Author | Zhiyu For more financial insights | BT Finance Data Pass This article contains 2,364 words and takes approximately 9 minutes to read.

In March of this year, Mao Shengyong, Deputy Director of the National Bureau of Statistics, revealed a striking figure at a press conference:

China's average daily AI token calls surpassed 140 trillion in March.

Many may glance at the figure '140 trillion' and quickly move on.

Yet, behind this staggering number lies a critical question that few have seriously considered: How are these 140 trillion tokens generated? What resources are consumed in the process? And who truly benefits from this immense activity?

1

How Significant Is 140 Trillion Tokens?

To put this into perspective: If we imagine each of these 140 trillion tokens as a person saying 'hello'—

Speaking continuously, with one 'hello' per second, it would take a staggering 4.4 million years to complete just a single day's worth of token calls.

This is not mere rhetoric; it reflects the actual computational scale of China's AI industry in just one day.

Moreover, this number is continuously growing. Data from the National Bureau of Statistics indicates a more than 40% increase from late 2025 to March.

The question then arises: What underpins such massive computational activity?

2

What Fuels a Single Token Call?

▍Three Essential Infrastructure Layers

An AI request, from initiation to completion, relies on three fundamental requirements:

The first layer is computational power. Every AI model inference operates on GPUs or NPUs. By 2026, NVIDIA's H-series chips dominate China's AI platforms, although Huawei's Ascend 910B is gaining traction. In March, Alibaba Cloud raised prices for some large models, citing escalating computational costs (Source: Sina Finance).

The second layer is electricity. Data centers are notoriously energy-intensive. A medium-sized (10MW) AI data center consumes approximately 87.6 million kWh annually. At an average cost of ~0.6 RMB/kWh (2026 average), electricity expenses exceed 50 million RMB yearly (Estimates based on public data; actual costs vary by region and energy contracts).

The third layer is networking. Every request and response travels through fiber-optic networks. When you ask your phone's AI, 'What's the weather today?', your query and the response traverse thousands of kilometers of fiber.

Daily token calls: 140 trillion (National Bureau of Statistics, March 2026 press conference)

Year-over-year growth since late 2025: +40%+ (Same source)

Alibaba Cloud's model price increases: March-April 2026, 20-30% hikes for select models (Sina Finance)

China's AI computing landscape: NVIDIA-dominated, with Huawei Ascend gaining market share (Public market data)

Data center PUE range: Approximately 1.35-1.5 domestically (Source: MIIT reports)

3

Who Benefits from These Three Layers?

▍Computational Layer: Chips Take Center Stage

NVIDIA emerges as the biggest winner in global AI computing, reporting $47.5 billion in data center revenue for FY2025 (a 142% year-over-year increase) (Source: NVIDIA FY2025 annual report). Despite being a U.S. company, its chip exports directly influence China's AI computational capabilities.

Huawei Ascend: The flagship for domestic computing substitution. The 910B's performance approaches that of NVIDIA's A100, driving its increasing adoption amid supply constraints.

Cambricon: Another key player in domestic AI chips, with revenue surging over 100% year-over-year in 2025 (Source: Company annual report), although still smaller than industry leaders.

▍Electricity Layer: Data Centers as Major Energy Consumers

AI data centers represent the most predictable new source of demand for the power sector in recent years.

Liquid cooling systems are crucial: High-density data centers utilizing liquid cooling can reduce PUE (energy efficiency) from 1.4 to below 1.2, resulting in significant cost savings (Vendor technical whitepapers; values vary by deployment). Orders for liquid cooling providers like Envicool and Shenling Environmental have surged alongside AI computing growth.

▍Networking Layer: Fiber Optics as the Information Arteries

Every AI request depends on fiber networks. YOFC, Hengtong Optic-Electric, and Zhongtian Technology are China's leading fiber suppliers, with orders growing rapidly in 2025 (Source: Company annual reports).

AI is not sustained by developers alone—fiber sellers, chip vendors, and power suppliers all reap the benefits.

4

Where Will the Next 10-Fold Growth Originate?

140 trillion is already an enormous figure. Yet, it is merely the beginning.

Currently, AI is primarily utilized by large enterprises and high-net-worth users. True mass adoption will occur when AI costs decrease to levels affordable for small and medium-sized enterprises (SMEs).

Industry trends suggest that mainstream large model token costs decline by approximately 50% every 6-12 months (Consolidated from multiple securities firms and research reports). When prices reach a critical threshold, 'AI for all' scenarios will explode among SMEs, driving new growth across the supply chain.

Behind the 140 trillion tokens lie three often-underestimated supply chains: computational power, electricity, and fiber optics. Their growth is no less rapid than that of AI itself.

To truly understand the AI industry, don't just focus on model parameters. Examine how much electricity it consumes, which chips it utilizes, and how much fiber it deploys—these are the real foundations.

This article is intended solely for informational and industry analysis purposes and does not constitute investment advice, analysis, or a solicitation to trade. Markets carry risks; invest cautiously. Any investment decisions made based on this content are at your own risk, and the author and publishing platform assume no liability for gains or losses.

Sources

1. National Bureau of Statistics: March 2026 press conference (Daily token call data) by Deputy Director Mao Shengyong

2. Sina Finance: Reports on Alibaba Cloud's model price increases (March-April 2026)

3. NVIDIA FY2025 Annual Report (Data center revenue)

4. YOFC, Hengtong Optic-Electric, Zhongtian Technology 2025 Annual Reports

5. MIIT: Data center energy consumption and PUE policy documents

6. Cambricon 2025 Annual Report

This article is an original work by BT Finance and may not be used, reproduced, disseminated, or adapted without permission. Infringement will result in legal action.