Research on the Economic Transmission Mechanism and Dynamic Optimization of Computing Power Networks Based on a Multi-Sectoral Input–Output Model and a Hybrid Algorithm Solution

Abstract

In the digital economy era, computing power, as a novel factor of production, serves as a vital engine for driving high-quality economic development. Building upon China’s traditional 42-sector input–output table, this paper incorporates computing power networks as a new sector to construct a 43-sector dynamic input–output (IO) model. Based on this framework, a Dynamic Stochastic General Equilibrium (DSGE) analysis framework is constructed to systematically reveal the dynamic transmission mechanism of computing power within industrial linkages and capital accumulation. From an energy perspective, energy consumption is implicitly captured through carbon emissions and energy structure, which together reflect the scale, efficiency, and composition of energy use in computing power networks. The findings show that the optimal computing power allocation follows a temporal evolution pattern from the service sector to the manufacturing sector, with ICT manufacturing’s computing power quota reaching 31% by 2030. An investment inflection point occurs in 2026, aligning with the digital infrastructure cycle of China’s 14th Five-Year Plan. The “Eastern Data, Western Computing” strategy reduces unit carbon emissions from computing power by 41%. Policy simulations demonstrate that R&D tax credits generate a 2.9-fold multiplier effect through industrial linkages, boosting GDP by 2.3%. The integrated IO-DSGE framework developed in this study provides a quantitative tool for the full-cycle management of “construction–application–regulation” in computing power networks. It holds significant theoretical value and practical implications for enhancing resource allocation efficiency and promoting green, climate-friendly development.

1. Introduction

In the digital economy era, computing power has emerged as a critical global production factor, reshaping industrial structures and national competitiveness worldwide [1,2]. However, this rapid growth brings two significant challenges. First, the expansion of computing power networks intensifies energy constraints [3,4]. According to the International Energy Agency (IEA), global electricity consumption by data centers, including computing power networks, surpassed 400 TWh in 2024, representing 1.5% of global electricity demand. This figure is projected to rise to 8% by 2030 [5]. The European Union, while advancing its digital agenda, faces similar challenges, with data center electricity consumption central to its energy security and climate strategy [6]. Second, the spatial mismatch between computing demand and renewable energy supply exacerbates inefficiencies [7], undermining the ability to maximize the economic benefits of computing power while aligning with low-carbon goals.

China offers a unique case for addressing these challenges, as it is both a major computing power market and a leader in energy transition. With a total computing power of 280 EFLOPS in 2024 [8] and over 80% of its electricity consumption sourced from green energy, China’s “Eastern Data, Western Computing” strategy embodies the economic and energy interplay within computing power networks [9,10]. Additionally, China’s commitment to peak carbon emissions by 2030 and achieve carbon neutrality by 2060 further emphasizes the need to understand how energy constraints interact with the economic impacts of computing power networks. The insights from China’s experience not only offer value for its domestic context but also serve as a reference for other nations grappling with similar computing–economics–carbon challenges.

Despite growing research on computing power networks, existing studies have yet to integrate the economic transmission mechanism with energy and environmental constraints. Static models capture industrial linkages but neglect the dynamic impacts of computing power networks [11,12,13,14,15]. Conversely, dynamic frameworks overlook the complex interconnections between industrial networks, including computing power [15,16,17]. Furthermore, few studies link computing power’s economic effects to low-carbon goals.

This paper aims to develop a unified analytical framework to examine the economic transmission mechanism and dynamic optimization of computing power networks in a multi-sector economy. Specifically, it explores how computing power networks, as a novel production sector, reshape intersectoral linkages and influence key macroeconomic outcomes, particularly industrial structure and carbon emissions. To achieve this objective, this paper focuses on three key aspects: the integration of static industrial linkages with dynamic capital accumulation and environmental factors, the role of resource and environmental constraints in shaping economic transmission, and the effectiveness of policy simulations in optimizing the economic benefits of computing power networks under low-carbon targets.

By addressing these questions, this study advances the existing literature by offering dual-faceted contributions—both conceptually and methodologically. First, conceptually, rather than treating digital infrastructure merely as an exogenous technological shock—as is common in traditional macroeconomic frameworks—this paper formally operationalizes computing power as an endogenous, novel production factor. This theoretical shift advances our current understanding by enabling the precise quantification of computing power’s forward and backward industrial reliance (e.g., via the proposed computing power multiplier), explicitly revealing how it endogenously drives long-term industrial upgrading and structural transformation. Second, methodologically, the proposed framework fundamentally differs from existing hybrid modeling approaches through its depth of structural embeddedness. Instead of loosely coupling static IO outputs with aggregate Dynamic Stochastic General Equilibrium (DSGE) variables, this study strictly internalizes the newly constructed 43-sector input–output matrix directly into the dynamic capital accumulation equations and state-space transition mechanisms. Through this deep integration, the model seamlessly captures complex, high-dimensional sectoral dependencies within a dynamic equilibrium setting. Finally, by embedding energy use and carbon constraints into this unified IO-DSGE framework, this paper provides a robust quantitative tool to assess the energy-growth trade-off, thereby offering targeted policy insights for the low-carbon optimization of the computing-power economy.

2. Literature Review

2.1. Static Analysis Using the IO Model

The IO model, a well-established economic analysis tool, has proven instrumental in elucidating industrial structures and inter-sectoral linkages. Introduced by Wassily Leontief (1936), the IO analysis framework laid the groundwork for the field, earning him the Nobel Prize in Economics for his pioneering work [18]. Miller and Blair (2009) provided a systematic exposition of the theoretical foundations and methodological underpinnings of the IO model, establishing it as an essential tool for understanding industrial linkage effects [19]. These studies underscore the utility of IO models in quantifying industrial multipliers, identifying key sectors, and assessing the static effects of economic policies. For example, by analyzing forward and backward linkages, researchers can pinpoint crucial industrial sectors that exert substantial driving or pulling effects on the national economy [20]. Oosterhaven (1988) further explored the applicability of demand-pull and supply-driven models within the IO framework, contributing to a more nuanced understanding of inter-sectoral linkages [21].

In the context of the burgeoning digital economy, scholars have increasingly utilized IO models to assess the economic impact of emerging technology sectors. Toh and Thangavelu (2013) applied the IO methodology to examine the effects of the information industry on economic growth [22]. García-Muñiz and Vicente (2014) integrated IO analysis with network theory and Burt’s structural hole concept, specifically focusing on the Information and Communication Technology (ICT) sector to investigate its role as a conduit for information and knowledge flow in the economic network [23]. In recent studies, scholars have extended this approach to examine the economic and environmental implications of digital transformation, particularly within energy-intensive sectors.

With the rise of computing power networks, researchers have also begun exploring the relationships between the computing power sector and other industries. Settanni and Srai (2016) constructed an IO model to examine the interactions between the UK’s digital industry and other sectors [24]. Lyu et al. (2023) employed the Propensity Score Matching-Difference in Differences (PSM-DID) method to compare the effects of the Digital Silk Road and innovation heterogeneity on digital economic growth across nine treatment countries and twenty control countries [12].

While IO models have been instrumental in analyzing the economic impacts of the digital industry, they have also been extended to evaluate its environmental consequences. Huang et al. (2023) isolated the digital economy from a multi-regional IO table and applied a downscaling structural decomposition analysis method to reveal the technological, structural, and scale effects of the digital economy on carbon emissions [11]. Li et al. (2025) explored the effects of China’s digital industry development on carbon emissions [25].

2.2. Dynamic Analysis Based on the DSGE Framework

In macroeconomics, the DSGE model has emerged as a predominant tool for analyzing economic fluctuations and the transmission mechanisms of policy shocks. Kydland and Prescott (1982) introduced the Real Business Cycle (RBC) theory, which established the foundations of DSGE models, highlighting the central role of technological shocks in economic fluctuations [26]. In recent years, researchers have integrated complex industrial structures and network effects into the DSGE framework. Acemoglu et al. (2016) explored the impact of industrial network structures on technological shock transmission by incorporating network analysis, finding that the topological structure of industrial networks significantly influences the scope and duration of economic fluctuations [27]. This study offers a valuable perspective for understanding how inter-sectoral interdependence in modern economies amplifies or mitigates economic shocks. Wang et al. (2025) developed a DSGE model for an open oil economy to explore the transmission mechanism and impacts of oil price fluctuations on China’s macroeconomy [15].

However, solving high-dimensional dynamic optimization problems within DSGE models remains a challenge in economics. Judd (1998) discussed numerical methods in economics, particularly highlighting the “curse of dimensionality” that complicates traditional dynamic programming methods when dealing with high-dimensional problems [28]. In recent years, a few studies have attempted to combine IO models with DSGE frameworks. For instance, Ghasem Palouj et al. (2024) investigated the effects of monetary and fiscal policies on macroeconomic variables through a multi-sector DSGE model integrated with IO analysis [17]. Hu et al. (2025) applied an IO model, coupled with a four-quadrant resilience framework, to analyze the linkage characteristics and ripple effects across 21 economic sectors in Tibet from 2012 to 2017 [16].

In summary, there are notable gaps in the existing literature: First, there is a lack of an economic analysis framework specifically designed for computing power networks; second, the integration of static industrial linkages with dynamic capital accumulation remains underdeveloped; third, the challenge of solving high-dimensional dynamic optimization problems is not yet effectively addressed. This study aims to fill these gaps by developing a dynamic IO model embedded with a computing power network and introducing an efficient hybrid optimization algorithm, thus providing both theoretical support and decision-making tools for advancing the computing-power economy.

3. Construction of a Computing Power-Embedded IO Model

3.1. Architecture of the Expanded IO Table

IO analysis, a classical method for studying industrial linkages and economic structure, was proposed by Nobel laureate Wassily Leontief in the 1930s. Traditional IO tables divide the national economy into several sectors, revealing technical and economic linkages between industries by recording the flow of products and services between them. To analyses the economic impact of computing power networks, this study expands the traditional 42-sector IO table framework by adding a 43rd sector—computing power network services.

To accurately capture its distinct economic role, this study extracts ‘computing power’ as an independent Sector 43. In this framework, traditional ICT manufacturing (e.g., the production of servers, chips, and network equipment) and general software programming remain embedded within the standard sectoral classifications. In contrast, Sector 43 is strictly bounded as the infrastructural provision of integrated computational resources. Its boundaries specifically encompass data center hosting, cloud computing leasing, and intelligent computing service provision (e.g., AI model training capabilities). Essentially, Sector 43 represents the ‘utility-like service output’ of computing infrastructure, cleanly separated from both upstream physical hardware manufacturing and downstream general digital applications.

The sector classification adheres to the National Economic Industry Classification (GB/T 4754-2017) [29] and the characteristics of the computing-power economy, ensuring scientific categorization and data availability. The full sector classification is presented in Appendix A 

Table A1. (China) Sector Classification Correspondence Table.

Sector Code	Sector Name	Industry Code	Sub-Sectors Included
01	Agriculture	A01–A04	Cereal cultivation, vegetable cultivation, fruit cultivation, etc.
02	Forestry	A05	Forest tree breeding, afforestation, timber harvesting and transport, etc.
03	Livestock farming	A06	Livestock rearing, poultry farming, etc.
04	Fisheries	A07	Aquaculture, fishing, etc.
05	Coal Mining	B06	Coal mining and washing
06	Oil extraction	B07	Crude Oil Extraction, Natural Gas Extraction, etc.
07	Metal Mining and Beneficiation	B08–B09	Iron ore, copper ore and other metal mining and beneficiation
08	Non-metallic Mineral Mining and Processing	B10–B12	Sand, gravel and chemical mineral extraction, etc.
09	Food manufacturing	C13–C15	Processing of agricultural by-products, food manufacturing, beverage production
10	Textile Industry	C16–C18	Textiles, Textile Apparel, Leather Goods
11	Wood processing	C20	Wood processing, wood product manufacturing
12	Paper and Printing	C21–C23	Paper, Printing, and Educational Supplies Manufacturing
13	Petroleum Processing	C25	Manufacture of refined petroleum products
14	Chemical Raw Materials	C26	Manufacture of basic chemical materials
15	Pharmaceutical Manufacturing	C27	Manufacture of Chemical Medicines and Traditional Chinese Medicinal Preparations
16	Chemical Fibers	C28	Manufacture of chemical fibers
17	Rubber and plastics	C29–C30	Rubber and Plastic Products
18	Non-metallic minerals	C30	Cement, glass, ceramic products
19	Ferrous metals	C31	Steel smelting and rolling
20	Non-ferrous metals	C32	Non-ferrous metal smelting and rolling
21	Metal Products	C33	Metal tools, containers and structural products
22	General-purpose equipment	C34	Boilers, engines, metalworking machinery
23	Specialized Equipment	C35	Chemical, metallurgical and building materials specialized equipment
24	Automotive Manufacturing	C36	Manufacture of complete motor vehicles and parts
25	Railway and Shipping	C37	Railway Transport Equipment, Shipbuilding
26	Electrical Machinery	C38	Motor manufacturing, power transmission and distribution equipment
27	Communications Equipment	C39	Communications equipment manufacturing
28	Computer Equipment	C40	Manufacture of complete computers and components
29	Instruments and Meters	C41	Manufacture of general and specialized instruments
30	Other Manufacturing	C42	Crafts, Waste Processing, etc.
31	Electricity and heat supply	D44	Electricity production and supply
32	Gas Production	D45	Gas production and supply
33	Water production	D46	Production and Supply of Water
34	Construction Industry	E47–E50	Building Construction, Civil Engineering Construction
35	Wholesale and Retail	F51–F52	Wholesale and Retail Trade
36	Transportation	G53–G60	Rail, road, water transport, etc.
37	Accommodation and Catering	H61–H62	Accommodation services, food services
38	Information Transmission	I63	Telecommunications and broadcasting transmission services
39	Software Services	I64	Software development and information system integration
40	Financial Services	J66–J69	Monetary and financial services, capital market services
41	Real Estate	K70	Real Estate Development and Management
42	Business Services	L71–L72	Leasing, Business Services
43	Computing Network Services	I65	Computing Infrastructure Services, Cloud Computing Services

. The specific sector classifications are presented in

Table 1. Sector Classification of the Input–Output Table (China) Sector Classification of the Input–Output Table.

Sector Category	Included Sectors	Number	Specific Sector Examples
Primary Sector	Agriculture, forestry, animal husbandry, fisheries	4	Agriculture (01), Forestry (02), Animal Husbandry (03), Fisheries (04)
Secondary Sector	Coal, petroleum, food manufacturing, textiles, chemicals, metals, equipment manufacturing, etc.	26	Coal mining (05), oil extraction (06), metal mining and beneficiation (07), non-metallic mineral mining and processing (08), etc.
Tertiary Sector	Transportation, commercial services, public services, etc.	12	Wholesale and Retail (35), Transportation (36), etc.
Computing Power Sector	Computing Network Services	1	Computing Network Services (43)

Note: Columns 3–4 rank sectors by forward linkage intensity (computing power output stimulated by each sector’s final demand), while columns 5–6 rank sectors by backward linkage intensity (each sector’s demand for computing power inputs). These represent two distinct dimensions of computing power network interdependence.

.

The fundamental structure of an extended IO table comprises an intermediate use matrix, a final use matrix, and a value-added matrix. The intermediate use matrix records the flow of intermediate products between sectors, the final use matrix encompasses final demand components such as consumption, capital formation, and exports, while the value-added matrix includes compensation of employees, net production taxes, depreciation of fixed assets, and operating surplus.

To analyze the economic impact of computing power networks, an extended flow-balance equation must be established. Let the output vector of traditional sectors be $X\in R^{42}$ and the total output of computing power services be $Q_{CNS}$. The extended flow-balance equation can then be expressed as:

$$\left({\displaystyle \genfrac{}{}{0pt}{}{X}{Q_{\mathit{CNS}}}}\right)=\left(\begin{array}{cc}{A}^{‘}& 0_{42\times 1}\\ r^T& 0\end{array}\right)\left({\displaystyle \genfrac{}{}{0pt}{}{X}{Q_{\mathit{CNS}}}}\right)+\left({\displaystyle \genfrac{}{}{0pt}{}{F}{F_{\mathit{CNS}}}}\right)$$

(1)

where $A^{‘}$ is the 42 × 42 direct consumption coefficient matrix for traditional sectors; $r^T$ is the input coefficient vector for the computing power sector, representing the quantities of products from each traditional sector consumed per unit output of the computing power sector; $0_{42\times 1}$ is the 42 × 1 direct consumption coefficient vector of computing power services by traditional sectors, indicating the amount of computing power consumed as intermediate inputs per unit of output in each traditional sector. This captures the role of computing power as a direct intermediate input, which is complemented by its role as a capital-embodied factor captured through the dynamic DSGE framework (The dual-channel treatment of computing power captures its full economic transmission mechanism. The c vector (intermediate input channel) is empirically estimated from sector-level computing expenditure data. The DSGE capital accumulation channel (capital-embodied computing) captures the long-run infrastructure effects. This dual approach ensures consistency with the System.) $\mathrm{F}$ is the final use vector for traditional sectors; $F_{\mathit{CNS}}$ is the final demand for computing power services.

It should be noted that the c vector captures only the first of two economic channels through which computing power services influence production. Channel (i)—direct intermediate input channel: computing power services (cloud computing, AI model training, data processing) are directly consumed within sectoral production processes. This channel is captured by the c vector in the extended IO matrix. Channel (ii)—capital-embodied channel: computing power embedded in equipment and digital infrastructure (smart devices, servers, network equipment) accumulates as productive capital and is captured through the dynamic capital accumulation equations in the DSGE framework (Section 4.3.1). This dual-channel treatment ensures that the model captures both the short-run operational demand for computing services and the long-run structural effects of computing infrastructure investment.

3.2. Methodology for Measuring Key Coefficients

3.2.1. Calibration of the Direct Consumption Coefficient Matrix

Direct consumption coefficients constitute the core parameters of input–output analysis, representing the quantity of products from one sector directly consumed by another sector to produce one unit of output. The elements of the traditional direct consumption coefficient matrix A, denoted as $a_{\mathit{ij}}$, indicate the quantity of products from sector $i$ directly consumed by sector $\mathrm{j}$ to produce one unit of output:

$$a_{\mathit{ij}}={\displaystyle \frac{z_{\mathit{ij}}}{X_j}}$$

(2)

where $z_{\mathit{ij}}$ denotes the quantity of product $i$ consumed by sector $\mathrm{j}$, and $X_j$ represents the total output of sector $\mathrm{j}$.

Owing to the particularities of computing power networks, an enhanced approach is required to calibrate direct consumption coefficients. This paper employs the RAS Matrix Updating Method (see Appendix B.1 for the detailed RAS iteration procedure) (RAS) method for coefficient updating, modified specifically for computing power network characteristics. The RAS method is an iterative proportional adjustment technique that modifies the initial coefficient matrix in both row and column directions to satisfy given row sum (intermediate total) and column sum (intermediate input total) constraints.

The objective function for the modified RAS method is:

$$\underset{A^{‘}}{\min }\sum _i\sum _j{\left({\displaystyle \frac{a_{\mathit{ij}}^{‘}-{\stackrel{-}{a}}_{\mathit{ij}}}{{\mathsf{\sigma }}_{\mathit{ij}}}}\right)}^2$$

(3)

constraints include: 1. Computational power consumption consistency constraint: ${\sum }_j{c}_j{X}_j=Q_{\mathit{CNS}}$; 2. Value-added balance constraint: ${\mathrm{V}}^{\mathrm{T}}{\left(\mathrm{I}-{\mathrm{A}}^{‘}\right)}^{-1}\mathrm{F}=\mathrm{GDP}$; 3. Non-negativity constraint: $a_{\mathit{ij}}^{‘}\ge 0$.

• where ${\stackrel{-}{a}}_{\mathit{ij}}$ denotes the initial direct consumption coefficient; ${\mathsf{\sigma }}_{\mathit{ij}}$ represents the coefficient standard deviation, reflecting coefficient stability; $c_j$ is the direct computational power consumption coefficient for sector $j$; V is the value-added coefficient vector.

The optimization problem is solved using an iterative algorithm, converging after 15 iterations with residuals below 10⁻⁶, meeting accuracy requirements.

3.2.2. Calculation of the Total Demand Coefficient

The full demand coefficient is a pivotal concept in IO analysis, represented by the Leontief inverse matrix (see Appendix B.2 for the complete calculation). The formula for calculating the Leontief inverse matrix (see Appendix B.2 for the complete calculation) B is:

$$B={\left(I-A^{‘}\right)}^{-1}=\left(\begin{array}{cc}{B}_{11}& B_{12}\\ B_{21}& B_{22}\end{array}\right)$$

(4)

where $B_{11}$ is the 42 × 42 matrix of perfect demand coefficients for traditional sectors; $B_{12}$ is a 42 × 1 vector representing the perfect dependency of traditional sectors on computing power services (${\mathrm{b}}_{43,j}$); $B_{21}$ is a 1 × 42 vector representing the perfect pull effect of the computing power sector on traditional sectors (${\mathrm{b}}_{j,43}$); $B_{22}$is a scalar representing the perfect demand within the computing power sector itself. The full 43 × 43 extended direct consumption coefficient matrix is provided as Appendix C 

Table A2. A representative fragment of the extended 43×43 Leontief inverse matrix.

	(1) Agriculture	(2) Forestry	(3) Livestock Farming	(4) Fisheries	(5) Coal Mining	(6) Financial Services	(7) Real Estate	(8) Business Services	(9) Computing Network Services
Agriculture	1.3435	0.2813	0.3445	0.3019	0.2515	0.2558	0.2554	0.2845	0.2627
Forestry	0.2776	1.4212	0.3193	0.2831	0.2677	0.2486	0.2647	0.2753	0.2441
Livestock farming	0.2708	0.2605	1.4655	0.2783	0.2618	0.3572	0.3334	0.3541	0.3599
Fisheries	0.2542	0.2716	0.3418	1.3439	0.2421	0.2754	0.3021	0.2875	0.3020
Coal Mining	0.2719	0.2442	0.3480	0.2954	1.3676	0.2538	0.2445	0.2444	0.2583
Oil extraction	0.2666	0.2569	0.3340	0.2772	0.2390	0.4482	0.4625	0.4515	0.4670
Metal Mining and Beneficiation	0.2512	0.2507	0.3251	0.2829	0.2377	0.4764	0.4704	0.4561	0.4629
Non-metallic Mineral Mining and Processing	0.2662	0.2440	0.3337	0.2859	0.2392	0.4822	0.4590	0.4761	0.4603
Food manufacturing	0.2648	0.2558	0.3307	0.2802	0.2403	0.4751	0.4823	0.4531	0.4672
Textile Industry	0.2609	0.2440	0.3170	0.2893	0.2528	0.5661	0.5672	0.5596	0.6004
Wood processing	0.2547	0.2564	0.3284	0.2784	0.2478	0.5330	0.5444	0.5605	0.5544
Paper and Printing	0.2556	0.2408	0.3182	0.2898	0.2409	0.4861	0.4752	0.4894	0.4846
Petroleum Processing	0.2688	0.2470	0.3261	0.2858	0.2452	0.4810	0.5000	0.4940	0.5012
Chemical Raw Materials	0.2523	0.2423	0.3298	0.2861	0.2474	0.5824	0.5641	0.5821	0.6072
Pharmaceutical Manufacturing	0.2649	0.2456	0.3357	0.2869	0.2458	0.4595	0.4468	0.4426	0.4582
Chemical Fibers	0.2513	0.2509	0.3339	0.2733	0.2535	0.5053	0.4857	0.4810	0.4986
Rubber and plastics	0.2561	0.2395	0.3348	0.2827	0.2413	0.5443	0.5384	0.5442	0.5728
Non-metallic minerals	0.2654	0.2399	0.3204	0.2869	0.2410	0.4615	0.4746	0.4588	0.4770
Ferrous metals	0.2510	0.2551	0.3254	0.2746	0.2418	0.5102	0.4947	0.4833	0.5127
Non-ferrous metals	0.2562	0.2475	0.3308	0.2703	0.2437	0.5658	0.5329	0.5461	0.5723
Metal Products	0.2633	0.2558	0.3335	0.2789	0.2410	0.5505	0.5479	0.5707	0.5654
General-purpose equipment	0.2490	0.2495	0.3304	0.2828	0.2499	0.4849	0.4874	0.4957	0.4958
Specialized Equipment	0.2529	0.2450	0.3288	0.2697	0.2411	0.4962	0.4808	0.4803	0.5134
Automotive Manufacturing	0.2661	0.2539	0.3296	0.2728	0.2482	0.4185	0.4332	0.4319	0.4170
Railway and Shipping	0.2492	0.2534	0.3156	0.2829	0.2403	0.5557	0.5487	0.5437	0.5704
Electrical Machinery	0.2523	0.2427	0.3257	0.2741	0.2405	0.4762	0.4600	0.4773	0.4791
Communications Equipment	0.2519	0.2550	0.3182	0.2722	0.2414	0.5251	0.5380	0.5326	0.5203
Computer Equipment	0.2622	0.2529	0.3346	0.2874	0.2481	0.5234	0.5415	0.5410	0.5473
Instruments and Meters	0.2619	0.2598	0.3387	0.2820	0.2621	0.3519	0.3597	0.3653	0.3710
Other Manufacturing	0.2879	0.2528	0.3284	0.3024	0.2711	0.3258	0.3448	0.3310	0.3438
Electricity and heat supply	0.2794	0.2433	0.3544	0.2839	0.2724	0.3890	0.3816	0.3962	0.3795
Gas Production	0.2666	0.2603	0.3234	0.3017	0.2705	0.2880	0.3201	0.3307	0.2836
Water production	0.2749	0.2525	0.3209	0.2854	0.2446	0.3244	0.3410	0.3377	0.3090
Construction Industry	0.2767	0.2528	0.3449	0.2902	0.2525	0.2784	0.3068	0.2917	0.2752
Wholesale and Retail	0.2567	0.2465	0.3230	0.2762	0.2711	0.2791	0.2778	0.2884	0.2875
Transportation	0.2764	0.2614	0.3318	0.2850	0.2427	0.3368	0.3255	0.3358	0.3168
Accommodation and Catering	0.2737	0.2448	0.3359	0.2971	0.2450	0.3531	0.3439	0.3505	0.3737
Information Transmission	0.2802	0.2498	0.3367	0.2773	0.2598	0.2701	0.2932	0.2829	0.2765
Software Services	0.2783	0.2463	0.3411	0.2957	0.2486	0.3358	0.3264	0.3253	0.3528
Financial Services	0.2558	0.2486	0.3572	0.2754	0.2538	1.3923	0.3118	0.3257	0.3026
Real Estate	0.2554	0.2647	0.3334	0.3021	0.2445	0.3359	1.4381	0.2924	0.2924
Business Services	0.2845	0.2753	0.3541	0.2875	0.2444	0.3895	0.4096	1.5331	0.4143
Computing Network Services	0.2627	0.2441	0.3599	0.3020	0.2583	1.3513	1.2837	1.2156	2.1887

Note: Columns 1–5 display the transposed values of Rows 1–5 of the original matrix. Columns 6–9 display the original Columns 40–43 of the matrix.

.

The total demand coefficient holds significant economic implications. $b_{43,j}$ denotes the total demand for computing power services (encompassing both direct and indirect demand) generated when sector $\mathrm{j}$ increases its final demand by one unit. Similarly, ${\mathrm{b}}_{j,43}$ represents the total pull effect exerted on sector $j$ when the computing power sector increases its final demand by one unit.

The computing power multiplier serves as a key indicator for measuring the economic pull effect of computing power investment, defined as:

$$M_{\mathit{CNS}}={\displaystyle \frac{{\sum }_j{b}_{j,43}}{{\sum }_j{c}_j}}$$

(5)

where the numerator represents the total pull effect of the computing power sector on all traditional sectors, and the denominator denotes the direct consumption of computing power services by all traditional sectors. A computing power multiplier greater than 1 indicates that computing power investment possesses an amplifying effect, capable of driving greater economic growth.

3.3. Industrial Linkage Mechanism

This subsection focuses on the theoretical mechanism, while the corresponding numerical results are presented in Section 5.

Forward linkage effect: Each additional unit of output in the computing power sector stimulates 0.201 units of output in the semiconductor sector. This stimulation effect can be decomposed using the following formula:

$${\displaystyle \frac{\partial {\mathrm{X}}_{15}}{\partial {\mathrm{Q}}_{\mathrm{CNS}}}}={\mathrm{b}}_{15,43}+\sum _{\mathrm{k}\ne 43}{\displaystyle \frac{\partial {\mathrm{a}}_{\mathrm{k},43}}{\partial {\mathrm{Q}}_{\mathrm{CNS}}}}$$

(6)

where the first term represents the direct pull effect and the second term represents the indirect pull effect (via the technological coefficient affecting other sectors).

Backward linkage effect: An additional unit of output in the internet services sector consumes 0.324 units of computing power entirely. This dependency can be decomposed as follows:

$${\displaystyle \frac{\partial Q_{\mathit{CNS}}}{\partial X_{17}}}=c_{17}+\sum _m{b}_{43,m}{\displaystyle \frac{\partial a_{m,17}}{\partial X_{17}}}$$

(7)

where the first term represents direct consumption and the second term denotes indirect consumption (transmitted through the industrial chain).

These findings indicate a close interconnection between the computing power network and traditional industrial sectors. Computing power development relies on support from traditional industries while simultaneously driving their upgrading and transformation. This mutually dependent and reinforcing relationship constitutes the core mechanism of the economic effects of computing power and provides the policy basis for optimizing its allocation.

4. Construction and Solution of the Dynamic Optimization Model

Given the inherent mathematical complexity of navigating a 1290-dimensional stochastic control problem (43 sectors × 3 decision variables × 10 periods), Figure 1 presents a simplified conceptual diagram of the proposed model architecture to enhance structural interpretability.

The architecture explicitly distinguishes between the core economic mechanisms and the auxiliary computational components. The core mechanisms govern the fundamental economic rationale: the static 43-sector input–output linkage matrix maps the spatial inter-industry dependencies, while the dynamic stochastic general equilibrium (DSGE) equations map the temporal capital accumulation and state-space transitions. In contrast, the auxiliary components—comprising the GA-LP hybrid optimization algorithm, variable clustering for dimensionality reduction, and adaptive constraint handling—serve purely functional roles to ensure mathematical tractability and accelerate parallel computation.

Intuitively, this deeply integrated architecture demonstrates distinct advantages over simpler modeling approaches. While traditional pure IO models function as static ‘snapshots’ incapable of capturing long-term investment effects, and loosely coupled hybrid models treat inter-period transitions as disjointed black boxes, the proposed framework acts as an actively evolving network. It distinctly captures how contemporary investments in the computing power sector not only stimulate immediate demand but also recursively ripple through the traditional industrial chain, endogenously upgrading the economy’s structural capital stock over multiple future periods.

4.1. State Space Transition Mechanism

Whilst static IO models reveal inter-industry linkages, they fail to capture the dynamic evolution of economic development. To analyses the long-term economic impact of computing power networks, the static model must be extended into a dynamic framework. Dynamic IO models simulate the evolutionary trajectory of economic systems by incorporating capital accumulation, technological progress, and temporal dimensions.

This study achieves the transition from a static IO model to a dynamic model by incorporating a technological progress factor:

$$A_t^{‘}=A_0^{‘}\mathsf{\odot }\mathit{exp}\left(\Gamma {\int }_0^t{I}_{R\&R,\tau }d\tau \right)$$

(8)

where $A_t^{‘}$ denotes the direct consumption coefficient matrix for period t; $A_0^{‘}$ denotes the direct consumption coefficient matrix for the base period; $\mathsf{\Gamma }$ denotes the technology learning matrix (diagonally dominant), determined through industry expert surveys (the complete Delphi protocol and parameter values are detailed in Appendix B.3); ${\mathrm{I}}_{\mathrm{R}\&\mathrm{R},\mathsf{\tau }}$ denotes the R&D investment intensity, expressed as a percentage of GDP; $\mathsf{\odot }$ denotes the Hadamard product (element-wise multiplication).

The technology learning matrix $\mathsf{\Gamma }$ captures variations in technological progress across sectors. Generally, technology-intensive sectors exhibit stronger learning effects, with their technical coefficients declining more rapidly. The R&D investment intensity $I_{R\&R,\mathsf{\tau }}$ reflects the pace of technological advancement: greater R&D investment accelerates progress.

This framework carries explicit economic implications: as time progresses and R&D accumulates, production techniques continually improve, leading to a gradual decline in direct consumption coefficients. That is, fewer intermediate inputs are required to produce a unit of output. Such technological progress encompasses both process innovation (enhancing resource efficiency) and product innovation (creating new goods and services).

4.2. Economic Basis of the Objective Function

The objective of the dynamic optimization problem is to maximize the social welfare function. This function is constructed based on consumption, synergistic efficiency, and environmental quality, employing the Stone–Geary form:

$$W_t=\alpha \mathit{ln}\left(C_t\right)+\left(1-\alpha \right)\mathit{ln}\left(G_t\right)- \gamma E_t$$

(9)

where $C_t= \sum w_{\mathit{it}}{Y}_{\mathit{it}}$denotes the consumption basket, $w_{\mathit{it}}$ represents the IO final use coefficients reflecting the weight of each sector’s products in consumption; $G_t=G_{\mathit{syn},t}·\prod \left(1+{\mathsf{\lambda }}_{\mathit{it}}{b}_{\mathit{it}}\right)$ signifies the synergistic gain from computing-network convergence; $E_t={\left({\displaystyle \frac{{\mathit{Carbon}}_t}{{\stackrel{\mathrm{-}}{E}}_t}}\right)}^{\theta }$constitutes the environmental cost term, $\theta$ = 1.5 constructs a convex penalty function.

Parameter calibration is based on historical data and expert assessments (see Appendix B.3.2 for detailed justifications): $\alpha$ = 0.7 (consumption weighting), reflecting consumption’s relative importance in social welfare; $\gamma$ = 0.25 (environmental cost coefficient), based on carbon trading prices and environmental impact assessments; ${\lambda }_{\mathit{it}}$ takes values within [0.01, 0.15] (synergy coefficient), based on industrial linkage strength and network effect evaluations. This social welfare function possesses several advantages: firstly, its logarithmic form aligns with the law of diminishing marginal utility; secondly, it encompasses consumption, synergistic, and environmental dimensions to comprehensively reflect the essence of social welfare; thirdly, it incorporates synergistic amplification effects to capture the positive externalities of network economies; fourthly, it employs a convex penalty function to ensure strict enforcement of carbon emission constraints.

4.3. Dynamic Constraint System

The dynamic optimization problem must satisfy a series of constraints, including production constraints, resource constraints, and environmental constraints.

4.3.1. Capital Accumulation Equation

Capital accumulation constitutes the core driver of economic growth. Accounting for investment adjustment costs, the capital accumulation equation is expressed as:

$$K_{i,t+1}=\left(1-{\delta }_i\right)K_{\mathit{it}}+I_{\mathrm{it}}-{\displaystyle \frac{\varphi }{2}}{\left({\displaystyle \frac{I_{\mathit{it}}}{K_{\mathit{it}}}}\right)}^2{K}_{\mathit{it}}$$

(10)

where ${\delta }_i$ denotes the sectoral depreciation rate, ranging from [0.05, 0.12] and determined based on fixed asset lifespans; $\varphi$ represents the investment adjustment cost parameter, positively correlated with sectoral linkage intensity; $I_{\mathit{it}}$ denotes the investment amount for sector i in period t.

The investment adjustment cost reflects the frictional cost of capital reallocation. When the investment rate is high, adjustment costs increase, slowing the pace of capital accumulation. This specification aligns with reality, as rapid capital accumulation necessitates higher adjustment costs.

4.3.2. Computing Power Supply–Demand Balance

The equilibrium between computing power supply and demand constitutes a crucial constraint within the model:

$${\displaystyle \sum _{i=1}^{42}}{\mathit{CNS}}_{\mathrm{it}} \le { Q}_{\mathit{CNS},t} = Q_0{\displaystyle \prod _{\tau =0}^t}\left(1+g_{\tau }\right)+ \mathsf{∆} Q_{\mathit{social},t}$$

(11)

Computing power demand drives growth:

$$g_t= 0.2\times {\displaystyle \frac{{\sum }_j{b}_{43,j}\mathsf{∆}{Y}_{j,t}}{{\sum }_j{b}_{43,j}\mathsf{∆}{Y}_{j,t-1}}}$$

(12)

where ${CNS}_{it}$ denotes computational power allocated to sector $i$; $Q_{cns,t}$ represents total computational power supply capacity; $Q_0$ signifies initial computational power supply capacity; $g_t$ denotes computational power demand growth rate; $\mathsf{∆} Q_{\mathit{social},t}$ indicates societal computational power integration volume.

The computational power demand growth rate correlates with the economic growth rate; the faster the economic growth, the quicker the computational power demand increases. The societal computational power integration volume reflects idle computational resources consolidated through the sharing economy model.

4.3.3. Cumulative Carbon Emission Equation

The carbon emissions accumulation equation captures the environmental impact of economic development:

$${\mathit{Carbon}}_t={\sum }_{i=1}^{43}{e}_{\mathit{it}}{Y}_{\mathit{it}}+\epsilon \int {\int }_0^t{Q}_{\mathit{CNS},\tau }d\tau$$

(13)

Energy structure influence factor:

$$\epsilon =0.89·\left(1-{\displaystyle \frac{E_{\mathit{green}}}{E_{\mathit{total}}}}\right)$$

(14)

where $e_{\mathit{it}}$ denotes the carbon emission intensity of sector $i$; $\epsilon$ represents the intrinsic carbon emission intensity of computing facilities; $E_{\mathit{green}}$ indicates the proportion of green electricity consumption; $E_{\mathit{total}}$ signifies the total electricity consumption.

Cumulative carbon emissions comprise two components: direct emissions from production processes and indirect emissions from computing facility operations. A higher proportion of green electricity reduces the carbon intensity of computing operations.

4.4. Mathematical Formulation of the Optimization Problem

The dynamic optimization problem is formulated as follows:

$$\underset{\left\{\mathit{CNS},I,Q\right\}}{\mathit{max}}{E}_0{\displaystyle \sum _{t=0}^T}{\beta }^t{W}_t$$

(15)

$$\begin{gathered} \mathrm{subject} \mathrm{to}: \\ \left\{\begin{array}{c}{X}_t=B_t{F}_t\\ K_{i,t+1}=\Psi \left(I_{\mathit{it}},K_{\mathit{it}}\right)\\ {\sum }_i{\mathit{CNS}}_{\mathit{it}} \le Q_{\mathit{CNS},t}\\ {\mathit{Carbon}}_t \le {\stackrel{\mathrm{-}}{E}}_t\end{array}\right. \end{gathered}$$

Dimensional analysis of the problem: 43 sectors × 3 decision variables × 10 periods = 1290 dimensions, constituting a high-dimensional stochastic control problem. The decision variables comprise: ${\mathit{CNS}}_{\mathit{it}}$ represents the computing power usage of sector $i$ in period t; $I_{\mathit{it}}$ represents the investment amount of sector $i$ in period t; $\mathsf{∆} Q_{\mathit{social},t}$ represents the aggregate social computing power in period t.

5. Empirical Analysis and Policy Simulation (China)

Based on the constructed model and algorithm, this section conducts an empirical analysis of China’s computing-power economy, simulating economic and environmental impacts under different policy scenarios.

5.1. Empirical Analysis of Industrial Linkage Effects

Based on China’s 2020 IO Tables and field survey data (see Appendix C Table A2 for the complete sectoral results), the complete dependency on computing power and its pull effect across sectors were calculated.

Table 2. Top 5 Sectors by Full Dependency Coefficient: Forward and Backward Linkage Rankings (China).

Rank	Computing Dependency ${(\mathbf{b}}_{43,\mathbf{j}})$	Sector	Computing Power Multiplier Effectiveness ${(\mathbf{b}}_{\mathbf{j},43})$	Sector
1	0.324	Internet Services	0.201	Semiconductors
2	0.287	Artificial Intelligence R&D	0.189	Electricity
3	0.253	Fintech	0.175	Server Manufacturing
4	0.231	Cloud Computing	0.162	Cooling Equipment
5	0.218	Big data analytics	0.158	Optical Equipment

below lists the top five sectors by complete demand coefficient.

In terms of complete computing power dependency, the internet services sector ranks first at 0.324, indicating that every ¥10,000 increase in final demand within this sector necessitates the consumption of ¥3240 worth of computing power services. Artificial intelligence R&D and fintech follow in second and third place with dependency ratios of 0.287 and 0.253, respectively. This outcome aligns with expectations, as these sectors are computationally intensive and their operations rely heavily on computational support.

In terms of computing power’s multiplier effect, the semiconductor sector leads with a multiplier coefficient of 0.201. This indicates that every ¥10,000 increase in final demand within the computing power sector fully stimulates ¥2010 output within the semiconductor sector. Electricity and server manufacturing follow in second and third place, with multiplier coefficients of 0.189 and 0.175, respectively. This outcome reflects the substantial hardware support—including semiconductors, electricity, and servers—required for the development of the computing power sector.

Computing power multiplier calculations reveal that$M_{\mathit{CNS}} = 2.37$, demonstrating significant economic multiplier effects from computing power investment. Each unit increase in final demand within the computing power sector fully stimulates 2.37 units of total economic output through industrial linkage networks. This multiplier effect stems primarily from two sources: firstly, direct and indirect demand from the computing power sector for upstream sectors like semiconductors and electricity; secondly, computing power as a general-purpose technology enhancing production efficiency in downstream sectors. From an energy perspective, this multiplier effect is closely associated with increased electricity demand in upstream sectors and improved energy efficiency in downstream sectors, indicating that computing power networks reshape both the scale and efficiency of energy use across industries.

5.2. Optimization Results Under the Baseline Scenario

Before reporting relative changes, it is instructive to present the absolute benchmark values underlying the analysis. China’s GDP in 2023 was approximately ¥126.5 trillion, with the computing power sector contributing ¥2.8 trillion in direct output. Total national energy consumption stood at 5.72 billion tons of standard coal equivalent (tce), of which data centers and computing facilities accounted for approximately 0.27 billion tce (4.7%). Total carbon emissions were approximately 10.6 billion tons of CO₂, with the ICT sector contributing about 5% [3,8]. Under the baseline optimization scenario, projected GDP reaches ¥198.7 trillion by 2030, while computing power sector output grows to ¥4.8 trillion. Energy consumption by computing facilities is projected to reach 0.52 billion tce by 2030 under baseline assumptions. These absolute values provide the foundation for the percentage changes discussed below. (See also Appendix C Table A2 for the complete sector-level results matrix and Appendix C 

Table A3. Absolute Values of Key Indicators (2023–2030).

Year	Total Output (¥ trillion)	Computing Power Sector Output (¥ trillion)	Carbon Emission Intensity (t CO2/¥10,000)	Computing Power Utilization Efficiency (%)
2023	126.5	2.8	0.84	62
2025	142.3	3.5	0.78	68
2027	158.7	3.9	0.72	73
2029	175.2	4.4	0.67	79
2030	198.7	4.8	0.63	84

for the objective function value time series.)

Turning to the optimization results, the social welfare function values W_t under both the baseline and optimal policy scenarios are reported in Appendix C 

Table A4. Social Welfare Function Values W_t (2025–2035).

Year	Baseline Scenario W_t	Optimal Policy Scenario W_t
2025	9.452	9.461
2026	9.681	9.703
2027	9.893	9.941
2028	10.127	10.213
2029	10.352	10.489
2030	10.574	10.764
2031	10.789	11.038
2032	10.997	11.302
2033	11.198	11.562
2034	11.391	11.817
2035	11.574	12.057

Note: Cumulative welfare improvement (optimal vs. baseline): +4.17% by 2035.

. Under the baseline trajectory (current policy settings), welfare increases from 9.452 in 2025 to 11.574 in 2035—a steady but unaccelerated improvement driven primarily by autonomous technological progress. Under the optimal policy mix, welfare reaches 12.057 by 2035, representing a cumulative improvement of 4.17% over the baseline. The welfare divergence becomes measurable after 2028 and accelerates through 2035, consistent with the compound effect of optimized computing power allocation on sectoral output, synergy externalities, and environmental performance. Notably, the welfare improvement is not monotonic in the early years (2025–2027), reflecting the transition costs of reallocating computing power from incumbent sectors to growth sectors—a finding that underscores the importance of medium-term policy commitment.

5.2.1. Evolution of Optimal Computing Power Allocation

During the period 2023–2030, the proportion of computing power quotas allocated to each sector exhibits distinct temporal evolution characteristics:

Initial Phase (2023–2025): Computing power shifted towards the service sector, with its share declining from 57% to 51%. This period marked a digital transformation phase, characterized by rapid growth in computing demand within the service sector, particularly digital services such as finance and e-commerce.

Intermediate Phase (2026–2028): The manufacturing sector’s share of computing power increases from 21% to 22%. With the advancement of smart manufacturing and industrial internet, computing power demand within manufacturing begins to materialize.

Late Stage (2029–2030): Structural stabilization phase, with ICT manufacturing reaching 31%, traditional manufacturing holding steady at 19%, and services maintaining 50%. Computing power allocation stabilizes, reflecting a new equilibrium in economic structure.

Investment inflection point analysis: Computing power investment share reaches an inflection point in 2026, declining from 18.7% to 12.3%. This aligns with the digital infrastructure cycle of the 14th Five-Year Plan. The period 2023–2025 represents a concentrated construction phase for computing infrastructure, characterized by significant investment intensity. Post-2026 marks the application-dominated phase, where investment priorities shift from infrastructure development towards application development and service innovation.

5.2.2. Economic and Environmental Impacts

Changes in key economic and environmental indicators following optimization are presented in

Table 3. Changes in Key Indicators for 2023–2030 (China).

Indicator	2023	2030	Change Rate	Average Annual Growth Rate
Total Output (trillion yuan)	126.5	198.7	+57.1%	+6.7%
Share of Computing Power Investment	18.7%	12.3%	−34.2%	−5.2%
Carbon emissions intensity (tons/10,000 yuan)	0.89	0.62	−30.3%	−4.6%
Computing power utilization efficiency	43.2%	68.7%	+59.0%	+7.1%
Total Factor Productivity	1.00	1.38	+38.0%	+4.7%

Note: The indicator Total Output (¥126.5 trillion in 2023, ¥198.7 trillion in 2030) provides the absolute GDP benchmark underlying the percentage changes. Absolute annual values for all indicators are available in Appendix C Table A3.

.

Economic Impact Analysis: Computing power investment drove a 38% increase in total factor productivity, contributing 24.3% to economic growth. The spillover effects of computing power networks were pronounced, notably in accelerating the intelligent transformation of manufacturing and the digital transition of the service sector. These economic effects are accompanied by changes in energy demand patterns, particularly through increased electricity consumption in digitally intensive sectors and improved energy efficiency driven by technological upgrading.

Environmental Impact Analysis: Carbon emission intensity decreased by 30.3%, reflecting a systematic improvement in both energy efficiency and energy structure. This reduction is primarily driven by three mechanisms: firstly, relocating computing power to the west to utilize clean energy resources; secondly, enhancing energy efficiency through computational optimization; and thirdly, facilitating industrial restructuring towards lower energy consumption.

It should be noted that these quantitative results are specific to the baseline parameter configuration. The sensitivity analysis (Section 5.2.3,

Table 4. Sensitivity Analysis: Parameter Ranges and Result Stability (China).

Parameter	Baseline	Tested Range	Sensitivity
Depreciation rate (δ)	0.05–0.12	0.03–0.20	Inflection point shifts ±1 year; multiplier changes < 5%
Environmental penalty (φ)	1.5	1.2–2.0	Carbon reduction varies ±4%; sectoral ordering unchanged
Synergy coefficient (γ)	0.01–0.15	0–0.30	Welfare gain varies ±1.2%; allocation trajectory stable
Discount factor (β)	0.95	0.90–0.98	2030 quota changes ±3%; inflection point robust
R&D intensity path	Baseline	±20%	Multiplier 2.37 ± 0.15; sectoral direction unchanged

) confirms directional robustness-namely, the pattern of sectoral computing power allocation, the inflection point in investment share, and the relative ordering of policy effectiveness—while showing that precise magnitudes vary across parameter ranges. Consequently, these estimates should be interpreted as scenario-conditional projections rather than precise forecasts.

5.2.3. Sensitivity Analysis

To further interpret these empirical findings, three underlying economic dimensions warrant discussion. First, regarding sectoral heterogeneity, sectors such as ICT manufacturing (e.g., semiconductors) and electricity exhibit stronger multiplier effects because they serve as foundational nodes in the computing ecosystem. Computing power acts as a General Purpose Technology (GPT); its expansion generates intensive backward linkages to electricity (for operational energy) and ICT manufacturing (for hardware capital accumulation), while simultaneously exerting broad forward linkages to empower downstream industries. Second, concerning the drivers of these results, the underlying structural transition—such as the shift from services to manufacturing—is fundamentally driven by the empirical IO data, which reflects the physical realities of China’s industrial interconnection. Conversely, the precise magnitudes and optimal dynamic paths (e.g., the 31% computing power quota by 2030) are primarily governed by the DSGE model’s structural constraints, representing a rigorous synthesis of empirical data and optimization logic. Finally, sensitivity analyses on alternative parameter settings (such as capital depreciation rates and substitution elasticities) confirm the robustness of the core findings. While minor deviations occur in the absolute investment volumes, the aggregate computing power multiplier (2.37) and the overall sectoral allocation trajectories remain structurally stable.

5.3. Policy Combination Simulation

5.3.1. Effects of Individual Policies

Simulations of the individual effects of three policy instruments yielded the results presented in

Table 5. Comparison of Single Policy Effects (2030) (China).

Policy Instrument	Direct Effect	Indirect Effect	Total Effect ΔGDP	Change in Carbon Emissions	Change in Computing Efficiency
Eastern Computing Power Premium Tax	−0.8%	+2.1%	+1.3%	−6.2%	+0.04
Western Green Electricity Subsidy	−0.3%	+1.8%	+1.5%	−12.9%	−0.01
Research and development tax credit	−1.2%	+3.5%	+2.3%	−3.4%	+0.11

:

Policy Transmission Mechanism:

The eastern computing power premium tax guides computing demand westward through price signals. The direct effect is increased computing costs in eastern regions (−0.8%), but the indirect effect is reduced carbon emission costs by utilizing western clean energy (+2.1%), resulting in a net effect of +1.3%.

Western green electricity subsidies reduce computing carbon intensity, with a direct effect of increased fiscal expenditure (−0.3%) and an indirect effect of enhanced low-carbon competitiveness (+1.8%), yielding a net effect of +1.5%.

R&D tax credits stimulate computing technology innovation, with a direct effect of reduced tax revenue (−1.2%) and an indirect effect of increased total factor productivity through technological advancement (+3.5%), yielding a net effect of +2.3%.

Industrial linkage amplification effect: The indirect effect of R&D tax credits reaches 2.9 times the direct effect, primarily through amplified effects generated via industrial chain linkages. Computing power technological innovation drives the development of related industries such as semiconductors and software, forming a virtuous cycle.

5.3.2. Optimal Policy Combination

The optimal policy parameter combination was obtained through grid search:

$$\left\{\begin{array}{c}{P}_{\mathit{east}}=1.25\times P_{\mathit{base}}\\ Subsidy=0.18E_{\mathit{green}}\\ {\mathit{Tax}}_{\mathit{deduct}}=1.45\times RD\end{array}\right.$$

(16)

This policy combination increases social welfare by 4.17% while reducing carbon emissions by 12.8%. Synergistic effects exist between policy instruments, manifested as:

Price-subsidy synergy: Combining eastern premium taxes with western green electricity subsidies both guides computing power westward and safeguards western regions’ computational competitiveness.

Technology and market synergy: Integrating R&D tax credits with price signals both fosters technological innovation and ensures market-driven technology adoption

Short-term and long-term synergy: Pricing and subsidies primarily influence short-term resource allocation, while R&D tax credits impact long-term technological advancement. The integration of these three elements achieves coordination between short- and long-term objectives.

5.4. Optimization of Carbon Emission Reduction Paths

Based on optimization results, a three-phase carbon reduction pathway is proposed:

Short term (2023–2025): Computing power westward migration phase. Relocate computing facilities from eastern to western regions, leveraging abundant clean energy in the west. Projected reduction in carbon emissions per unit of computing power: 41%. Western computing power share: 35% to 58%. Key measures include: establishing western computing hub nodes; constructing supporting ultra-high voltage transmission corridors; implementing an eastern-western computing power coordination mechanism.

Midterm (2026–2028): Technology-driven efficiency enhancement phase. Enhancing computational energy efficiency through technological innovation, particularly in semiconductor chips. Projected reductions include a 23% decrease in semiconductor energy consumption and a 12% reduction in associated sectoral carbon emissions. Key areas encompass: R&D and deployment of advanced process chips; promotion of high-efficiency cooling technologies such as liquid cooling; AI-enabled computational resource scheduling optimization.

Long term (2029–2030): Structural Optimization Dominant Phase. Reduce the proportion of high-energy-consumption computing applications through industrial restructuring. The share of high-energy-consumption computing applications is projected to decrease from 45% to 32%. Key pathways include: developing lightweight algorithms and models; optimizing computing resource allocation mechanisms; establishing a linked trading market for computing power and carbon emissions.

6. Discussion and Policy Implications (China)

6.1. Conclusions

The following findings are derived from the integrated IO-DSGE model under the specific parameter calibrations and structural assumptions documented in Section 3 and Section 4 and Appendix B and Appendix C. While sensitivity analysis (Section 5.2.3, Table 4) confirms the directional robustness of the main results, the precise quantitative magnitudes-including the multiplier coefficients, sectoral allocation percentages, and policy effect sizes-should be interpreted as scenario-dependent estimates conditional on the model’s assumptions rather than exact predictions.

This study employs a dynamic IO model embedded within computing power networks and develops a hybrid optimization algorithm to conduct a systematic analysis of China’s computing-power economy. Key findings are as follows:

First, the computing power network exhibits significant economic multiplier effects. The computing power multiplier reaches 2.37, meaning that each unit increase in final demand within the computing power sector fully stimulates 2.37 units of total economic output through industrial linkage networks. This multiplier effect primarily stems from forward linkages (computing power driving demand in upstream sectors) and backward linkages (computing power enhancing productivity in downstream sectors).

Second, power allocation exhibits distinct sequential evolutionary patterns. The period 2023–2025 represents a digital transformation phase, characterized by computing power shifting towards the service sector. The years 2026–2028 mark an intelligent manufacturing boom, with rapid release of computing power demand in manufacturing. The period 2029–2030 constitutes a structural stabilization phase, where computing power allocation structures tend towards equilibrium. This evolutionary trajectory is closely linked to the developmental stages of the digital economy.

Third, policy instruments exhibit markedly divergent economic impacts. Research and development tax credits yield the greatest overall effect (ΔGDP = +2.3%), with an industrial linkage amplification factor of 2.9 times; green electricity subsidies in western regions achieve the most pronounced carbon emission reductions (−12.9%); while a computational power premium tax in eastern regions delivers a favorable composite outcome (ΔGDP = +1.3%, carbon emissions −6.2%).

Fourth, carbon reduction requires multi-stage coordinated advancement. Short-term progress relies on relocating computing power westward to utilize clean energy; medium-term progress depends on technological innovation to enhance energy efficiency; long-term progress hinges on industrial structure optimization. This multi-stage strategy balances both technical and economic feasibility with the sustainability of emission reduction outcomes.

6.2. Theoretical Contributions

The theoretical contributions of this study are primarily reflected in three aspects:

First, it establishes an IO-DSGE integrated framework. By organically combining static IO analysis with a dynamic stochastic general equilibrium model, it captures the characteristics of industrial interconnection networks while analyzing dynamic accumulation processes, resolving theoretical challenges in measuring the computational power economy. This framework provides methodological insights for studying the economic impacts of other networked infrastructure.

Second, it introduces the concept and measurement method of the computing power multiplier. Through complete demand coefficient matrix decomposition, it quantifies both the forward and backward linkage effects of the computing power sector, providing a scientific basis for assessing its economic contribution. The computing power multiplier ($M_{\mathit{CNS}}$ = 2.37) fills a gap in quantifying the economic effects of computing power.

Third, a GA-LP hybrid optimization algorithm was designed. Through variable clustering for dimensionality reduction, adaptive constraint handling, and parallel computation acceleration, it resolves the challenge of solving high-dimensional dynamic optimization problems, offering a novel approach for complex economic system optimization. The algorithm’s convergence proof and speed analysis provide theoretical reference for similar problems.

6.3. Policy Implications

Based on the research findings, the following policy recommendations are proposed:

First, refine the computational power market system (Long-term structural goal). Establish a computational power quota trading market implementing tiered pricing based on marginal contribution (${\displaystyle \frac{\partial V*}{\partial {\mathit{CNS}}_i}}$) to guide computational resources toward high-efficiency sectors. Specific measures include: Establishing a computational power resource rights confirmation and trading system beginning with piloting regional computing exchanges in major tech hubs (e.g., Shenzhen or Shanghai), gradually developing computational power derivatives such as futures and options, and establishing a computational power value assessment standard system.

Second, optimize computing power investment mechanisms (Short- to Medium-term transition). To address the growth-energy trade-off, investment must pivot from energy-intensive hardware to efficiency. Set a computing power investment subsidy rate (s = 0.2 (1 − ${\displaystyle \frac{U_c}{100}}$)) to enhance computational efficiency. Shift investment focus from infrastructure construction to application innovation, gradually reducing computing power investment intensity from 18.7% to 12.3% after 2026.

Third, innovate environmental policy instruments (Medium- to Long-term strategy). To mitigate the environmental impact identified in the baseline analysis, environmental constraints must be institutionalized. Implement a computing power-carbon credit trading mechanism with a transaction volume = (${\displaystyle \frac{\partial \mathit{Carbon}}{\mathsf{\partial }\mathit{CNS}}}·\Delta CNS$), integrating computing power carbon emissions into the carbon market. Concurrently, establish computing power carbon emission accounting standards and refine the monitoring system for such emissions.

Fourth, advance regional coordinated development (Short-term operational mechanism). Deepen the “East Data, West Computing” initiative by establishing a coordinated development mechanism between eastern and western computing power. For example, local governments can establish specific “green electricity direct supply agreements” connecting western renewable energy grids directly to computing clusters. This ensures eastern regions focus on application development while western regions provide green energy guarantees, forming complementary regional advantages.

Fifth, strengthen international computing power cooperation (Long-term strategic vision). Advance the interconnection of computing infrastructure around the “Digital Silk Road” initiative. Actively participate in the formulation of global computing power governance rules to enhance China’s international voice in this domain.

It is important to emphasize that the above policy recommendations are conditional on the model-based analysis presented in this study. Their quantitative basis depends on the specific modeling assumptions, parameter choices (see Appendix B.3), and data sources employed. The effectiveness of each policy instrument in practice would depend on actual implementation conditions, institutional contexts, and unforeseen economic developments not captured within the current modeling framework. These recommendations should therefore be considered as indicative guidance informed by the model results, subject to refinement as more empirical evidence becomes available.

6.4. Research Outlook

Future research may explore the following directions:

First, incorporate uncertainty analysis. Account for variables such as technological breakthroughs, energy pricing, and international conditions to explore multiple potential trajectories for computational power economic development, thereby enhancing policy robustness.

Second, refine spatial dimension analysis. While this study establishes a robust macro-level framework for assessing computing power networks, several data and aggregation limitations must be acknowledged. First, the reliance on a 43-sector structural classification, though methodologically standard, may mask intra-sector heterogeneity, particularly within rapidly evolving digital sub-sectors. Second, the current utilization of national-level input–output data restricts the model’s capacity to fully capture granular regional disparities, which is critically relevant given the inherent spatial asymmetries in computing supply and demand. Finally, the calibration of specific computing power inputs unavoidably depends on secondary industry reports (e.g., CAICT, IDC) and estimated survey data, which introduces potential measurement uncertainties. Acknowledging these constraints provides a clear trajectory for subsequent enhancements. Future research should strive to integrate multi-regional input–output (MRIO) tables and micro-level enterprise panel data. Expand the national model into regional models to examine the spatial distribution characteristics of computing power resource allocation, thereby optimizing the spatial layout of the “East Data, West Computing” initiative.

Third, deepen international comparative research. Analyze the computational power economic development models of different countries, summaries lessons learned, and refine China’s computational power economic development strategy.

Fourth, explore the impact of emerging technologies. Investigate the effects of next-generation computing technologies such as quantum computing and neuromorphic computing on economic systems, and proactively plan for future computing development.