Computing Power Stocks Surge! Brokers Mine Token "Word Element Economy" Investment Thesis

On March 25th, all three major A-share indices closed higher. The computing power sector experienced a collective surge, with the Wind East-Number West-Calculation and IDC (computing power leasing) concept indices rising by 4.98% and 4.25%, respectively. Several stocks, including Tianji Technology, TrueView, 263, Eurotech, and Dawi Technology, hit the daily limit-up.

In terms of news, the token officially received its Chinese name “词元” (word element). During the recent China Development High-Level Forum 2026 Annual Meeting, Liu Liehong, Director of the National Data Bureau, used “词元” as the Chinese translation for token in his speech.

Liu Liehong stated that in early 2024, China’s daily token call volume was 100 billion; by the end of 2025, it soared to 100 trillion; and in March this year, it surpassed 140 trillion, representing over a thousand-fold growth in two years.

He explained that “词元” (word element) is not only a value anchor in the intelligent era but also a “settlement unit” connecting technological supply and commercial demand, providing quantifiable possibilities for the implementation of business models.

The industry also sent strong signals. Just last week, NVIDIA CEO Jensen Huang proposed the concept of “Token Factory Economics” at GTC, stating that tokens will become a new commodity in the AI era. Future data centers will serve as factories producing tokens, with performance per watt becoming the core competitive advantage for commercial monetization.

Why is the “词元” (word element) igniting the capital markets? Under the “词元 economy” wave, how should investors seize this opportunity? Several brokerages have recently published research reports exploring the “词元 economy,” with key themes including computing infrastructure, model export, and compute-electric synergy.

From technical concepts to market focus

Why is the token “词元” so popular?

What exactly is the token “词元” that has sparked intense discussion in the capital markets?

A token is the smallest unit of information processed by large models. Technically, it segments natural language text into units that AI can understand, facilitating model computation; commercially, it is a measure of AI computing power costs, significantly impacting AI service pricing.

This concept, combining technical and economic attributes, has recently been widely discussed, reflecting a profound shift in the business logic of the AI industry.

The demand side’s explosion is the most intuitive. Recently, the phenomenal popularity of AI agent frameworks like OpenClaw has been seen as a direct driver of rapid token demand expansion.

Data from third-party AI model aggregation platform OpenRouter shows that during the week of March 9-15, 2026, OpenClaw contributed 20% of the platform’s token consumption, with weekly consumption equivalent to 60% of the platform’s total token usage in Q4 2025.

Price changes have also quietly begun. Since the start of 2026, the compute leasing market has entered a price increase cycle.

By the end of February, rental prices for high-end GPUs like NVIDIA H200 and H100 increased by 15%-30% month-over-month. Domestically, the expansion of token demand has also driven collective price hikes among vendors from model layers to cloud services. Major model companies like Zhipu, as well as cloud providers Alibaba Cloud and Baidu Cloud, have recently announced price increases for AI computing products.

Zhou Cheng, a computer industry analyst at Xiangcai Securities, summarized this trend as “volume and price rising together,” further noting that the nonlinear growth in token demand during the AI Agent era has disrupted the supply-demand balance of computing power, directly affecting costs for upstream hardware such as GPUs, enterprise storage, and CPUs. Under the dual pressure of rigid downstream demand and upstream hardware cost inflation, the cloud computing industry’s pricing logic is shifting toward premium monetization.

Regarding whether this price hike cycle can continue, many institutions believe that the supporting factors are unlikely to reverse in the short term.

CITIC Securities’ chief analyst team led by Ying Ying predicts that with higher-frequency inference requests and longer context requirements brought by OpenClaw, cloud resource utilization will further increase, and demand surges along with upstream cost transmission are expected to continue pushing cloud service prices upward.

Jiang Ying, Chief Analyst of the Communications Industry at Open Source Securities, also suggested that the proliferation of AI applications and the OpenClaw framework could trigger a surge in inference demand. Coupled with NVIDIA’s limited capacity, rising hardware costs, and the gap in domestic independence, the market is entering a “seller’s market,” and the price increase trend may persist.

Looking at the longer term, industry insiders and institutions generally believe that the rising prosperity of the token market is not a short-term pulse but a trend driven by the widespread adoption of AI applications.

At GTC, Jensen Huang proclaimed “Token is king.” He envisions future data centers as factories producing tokens, with performance per watt becoming the core competitive advantage for commercial monetization. Traditional architectures centered on server counts and storage capacity will gradually give way to new architectures focused on token generation rate and energy efficiency.

In terms of commercial implementation, Huang believes that tokens will become a new commodity, with tiered pricing based on speed and intelligence once mature—ranging from free to ultra-high-speed tiers (about $150 per million tokens), opening broader commercial space for inference scenarios.

Brokerage firms are exploring “词元经济” investment themes:

• Computing infrastructure, model export, and compute-electric synergy

From an investment perspective, Lu Wei, Chief Analyst of the Computer Industry at Guolian Minsheng Securities, pointed out that token demand “inflation” is likely to become a core AI investment theme this year, with related opportunities focusing on rapid growth in inference token demand.

How to grasp the “词元经济” investment opportunities amid demand “inflation”? Several brokerages have recently outlined related beneficiary sectors and targets in their research reports.

Specifically, the most directly benefiting sectors from the surge in token calls are the infrastructure and hardware segments, which currently enjoy high consensus among institutions.

Jiang Ying summarized the three core themes of the “Token Factory” as AIDC (AI Data Center), compute leasing, and CDN (Content Delivery Network). She sees “Token = AI chips (domestic compute + compute leasing) = AIDC.” Additionally, as token demand continues to grow, CDN demand is also expected to increase significantly.

Based on these three themes, Jiang Ying further identified five sub-sectors worth attention: AIDC data centers, liquid cooling for AIDC, power supply for AIDC, CDN, and AIDC computing and networking.

CITIC Securities’ TMT team led by Yan Guicheng also stated that “short-term fluctuations in the compute sector do not change the long-term growth logic,” continuing to recommend related stocks in the AI compute industry chain, including GPUs/CPUs, optical modules, optical chips, liquid cooling, and fiber optic cables.

Beyond underlying infrastructure, large model vendors as the upper application layer of compute resources are also expected to see new investment opportunities.

Lu Wei pointed out that large model vendors are shifting toward “selling tokens as fuel + selling results.” As inference becomes a production material, model vendors can convert “scarcity of compute” into profits and cash flow through tiered pricing and subscription products.

Notably, under this logic, domestic models are showing strong competitiveness, and “overseas token export” has been frequently mentioned in recent brokerage reports.

Shenwan Hongyuan Securities’ computer industry team led by Huang Zhonghuang estimates that domestic models are highly cost-effective compared to overseas counterparts, with overall costs about 1/6 to 1/10 of foreign models. This cost advantage stems from architecture improvements brought by DeepSeek and similar technologies, especially MLA and sparse architectures, which significantly reduce inference costs.

Lu Wei suggests continuing to monitor high-quality large model vendors. He believes that maintaining subscription retention and enterprise expansion in high-ROI scenarios like programming, agents, and enterprise workflows can convert “token usage” into tangible delivery value—saving labor, time, and rework—thus gaining resilience against open-source and price wars.

He also mentioned that “AI firewall” targets are worth attention. As enterprises embed AI into workflows, risks like data leaks and proxy overreach may drive demand for “AI security platforms/governance platforms.”

Additionally, “compute-electric synergy” is considered a key industry advantage supporting “overseas token export.”

Dongwu Securities’ computer team believes that green power hubs effectively reduce electricity costs, making low-cost electricity a core competitive advantage for token export. The new digital trade model of “electricity not crossing borders, compute value crossing borders” is becoming a key barrier for China’s AI participation in global competition.

The team further pointed out that within the compute-electric synergy sector, four types of targets hold core value: traditional power transformation companies that build data centers based on energy resources, with the highest valuation uplift; green power operators relying on low-cost renewable energy to supply long-term green power for compute clusters; dispatch software service providers that use algorithms to match load and electricity prices in real-time, improving operational efficiency; and power engineering leaders with experience in ultra-high-voltage and source-grid-storage construction, solidifying the physical foundation of synergy. Together, they form a “energy—compute” closed loop.