The ultimate limit of AI is computing power, and the ultimate limit of computing power is electricity
In March 2026, the "Report on the Work of the Government" proposed for the first time to "build a new form of intelligent economy" and made relevant deployments: deepening and expanding "AI+", promoting the accelerated adoption of new-generation intelligent terminals and intelligent agents, driving the commercial and large-scale application of AI in key industries, and cultivating new business forms and models native to intelligence; implementing new infrastructure projects such as ultra-large-scale intelligent computing clusters and computing-power-electricity coordination, strengthening the national integrated monitoring and scheduling of computing power, and supporting the development of public clouds. This policy orientation indicates that promoting the upgrading of infrastructure such as electricity and computing power to support the large-scale and inclusive application of AI has become a new focus of China's economic growth.
This article briefly elaborates on the logical connection between computing power, electricity, and Tokens, systematically reviews the quantitative relationship between computing power, electricity, and Tokens of mainstream large models, and preliminarily constructs a Token pricing calculation model. The aim is to clarify the value conversion chain from computing power and electricity to Tokens, providing a reference for promoting the development of China's intelligent economy.
01. Computing Power, Electricity, and Tokens—Inputs and Outputs of the Intelligent Economy
From the supply side perspective, the intelligent economy requires a solid foundation of computing power and electricity as its physical base. "The ultimate limit of AI is computing power, and the ultimate limit of computing power is electricity" has become an industry consensus, vividly illustrating that the development of AI is inseparable from efficient and stable computing support and sufficient and reliable energy supply. By the end of March 2026, China's intelligent computing scale had reached 1,882 EFLOPS, with a year-on-year growth expected to exceed 50% for the whole year. According to data from the China Academy of Information and Communications Technology (CAICT), the electricity consumption of China's computing centers in 2025 was about 196 billion kWh, a year-on-year increase of 18.1%; it is estimated that the relevant electricity consumption will reach 500 to 700 billion kWh by 2030. In the future "factory" of the intelligent economy, electricity provides the "fuel" and computing centers serve as the "factory buildings," jointly supporting the operation of various large model "assembly lines." This means that computing power and electricity will become the underlying driving force of the intelligent economy, representing the core input of the value chain.
From the application side perspective, the rapid popularization of intelligent agents is rewriting the logic of Token consumption. Intelligent agents capable of autonomously executing tasks complete operations such as planning, tool invocation, result evaluation, and self-correction through multiple rounds of large model calls. A single model call consumes thousands to tens of thousands of Tokens, and the Token consumption for an entire task cycle can even reach the million level. In March 2026, the National Data Administration announced that "China's average daily Token invocation volume has exceeded 140 trillion, an increase of over a thousand times from 100 billion at the beginning of 2024." Data released by OpenRouter in June showed that the Token invocation volume of Chinese large models has surpassed that of the U.S. for six consecutive weeks, firmly ranking first globally. The surge in Token consumption has prompted its role positioning to shift from a "measurement unit for models" to a "pricing unit for users." Therefore, Tokens will become the standard output of the intelligent economy, representing the native output of the value chain.
It is precisely because of this "input-output" relationship linked by models that computing power, electricity, and Tokens have gradually become core elements of competition in the AI industry.
On the input end, the industry currently focuses primarily on the coordination of computing power and electricity at various levels, dedicated to improving the energy efficiency per unit of computing power. Research institutions such as the China Academy of Information and Communications Technology (CAICT) focus on tackling key technologies and promoting the construction of a batch of computing-power-electricity coordination demonstration projects.
On the output end, the domestic and international industry targets the commercial attributes of Tokens and actively promotes inclusive Token services. In March 2026, Jensen Huang, at the NVIDIA 2026 Developer Conference, defined Tokens as "the new commodity." In the same month, the National Data Administration officially determined the Chinese translation of "Token" as "词元" (Ci Yuan) and pointed out that "Tokens are the value anchor in the intelligent era, which can serve as the 'settlement unit' connecting technology supply and commercial demand, providing quantifiability for the implementation of business models."
However, the value connection between the input and output ends remains to be clarified, and how computing power and electricity determine the actual and commercial value of Tokens still requires further exploration.
02. Quantitative Relationship Between Computing Power, Electricity, and Tokens—Core Metrics for Evaluating Value Conversion Efficiency
Currently, there are various ways to express the quantitative relationship between computing power, electricity, and Tokens, such as Tokens/kWh and FLOPS/Token. This article uses the energy consumption of a model processing one million Tokens to represent the quantitative relationship between computing power, electricity, and Tokens. On the one hand, this is because most models use "one million Tokens" as the performance evaluation benchmark, and using "one million Tokens" as a unified settlement unit conforms to measurement habits. On the other hand, considering that the energy efficiency relationship between computing power and electricity is relatively complex and difficult to describe accurately, treating computing power and electricity as a whole and characterizing them with energy consumption helps simplify the analysis.
Multiple Factors Affect the Energy Consumption per Unit Token
The energy consumption per unit Token is mainly related to three aspects: the model, the task, and the hardware.
First, the impact of model scale and architecture. The number of floating-point operations required for a trillion-parameter model to process each Token may be more than 10 times that of a hundred-billion-parameter model. Besides the parameter scale, architecture design is equally important for the energy consumption of the model. For example, Mixture of Experts (MoE) models such as DeepSeek-V3 activate only a portion of the parameters for each inference. Compared with "dense models" that activate all parameters each time, they can reduce energy consumption by 50% to 70% while maintaining performance.
Second, the impact of task types. There are significant differences in the energy consumption of models processing different tasks. For example, the energy consumption of GPT-4o processing complex reasoning tasks such as code debugging and mathematical proofs is 4 to 6 times that of processing simple dialogue tasks such as daily Q&A and information retrieval.
Third, the impact of hardware configuration and precision. Hardware configuration determines how much computation can be completed per kWh of electricity. Taking the comparison between the enterprise-grade NVIDIA H100 and the consumer-grade RTX 4090 as an example, in multi-model inference scenarios, the average energy consumption of the RTX 4090 per unit Token is 3.1 times that of the H100. Adopting low-precision computing to reduce the number of bits per data can significantly reduce the circuit loss of chip memory read/write and data transmission, thereby significantly lowering hardware energy consumption.
Based on open-source data, this article mainly selects six mainstream large models from China and the U.S. as research objects to analyze the energy consumption per unit Token of different models. Among them, the selected U.S. large models include OpenAI's GPT-5.5, Anthropic's Claude Opus 4.7, Google's Gemini 3, and Meta's Llama 3.1. The Chinese large models include DeepSeek-V3.1 and Alibaba's Qwen 3. The data on the energy consumption of one million Tokens for the six large models when processing text dialogue tasks are shown in Figure 1.
Figure 1 Energy Consumption of One Million Tokens for Mainstream Models Processing Text Dialogue Tasks
By type, the energy consumption of one million Tokens for large models is positively correlated with their parameter scale. The energy consumption of one million Tokens for ultra-large models (parameter count greater than 1000B) is higher than 1 kWh, while that for large models (parameter count between 100B and 1000B) is lower than 0.5 kWh.
By country, under the same scale level, the energy consumption of one million Tokens for Chinese models is lower than that of U.S. models. For example, DeepSeek-V3.1 has more parameters than Llama 3.1, but its energy consumption for one million Tokens is lower. This may be attributed to the widespread adoption of the Mixture of Experts (MoE) architecture in Chinese models. Meanwhile, the U.S. still maintains a performance advantage in ultra-large models. Although the energy consumption of these models is relatively high, they can effectively handle high-value complex tasks.
By model, DeepSeek has the best overall energy consumption for one million Tokens, while Google's Gemini has a clear advantage in the energy consumption comparison among U.S. models. The reason is that DeepSeek and Google have adopted two different energy reduction routes: the former is committed to improvements in model architecture, achieving energy efficiency improvements at the software level through MoE architecture, quantization techniques, etc.; the latter continues to make breakthroughs in hardware performance, relying on its self-developed TPU v6 to build infrastructure, showing excellent energy efficiency performance under large-scale concurrency.
Multi-dimensional Matching and Full-stack Optimization Are Key Means to Reduce Energy Consumption per Unit Token
Reducing the energy consumption per unit Token cannot rely on a single technological breakthrough but requires systematic optimization of "software-hardware collaboration and task awareness" to improve the energy utilization efficiency in all aspects of model operation. Specifically, it involves matching at three levels: model-hardware matching, model-task matching, and task-hardware matching.
In terms of model-hardware matching, model architectures vary widely, and AI accelerators each have their own strengths. If a model is deployed on an unmatched hardware system, it will inevitably lead to problems such as idle computing power and VRAM bottlenecks, resulting in soaring energy consumption. Therefore, the density of the model should be considered first. Traditional dense models activate all parameters for each inference and are quite sensitive to computing density and VRAM bandwidth, making them suitable for general-purpose high-computing-power GPUs equipped with large-bandwidth memory. In contrast, sparse models such as MoE models activate only a portion of the parameters each time and rely more on inter-chip communication and random memory access capabilities. For example, Google's TPU v5p is designed to optimize MoE models and integrates an efficient All-to-All communication unit, which can significantly reduce inter-chip communication latency and energy consumption per unit Token. Secondly, attention should be paid to the alignment of computing precision. If the model has been quantized and compressed, there is no need to use high-precision scientific computing cards; conversely, if a consumer-grade GPU is used to run a high-precision model, energy consumption may surge due to insufficient VRAM. In addition, when the model parameters are too large, memory capacity and layer partitioning also need to be key considerations. Different model partitioning strategies have significant differences in their requirements for inter-chip communication bandwidth.
In terms of model-task matching, users' task requests are complex and variable. If the same model is used to process all requests, originally simple tasks may also consume a large amount of computing power, leading to a significant increase in the actual energy consumption for producing the same amount of Tokens. In this regard, hierarchical scheduling is the most direct and effective strategy. A service cluster containing models of different scales can be constructed to automatically select the best-matched model according to task complexity. For example, in Microsoft's Azure OpenAI service, when a user initiates a "simple translation" request, the system transparently calls an internal small-scale dedicated model, which can reduce inference costs by more than half. Dynamic deep reasoning requires a more precise matching method, where the decision to exit the model early can be made based on the confidence of intermediate results to avoid calling deep networks for simple tasks. Under this "early exit" mechanism, the average reasoning depth can be reduced by 40% to 50%, and energy consumption can also decrease proportionally. In addition, prompt engineering optimization is also an important method to improve the matching degree between models and tasks. The model can automatically rewrite user prompts to avoid triggering unnecessary chains of thought and prevent the introduction of meaningless reasoning processes.
In terms of task-hardware matching, real computing power clusters usually mix various types of hardware. If tasks and hardware cannot be reasonably matched, it will greatly affect the overall utilization rate and energy efficiency level of the system. Therefore, for latency-sensitive tasks such as online inference, high-performance computing resources should be allocated first to reduce queuing and execution latency, avoiding resource expansion and additional energy consumption caused by SLA (Service Level Agreement) constraints. For throughput-priority tasks such as offline training and batch inference, computing resources that meet performance requirements and are in the optimal energy efficiency range can be called. If classified by memory-intensive and compute-intensive, large-bandwidth memory chips can be matched for large-batch, large-model inference, while consumer-grade chips with strong computing power but slightly smaller bandwidth and memory can be used for small-batch, small-model inference. In addition, hardware-aware compilation and execution optimization provide performance profiles and decision-making basis for the precise matching of tasks and hardware. For example, modern inference frameworks such as TensorRT-LLM can automatically fuse operators, rearrange memory access, and select optimal parallel strategies according to the characteristics of the target hardware when executing compilation tasks.
03. Token Pricing Calculation Model—An Effective Tool for Measuring the Value of Computing Power and Electricity
No matter how complex the logical connection and quantitative relationship between computing power, electricity, and Tokens are, in the current context of emphasizing the commercial attributes of Tokens, the value conversion from computing power and electricity to Tokens will ultimately be reduced to a simple external price. Currently, the lack of unified Token pricing rules and the privatization of pricing authority lead to a lack of transparency in the market value of Tokens, which differs greatly from the actual cost.
Therefore, based on the above analysis of energy consumption per unit Token, this article will also explore the actual cost and pricing calculation model of Tokens.
Investment in Computing Power and Electricity Affects the Actual Cost of Tokens
The actual cost of Tokens mainly includes three major parts: hardware depreciation, electricity consumption, and infrastructure-related expenses, all of which are closely related to the investment in computing power and electricity.
First, hardware depreciation is the largest proportion of the actual cost of Tokens, accounting for approximately 50% to 70% of the total cost. Taking the NVIDIA H100 as an example, the procurement cost of an 8-card H100 server is as high as $300,000, and its technical lifespan is 3 to 5 years, requiring the allocation of hundreds of dollars in depreciation expenses daily. Second, electricity consumption is the second largest expenditure item in the actual cost of Tokens, accounting for approximately 12% to 15% of the total cost. The proportion of this part of the cost fluctuates greatly. In addition to being related to the energy consumption per unit Token, it is also affected by real-time electricity prices. Finally, the actual cost of Tokens is also related to infrastructure-related expenses, including investments in network bandwidth, data center operation and maintenance, and cooling systems.
Therefore, the calculation formula for the actual cost of Tokens is:
Specifically, the cost part determined by hardware depreciation is greatly affected by Token consumption. Because the daily depreciation cost of hardware is relatively fixed given a determined hardware lifecycle, the greater the Token consumption, the lower the corresponding cost. For the part determined by electricity consumption, the cost is related to the specific model being run and can be comprehensively calculated through the corresponding energy consumption per unit Token and real-time electricity prices. For the part determined by infrastructure-related expenses, fixed depreciation expenses and operational expenses related to the actual computing volume should be considered. Therefore, the daily related expenses of infrastructure generally show a trend where the total cost rises slowly with the increase in Token consumption, while the corresponding cost per average Token decreases.
Supply-Demand Relationship and Business Models Influence the Market Pricing of Tokens
Theoretically, the actual cost of Tokens provides a reference for the lower limit of their pricing, but the supply-demand relationship and business models also have a significant impact on the market pricing of Tokens.
On the one hand, the large model Token market shows a clear differentiated situation. The upstream supply side faces tight computing power and high costs, while the downstream demand side sees a surge in usage and obvious differentiation. The supply of low- to medium-performance models is relatively sufficient, while high-performance models and high-quality Tokens are in short supply. Ultra-large models capable of completing tasks such as complex reasoning and ultra-long text analysis, although having obvious Token premiums, still hold significant market initiative. For example, ByteDance's Doubao officially launched commercial paid services in late June, and its advanced features such as long document parsing, data analysis, and AI video generation will be charged through membership subscriptions.
On the other hand, the pricing strategies of model manufacturers are not passive cost reactions but proactive choices driven by interests. Manufacturers relying on technological cost reduction are more likely to choose permanent discounts to build ecological barriers; while manufacturers relying on capital investment may adopt short-term large subsidies, which will inevitably face huge price increase pressure in the future. For example, DeepSeek announced a permanent price reduction for its flagship model V4-Pro in May, with an average reduction of up to 75%, while most mainstream large model manufacturers are raising prices.
Based on open-source data, this article selects six mainstream large models from home and abroad as research objects to analyze the unit Token price (input) of different models. The data on the price of one million Tokens for the selected large models are shown in Figure 2.
Overall, the current price of one million Tokens for domestic large models is significantly lower than that of foreign large models. The main reason is the influence of business models, which have certain advantages in terms of energy consumption for one million Tokens, electricity prices, and Token supply. The comparison of U.S. large models shows that Google's Gemini model has a large price advantage for one million Tokens, and the unit Token price of the ultra-large model Gemini 3 is even lower than that of the large model Llama 3.1. The most likely reason for this phenomenon is that Google has an independent and controllable hardware ecosystem, which can support its business model of "reducing costs without reducing efficiency."
Figure 2 Comparison of the Price of One Million Tokens (Input) for Mainstream Models
04. Promoting the Value Leap of Computing Power and Electricity to Support the Inclusive Development of Tokens
Clarifying the value conversion process from computing power and electricity to Tokens not only lays the foundation for achieving a higher level of value leap but also helps build a complete Token business logic. The large-scale application of Tokens relies on the reduction of the unit Token price and will also be affected by the energy consumption per unit Token to a certain extent. Therefore, the industry needs to make coordinated efforts in multiple value links.
First, in the energy supply link, actively explore models such as direct connection to green electricity and "source-grid-load-storage," participate in electricity market transactions, and reduce the electricity price for computing power.
Second, in the "electricity-computing power" conversion link, promote the application of computing-power-electricity coordination technologies at all levels, improve the matching degree of multi-dimensional elements such as models, hardware, and tasks, and reduce the energy consumption per unit Token as much as possible.
Third, in the computing usage link, improve the utilization rate of hardware equipment and infrastructure in computing power clusters, enable them to operate efficiently in the optimal load range, and reduce the occurrence of idle phenomena.
Fourth, in the market pricing link, objective supply and demand relationships should be followed, sustainable business models should be adopted, and healthy competition in the Token market should be maintained.
This article was published in the 7th issue of "Communications World" in 2026; the original title is "From 'Computing Power and Electricity' to 'Tokens': The Underlying Logic and Inclusive Path of the New Form of Intelligent Economy."
Authors: Wang Yue, Zhu Guojun, Bai Liting, China Academy of Information and Communications Technology
Managing Editor/Layout: Zhu Wenfeng
Proofreader: Mei Yaxin
Supervisor: Liu Qicheng