China's Telecoms Industry Giants Introduce AI Token Subscriptions, Redefining the Telco AI Monetization Playbook

China Telecom, China Mobile, and China Unicom have begun rolling out nationwide Artificial Intelligence (AI) token subscription plans, marking the first large-scale commercial deployment of tokenized AI consumption within the telecoms sector. Rather than billing purely for connectivity, all three operators are now bundling AI inference tokens into monthly consumer and enterprise plans, effectively monetizing access to compute and AI services alongside traditional data packages.

Each operator is pursuing a distinct model for delivering these services, testing where value may accrue in the AI stack. China Telecom has launched tiered token packages nationally for both consumers and enterprises. It bundles these plans with access to its proprietary TeleChat Large Language Model (LLM) and DeepSeek, and it has also launched the China Telecom Token Ecosystem Alliance, covering partners across token production and distribution. China Mobile is positioning itself as a model marketplace and third-party AI service aggregator, pursuing a regional pilot approach. In Shanghai, it has partnered with Tencent, while in Beijing, it offers monthly token packages. China Unicom is taking a lighter approach as a third-party service aggregator, with its Shanghai branch launching consumer token packages. Early token bundles appear relatively modest, with consumer plans typically including on the order of 1 million to 5 million tokens per month, scaling higher for enterprise tiers depending on contract structure. At these levels, usage is primarily suited to conversational AI, content generation, and lightweight enterprise workflows such as customer service automation and document summarization, rather than large-scale inference. This positions current offerings as entry-level consumption bundles, similar to early mobile data plans, designed to introduce users to metered AI usage.

The near-simultaneous rollout of these differing approaches across operators operating within a tightly coordinated national ecosystem further signals China’s structured exploration of commercial and technical models for AI monetization. This stands in contrast to most global markets, where telcos have yet to move beyond early-stage AI service experimentation and have not introduced token-based pricing constructs at scale.

The Chinese operators’ implementation of token-based pricing and the near-simultaneous rollout across all three carriers are both strategically significant and together point to a broader shift in how telco operators may participate in the AI economy. The timing and structure of these launches are difficult to interpret through conventional competitive dynamics, given that China Telecom, China Mobile, and China Unicom typically compete directly across pricing, coverage, and enterprise contracts. Instead, this reflects coordinated direction within a state-led policy framework for AI infrastructure and commercialization.

The coordinated deployment of three distinct approaches, vertically integrated (China Telecom), marketplace platform (China Mobile), and aggregation layer (China Unicom), creates a structured environment in which different models can be evaluated at scale. The conditions enabling this rollout are also market-specific. State-ownership, coordinated infrastructure investment, and a domestic AI ecosystem that supports the introduction of new billing models and integrating compute into telco networks. These factors are not easily replicated in other regions, particularly where hyperscalers dominate AI infrastructure and telco operators have less control over service layers.

These operators are now positioning themselves within the inference value chain by running workloads on their own infrastructure and billing for consumption directly. This marks a shift from connectivity provider to compute provider. Compared to markets where operators bundle third-party AI subscriptions such as ChatGPT, owning the inference layer allows greater control over unit economics, latency optimization through localized compute, and access to usage data that can inform both network planning and service development. Tokens, in this context, function as a metered unit of consumption similar to mobile data, integrating into existing billing systems through tiering, bundling, and overage models.

No operators outside China have yet introduced AI tokens as a bundled subscription service, and the conditions enabling this rollout are not easily replicated. Delivering token-based services also introduces technical requirements, including visibility into inference-level metrics such as throughput and latency per token. Meeting these requirements will likely depend on deeper collaboration with infrastructure providers such as NVIDIA to support token-level telemetry and service-level guarantees.

In markets where hyperscalers dominate AI infrastructure, operators face trade-offs between building on-network inference capacity or relying on external platforms, each with implications for cost, control, and margins. As a result, while the Chinese model provides a clear directional signal, its applicability elsewhere remains constrained by market structure and ecosystem dynamics.

China’s rollout of AI token subscriptions should be viewed as an early indicator of how telco operators may attempt to capture value in the AI economy. Chinese operators are actively being repositioned, under state direction, and operators outside China must treat this as a signal to reassess where value is actually captured in an AI-native telecoms market.

For operators, the starting point is an assessment of the asset base they already control in an AI infrastructure context. Telcos collectively operate geographically distributed data centers and edge infrastructure, last-mile fiber connectivity reaching enterprise premises, and the physical network assets that high-density AI inference deployment requires. Operators need to activate these assets as compute platforms, rather than using them primarily as extensions of connectivity services.

Today, many operators are monetizing AI infrastructure through bare-metal Graphics Processing Unit (GPU) hosting or Compute-as-a-Service (CaaS) offerings, typically centered on hourly access to H100 or B200-class resources. However, this model sits at the lowest value layer of the stack. It is exposed to utilization volatility, structurally capped by hours sold, and increasingly vulnerable to price compression as GPU supply expands globally. More importantly, it misaligns with how enterprises actually consume AI. Enterprises are seeking predictable, contractually guaranteed AI outputs with defined performance characteristics.

This is where tokenization becomes key. For instance, a token-based model built on inference endpoints and microservice-based delivery, such as NVIDIA NIM, reframes compute as an outcome, rather than a resource. Instead of selling time on a GPU, operators sell a metered unit of AI production with embedded performance guarantees. NVIDIA's recently published Cloud Partner reference architecture provides operators with a blueprint for deploying high-yield AI inference factories on infrastructure they already own. The architectural and environmental prerequisites for this model are already largely in place. In this structure, operators become the commercialization layer that converts infrastructure into a managed AI service. Those that transition successfully move into the highest-yield segment of the AI value chain, while those remaining in hourly compute risk disintermediation by hyperscalers or systems integrators moving up the stack.

However, before this is commercially viable for telcos, there must be structural changes. First, infrastructure must be fit for token-level visibility, such as tracking throughput, latency per inference, and cross-session consistency in real time. Second, the commercial model must evolve from capacity sales to Service-Level Agreement (SLA)-backed AI services, requiring integration across GPU platforms, orchestration layers, and Operating Support System (OSS)/Business Support System (BSS) solutions, while also navigating the risk of a deep dependency on companies like NVIDIA. Operators must begin aligning with these requirements. Those that delay will find that enterprise customers and developers have already locked in relationships with providers that can guarantee performance and pricing stability at scale.