reveals how many layers require agreement.

The standards process also exposes the distance between policy language and a functioning market. The connection document remains a registered project under consultation, rather than a finished universal standard.

Technical compatibility alone will not guarantee participation. Providers must decide which resources to expose, how much operational data to share, and whether external schedulers can control their capacity.

Operators may hesitate to reveal utilization or performance information that affects pricing. They may also prefer customers to remain within their own cloud environments.

Security requirements create legitimate limits. A unified system needs strong identity controls, tenant isolation, data governance, logging, and responsibility rules across organizational boundaries.

Failure attribution becomes difficult when a workload crosses several networks and facilities. Customers need to know which provider is responsible when performance drops or data becomes unavailable.

This is why the national computing power network resembles a market design project as much as a telecommunications project. It requires trusted measurements and enforceable service expectations.

For enterprise buyers, the promised benefit is broader access without separately integrating every provider. A customer could select capacity by performance, location, energy source, or model compatibility.

The risk is that a nominally unified marketplace hides incomparable products. An advertised unit of AI capacity can deliver different results across models, precision formats, and software stacks.

Transparent workload benchmarks would help buyers evaluate those differences. So would public definitions for availability, energy use, latency, and delivered token throughput.

Teams assessing multiple providers will also need disciplined technical records. A searchable engineering knowledge base can preserve benchmark assumptions, deployment decisions, and incident evidence.

The wider lesson is simple. A national scheduler becomes valuable only when connected resources are sufficiently visible, compatible, and accountable.

Domestic AI Hardware Must Deliver Useful Tokens

Liu's emphasis on token efficiency acknowledges that installed accelerator capacity is not the same as productive AI output.

China Telecom reported 46 EFLOPS of self-owned intelligent computing capacity for 2025. Its self-owned and externally accessed capacity reached 91 EFLOPS, according to company disclosures.

The carrier's 2025 annual report also described more than 3.2 gigawatts of reserved rack power capacity. It cited provincial inference pools and a commercial super-node cluster.

Those numbers indicate substantial infrastructure. They do not show how efficiently every model runs across each connected accelerator type.

AI performance depends on more than raw arithmetic capability. Memory bandwidth, inter-chip communication, software kernels, model architecture, batch size, and numerical precision all affect delivered output.

Training and inference also impose different demands. Training emphasizes sustained computation and fast communication among large accelerator clusters.

Inference serves completed models and often prioritizes response time, availability, and operating cost. Performance can change significantly with prompt length and concurrent requests.

Liu's call for deeper model and hardware adaptation targets this problem. The carrier wants partners to optimize the complete path from computing resources to delivered tokens.

That effort could include model compression, lower-precision processing, workload routing, caching, compiler improvements, and accelerator-specific software. China Telecom did not provide a detailed implementation schedule.

The company's Xirang platform uses what it calls a Triless architecture. China Telecom says this design separates resources, software frameworks, and tools to support heterogeneous capacity.

That separation is useful in theory because customers should not need to redesign applications for every resource pool. The harder question is how much performance survives that abstraction.

A broad compatibility layer can simplify access while hiding hardware-specific features. Direct optimization can deliver better performance but make workloads less portable.

This is another practical tradeoff within the coordinated model. The network must offer common access without reducing every resource to the least capable shared interface.

China Telecom's recent financial disclosures show why management is emphasizing this area. During the first half of 2026, computing-network investment rose 97 percent year over year.

Its total capital expenditure reached 32.4 billion yuan during the period. The carrier said computing infrastructure took a significantly larger share of that spending.

At the same time, eSurfing Cloud revenue reached 61.8 billion yuan, up 7.8 percent. Intelligent business revenue increased 7.1 percent to 31.1 billion yuan.

These figures, reported in the company's interim results, show investment growing faster than the disclosed cloud revenue rate.

That comparison does not prove weak returns. Capital projects serve multiple years, while revenue reflects services already in operation.

It does establish a clear execution test. China Telecom must translate higher infrastructure spending into growing workloads, stronger utilization, and commercially valuable services.

Competition will remain intense. China Mobile reported 52.9 billion yuan in computing-services revenue for the first half of 2026, with data center revenue reaching 19 billion yuan.

Alibaba Cloud approaches the market from another direction. It controls a large public cloud platform, develops AI models, and can optimize software alongside infrastructure.

China Telecom brings nationwide connectivity, regulated-industry relationships, data centers, and public infrastructure responsibilities. Those strengths matter when workloads cross regions or require dedicated networks.

Public cloud providers may move faster on developer tools and model services. Carriers may hold advantages in network control, geographic coverage, and infrastructure coordination.

Liu's proposals attempt to make those carrier assets central to the AI market. Success depends on whether customers value coordinated access more than tightly integrated cloud environments.

The Standards Gap Could Slow the Unified Market

The plan remains credible as an infrastructure direction, but its key performance claims need comparable and independently testable measurements.

China Telecom's figures come mainly from corporate speeches and filings. They describe capacity, network reach, latency, energy use, and operating improvements using company-selected methods.

The disclosures provide useful evidence, yet they do not offer a complete basis for comparison across providers. EFLOPS totals can reflect different precision levels and hardware mixes.

The 87-EFLOPS schedulable figure also includes capacity accessible through a platform. It should not be treated as identical to wholly owned, continuously available hardware.

Energy claims need similar care. A facility's renewable electricity share does not automatically describe hourly carbon intensity or the emissions of each completed workload.

A direct connection to renewable generation can improve traceability. However, periods of low output may require grid electricity, stored energy, workload shifting, or reduced operations.

National monitoring standards could make these figures more meaningful. Buyers need consistent definitions for capacity, delivered performance, utilization, availability, and energy sourcing.

The system also needs reliable ways to measure useful AI output. Tokens per second provide one signal, but that figure changes with models, prompts, context lengths, and quality settings.

Cost per completed task may offer a more practical measure. Even that metric requires agreement about output quality, latency, and failure rates.

Security represents another unresolved area. Cross-provider scheduling increases the number of control systems, interfaces, and administrative boundaries involved in every workload.

A compromised scheduler could expose information about sensitive jobs or redirect workloads. Weak identity controls could allow unauthorized access to connected computing pools.

Regional rules may also restrict where certain datasets can move. Scheduling decisions must therefore account for data residency and sector-specific compliance requirements.

China Telecom has promoted security across its cloud and AI systems. Yet national interconnection requires shared protections among participants with different operational practices.

Governance must determine who certifies providers, audits measurements, investigates incidents, and compensates customers. Technical interoperability cannot answer those questions alone.

The standardization effort remains an encouraging signal because it includes carriers, equipment vendors, cloud organizations, research institutes, and public bodies.

Still, participation in drafting does not guarantee uniform deployment. Providers can implement the same specification differently or expose only limited capacity.

Commercial incentives also matter. A provider with strong demand may see little reason to make its best resources available through an external marketplace.

Less utilized operators may participate more readily, but their capacity might be older, remote, or poorly matched to current AI workloads.

The unified market will therefore need more than supply. It requires enough high-quality demand and enough attractive capacity to produce repeat transactions.

Public reporting should distinguish registered capacity from capacity that customers actually reserve and use. It should also disclose successful cross-provider workload completion.

Without those measures, the network risks becoming a catalog of resources rather than an operating market. That would weaken Liu's central argument for coordination.

Three Signals Will Show Whether the Plan Is Working

The next stage should be judged through deployed standards, cross-provider workloads, and measurable gains in energy-aware scheduling.

The first signal is the transition from draft specifications to operational regional and industry nodes. China Telecom's position strengthens if multiple nodes adopt compatible identifiers and interfaces.

Deployment should include visible resource registration, monitoring, selection, and scheduling functions. A list of participating organizations would not provide equivalent evidence.

The strongest proof would be a workload moving between independently operated pools without extensive manual integration. Published service commitments would make that result more credible.

Failure to advance common interfaces would weaken the coordinated model. Providers would remain dependent on private connections and custom agreements.

The second signal is reported utilization of connected domestic AI systems. Useful disclosures should cover delivered tokens, completed workloads, latency, and availability across different hardware.

China Telecom's case strengthens if optimization improves output without locking customers into one accelerator or model. Comparable benchmarks would carry more weight than aggregate EFLOPS.

Watch for production deployments in government, manufacturing, healthcare, education, and telecommunications. These environments can reveal whether heterogeneous scheduling works under real security and reliability requirements.

Evidence limited to demonstrations would leave the main question unresolved. National-scale infrastructure must operate consistently beyond controlled conference settings.

The third signal is verifiable energy-aware scheduling. The market needs examples showing workloads shifting by electricity availability while still meeting performance commitments.

Those examples should identify the workload type, participating locations, energy conditions, and measured operational outcome. Aggregate renewable percentages provide less insight.

China Telecom's argument becomes stronger if energy coordination reduces curtailment or operating costs without increasing failures. It weakens if power constraints continue to strand computing capacity.

Investors should also compare infrastructure investment with cloud and intelligent-service growth. Rising capital expenditure needs a path toward sustained demand and utilization.

Enterprise buyers should ask whether unified access simplifies procurement and deployment. They should also demand clear responsibility when a workload crosses provider boundaries.

Developers should watch the abstraction layer. A useful computing power network should improve portability while preserving access to accelerator-specific optimization when needed.

Liu's September speech defined a coherent destination: computing, networks, electricity, and domestic hardware coordinated as one service system.

The remaining question is whether China Telecom can convert that architecture into repeatable customer outcomes. That requires standards, incentives, transparency, and operational discipline.

For the next quarter, watch actual node launches, cross-provider workload evidence, and energy-aware scheduling results. Those signals will reveal whether the China Telecom computing power network is becoming infrastructure or remaining a policy-backed ambition.