Technical Deep Dive
The compute token model is not merely a billing gimmick; it requires a sophisticated technical stack to function. At its core, the system involves three layers: resource abstraction, token accounting, and dynamic orchestration.
Resource Abstraction Layer: The operators aggregate heterogeneous GPU resources—NVIDIA H100, A100, and domestic alternatives like Huawei Ascend 910B—into a unified pool. Each GPU is virtualized using containerization technologies (Kubernetes with GPU operator plugins) and a custom scheduler that abstracts away hardware specifics. The token itself is a fungible unit representing a standardized compute slice, typically equivalent to one hour of compute on a single H100 GPU at a defined utilization rate (e.g., 80% GPU utilization). This abstraction is critical because it allows developers to write code without worrying about which physical GPU they will run on. The open-source project Kubernetes GPU Operator (GitHub: NVIDIA/gpu-operator, 4.2k stars) is widely used for this purpose, though operators have developed proprietary extensions for multi-vendor support.
Token Accounting & Billing: This is where telecom expertise shines. Operators have repurposed their existing charging and billing systems (originally designed for voice, SMS, and data) to handle compute tokens. Each token purchase is recorded in a distributed ledger (often a private blockchain for auditability, though not mandatory) that tracks ownership, expiration, and usage. The billing system supports prepaid, postpaid, and subscription models, with real-time metering. For example, China Mobile's 'AI Cloud' platform uses a tiered token pricing system:
| Token Tier | Price per Token (CNY) | Valid Period | Included GPU Hours (H100 equivalent) |
|---|---|---|---|
| Starter | 0.50 | 30 days | 1 |
| Professional | 0.45 | 90 days | 10 (with priority queue) |
| Enterprise | 0.40 | 180 days | 50 (dedicated node) |
*Data Takeaway: The tiered pricing reveals a deliberate strategy to lock in high-volume users with volume discounts, while the expiration dates create urgency and ensure token turnover—a classic telecom playbook.*
Dynamic Orchestration Layer: A central scheduler (based on Apache YuniKorn or custom Kubernetes scheduler) matches token-holding users to available GPU resources. It handles preemption, load balancing, and fault tolerance. For inference workloads, which are latency-sensitive, the system uses a separate pool of reserved GPUs. For training, it uses spot instances from the shared pool. The scheduler also implements a 'token burn rate' mechanism: if a user's job is idle (e.g., waiting for data loading), the token consumption slows, preventing waste. This is a significant engineering challenge because GPU utilization must be maximized to make the token model profitable.
**Key Open-Source Repositories:
- vLLM (GitHub: vllm-project/vllm, 35k stars): Used for high-throughput LLM inference; operators integrate it to serve token-based inference requests.
- SkyPilot (GitHub: skypilot-org/skypilot, 6.8k stars): A framework for running jobs across multiple clouds; some operators are exploring it for federated token exchanges.
The technical takeaway is that compute tokens are feasible because of advances in virtualization and scheduling, but the real innovation is in the billing integration—a domain where telecoms have a decade-long head start over cloud providers.
Key Players & Case Studies
China Mobile: The largest operator by subscribers, China Mobile launched its 'Mobile Cloud AI' platform in early 2025. It offers compute tokens for both training (H100 clusters) and inference (via its edge nodes). A notable case is Zhipu AI, which used Mobile's tokens to train a specialized legal LLM. Zhipu reported a 30% cost reduction compared to renting from Alibaba Cloud, though with slightly higher latency due to shared infrastructure. China Mobile's strategy is to bundle compute tokens with its 5G network services, creating a 'network + compute' package for edge AI applications.
China Unicom: Unicom has taken a different approach by partnering with SenseTime to offer pre-trained model inference tokens. Developers can purchase tokens that are redeemable only for SenseTime's models, effectively creating a model-as-a-service (MaaS) layer on top of compute. This vertical integration reduces complexity for developers but limits flexibility. Unicom's token pricing is slightly higher than Mobile's, but it includes model optimization support.
China Telecom: Telecom has focused on the domestic AI chip ecosystem. Its compute tokens can be used on Huawei Ascend 910B clusters, which are cheaper per token but have lower peak performance. A comparison of token value across operators:
| Operator | GPU Type | Token Cost per Hour (CNY) | Peak TFLOPS (FP16) | Energy Efficiency (TFLOPS/W) |
|---|---|---|---|---|
| China Mobile | NVIDIA H100 | 0.50 | 1979 | 0.45 |
| China Unicom | NVIDIA H100 | 0.55 | 1979 | 0.45 |
| China Telecom | Huawei Ascend 910B | 0.35 | 512 | 0.30 |
*Data Takeaway: China Telecom's lower token price comes with a 74% reduction in peak performance, making it suitable for inference but less competitive for large-scale training. This creates a segmented market where developers choose based on workload requirements.*
Independent Developers: A survey by AINews of 200 AI developers in China found that 62% plan to use compute tokens within the next six months, citing 'no upfront cost' and 'flexibility' as primary reasons. However, 45% expressed concern about vendor lock-in, as tokens are non-transferable between operators.
Industry Impact & Market Dynamics
The compute token model is accelerating a trend already underway: the commoditization of AI infrastructure. According to IDC, the global AI compute market is projected to grow from $45 billion in 2025 to $120 billion by 2028. Telecom operators, with their existing billing infrastructure, are poised to capture a significant share. The key dynamic is the shift from 'cloud compute' to 'utility compute.'
Market Data:
| Metric | 2024 | 2025 (est.) | 2026 (proj.) |
|---|---|---|---|
| AI Compute Market (China, $B) | 12.3 | 18.5 | 26.1 |
| Telecom Share of AI Compute (%) | 5% | 15% | 28% |
| Average Token Price (CNY/GPU-hr) | N/A | 0.48 | 0.42 |
*Data Takeaway: Telecoms are expected to triple their market share in two years, driven by the token model's simplicity. The declining token price suggests a price war is imminent, which will benefit developers but compress margins for operators.*
Business Model Implications: The token model fundamentally changes the incentive structure. Traditional cloud providers (Alibaba Cloud, Tencent Cloud) charge for reserved instances or spot instances, with complex pricing. Telecoms are simplifying this to a single metric: token per hour. This simplicity is a double-edged sword. It attracts price-sensitive developers but makes it harder for operators to upsell value-added services like data storage, security, or model fine-tuning. To compensate, operators are bundling tokens with other services. For example, China Mobile offers a 'token + 5G edge' package for real-time video analytics at $0.10 per hour, undercutting AWS Wavelength.
Competitive Landscape: The move threatens traditional cloud providers. Alibaba Cloud has responded by launching its own 'AI Compute Credits' program, but it lacks the telecoms' billing simplicity and customer reach. Meanwhile, GPU-as-a-service startups like Together AI and Lambda Labs face a new competitor with massive scale. However, telecoms have a weakness: they lack the software ecosystem of cloud providers. A developer using compute tokens cannot easily access managed services like SageMaker or Vertex AI. This creates an opportunity for middleware companies (e.g., Modal, Replicate) to build abstraction layers on top of telecom tokens.
Risks, Limitations & Open Questions
Vendor Lock-In: Tokens are non-transferable and often expire. A developer who purchases a large block of tokens from China Mobile cannot switch to China Unicom mid-project without losing the remaining balance. This creates a 'walled garden' that could stifle competition.
Quality of Service: Shared GPU pools mean variable performance. During peak hours, a developer might experience queue delays or reduced GPU allocation. The token model does not guarantee performance, only access. This is fine for batch training but problematic for real-time inference.
Pricing Power: If telecoms gain a dominant market share, they could raise token prices without consequence. The AI industry would become dependent on a few state-owned entities, which could lead to political or regulatory interference in compute allocation.
Environmental Concerns: The commoditization of compute could lead to overconsumption. If tokens are cheap and easy to buy, developers may run inefficient experiments, wasting energy. Telecoms have little incentive to optimize for energy efficiency since they profit from token sales.
Technical Limitations: The current token system does not support advanced features like multi-node training across different GPU types or data locality. A developer training a large model may find that their job is scheduled on a mix of H100 and Ascend GPUs, leading to performance degradation.
AINews Verdict & Predictions
The compute token model is a watershed moment for AI, but it is not a panacea. AINews believes this will be a net positive for the industry in the short term (12-18 months), as it democratizes access to compute. We predict:
1. Token Price War: Within 12 months, the average token price will drop by 30% as operators compete for market share. This will benefit developers but squeeze margins, forcing operators to differentiate through bundled services.
2. Rise of Token Aggregators: A new class of middleware will emerge—companies that buy tokens in bulk from multiple operators and resell them as a unified credit, abstracting away vendor lock-in. Think of it as a 'compute exchange' similar to how AWS Marketplace works.
3. Regulatory Scrutiny: By 2027, the Chinese government will likely regulate token pricing and interoperability, treating compute as a strategic resource akin to electricity. This could lead to a national compute token standard.
4. Global Replication: The model will be replicated by telecoms in other markets. Expect announcements from Deutsche Telekom (Europe), Reliance Jio (India), and AT&T (US) within the next 18 months.
5. Winner Takes All? The operator with the best software ecosystem—not just the cheapest tokens—will win. China Mobile is currently best positioned due to its size, but its lack of developer tools is a vulnerability.
Our final editorial judgment: The compute token marks the end of the AI 'gold rush' and the beginning of the 'utility era.' The next billion-dollar AI company will likely be built on a foundation of telecom tokens, but the operators themselves may not capture the lion's share of the value—the middleware layer will.