Compute Token: The "New Oil" of the AI Era and a Trillion-Dollar Business

I. What is a Token? The "Unified Measure" of the Computing World

In the age of large AI models, a fundamental question has persistently remained unanswered: How do we measure the value of computing power? For a long time, the industry's evaluation metrics were blunt: how many GPUs, how many petaflops, how many parameters. But these figures fail to answer a core question: how much actual value has the user derived from the computing cost they paid?

The emergence of the token is precisely the answer to this "value question."

A token is the smallest unit of information that AI large models use to efficiently process data by breaking it down. Whether text, speech, or image/video, all are uniformly decomposed at the AI's underlying level into machine‑readable minimal processing units—tokens. Huang Wei, an expert at the China Academy of Information and Communications Technology (CAICT), points out that the token serves as a quantifiable voucher for computing resources, acting as a pricing medium for intelligent services. It transforms intangible technical capabilities into tangible, cost‑accountable commodities, bridging AI technology and economic value.

Think of it this way: just as the electricity industry uses the "kilowatt‑hour" to measure energy consumption and the "watt" to denote power, the computing industry also needs a common "language of value." Tokens are precisely that language—they transcend all data modalities—text, image, speech—and become the fundamental unit for measuring consumption of AI computing services.

In July 2026, at the Guanghe Organization Intelligent Computing Application Conference, the "Open Computing Token Taxonomy" initiative was officially launched, with nearly a hundred partners joining. This taxonomy covers three layers: compute production, resource scheduling, and value delivery. Through unified interfaces and open standards, heterogeneous computing resources can be encapsulated, measured, scheduled, and circulated. Tian Lihui, Dean of the Institute of Finance and Development at Nankai University, offers a vivid analogy: tokens "are like the electricity bill or fuel cost of the AI era."

II. Token Business Models: From "Selling Cards" to "Selling Tokens"

The rise of the token economy is fundamentally reshaping the business models of the computing industry.

Traditional model: compute leasing

Over the past two years, compute leasing has been one of the hottest segments in the capital market. In 2024, the global GPU server leasing market was valued at $5.912 billion, and is projected to expand to $23.62 billion by 2031. In the first quarter of 2026, the domestic compute leasing market in China had already reached tens of billions of yuan.

However, traditional compute leasing has clear pain points: users purchase fixed hardware units—a single server, a single GPU, or a fixed‑term compute capacity. The result is insufficient capacity during peak times and massive idleness during troughs. More critically, the most costly tasks—model deployment, performance tuning, power management—do not move with the lease contract. The average rack utilisation rate in many domestic AI data centres hovers only around 20‑30%, with some enterprise‑grade centres falling below 10%.

New model: Token factories

Token factories change the transaction object—customers directly purchase the token outputs produced by AI models, using them on demand, paying by volume, and settling based on actual usage. Users no longer need to pay a hefty fee for a specific server over a fixed period; they can pay as they would for water or electricity.

Currently, the participants in token services have formed a diverse landscape:

1. Cloud vendors and AI infrastructure platforms. Take a certain AI computing service platform as an example. Founded less than three years ago, it positions itself as an "open, independent token supply platform." Through its proprietary inference engine and compute orchestration system, it aggregates heterogeneous computing resources from different vendors and architectures, packaging them into standardised, meterable token services. By April 2026, its registered users had exceeded 10 million, with a daily token throughput of approximately 578.5 billion. In terms of annual token throughput, it was the fourth‑largest token supply platform in China in 2025.

2. The three major telecom operators. In 2025, all three major operators pivoted to token services as their core business line. China Telecom launched a personal package at 9.9 yuan/month for 10 million tokens; China Mobile turned AI model compute into a monthly subscription similar to a data plan; China Unicom rolled out Token Plan personal and team editions in multiple regions. Leveraging their nationwide backbone networks, abundant compute nodes, and a combined user base of 1.9 billion, the operators are transforming from "traffic pipe" providers into "compute service" providers.

3. Token aggregation and distribution platforms. Aggregators like OpenRouter have grown rapidly by relying on price spreads and commissions; there are also AI relay stations operating on cross‑border price differences. Currently, token pricing is extremely fragmented—in drug discovery, token prices average $1,000 per million, while casual chat costs only $0.01 per million, a gap of up to 100,000 times.

4. "Token export" as a new business. Locations such as Shanghai Lingang, Guangzhou Nansha, and Shantou have successively laid out token export businesses. The logic is to keep computing power, servers, and electricity within the country, while overseas customers remotely call API interfaces via fibre optics and pay based on usage.

5. Token revenue‑sharing models. Under the token revenue‑sharing model, compute providers and model developers enter a negotiation: the compute side, holding scarce GPU resources, has the right to determine base service prices and sharing ratios. After this model is implemented, the net profit margin of compute‑leasing companies is expected to rise from 15% to over 20%.

III. Prospects and Challenges: The Light and Thorns of a Trillion‑Dollar Market

Soaring demand: over 1,000‑fold growth in two years

China's average daily AI token calls surged from approximately 100 billion in early 2024 to over 140 trillion by March 2026—a more than 1,000‑fold increase in two years. In the first week of July 2026, China's AI large‑model calls reached 23.45 trillion tokens, ranking first globally for ten consecutive weeks. Projections indicate that China's total token calls will reach 306.3 trillion per day in 2026, and could hit 19,306 trillion by 2030. IDC forecasts that China's MaaS market token calls will reach 40,000 trillion in 2026, generating around 18.6 billion yuan in revenue.

Three main drivers are fuelling the sustained surge in token demand: continued skyrocketing compute procurement by leading AI labs for training and inference; vertical‑industry model companies entering a phase of large‑scale investment; and the rise of AI Agents triggering a new wave of compute consumption.

Challenges and concerns

However, the token economy is not without its hurdles—several challenges remain to be overcome.

1. Structural mismatch between supply and utilisation. On one hand, compute leasing prices remain high due to chip shortages; on the other, expensive GPUs sit idle in data centres. "Compute is getting more expensive, but tokens are getting cheaper"—domestic high‑quality compute resources remain tight, with prices having risen over 50% since before the Spring Festival of 2026.

2. Inconsistent pricing standards. Currently, token pricing in China is still primarily cost‑based, without a pricing mechanism linked to task completion or output value. Different vendors use different tokenisation rules for text and multimodal content, and the token consumption for the same content can vary by several times across vendors.

3. Profitability not yet proven. Take the aforementioned AI computing service platform as an example: in 2025, its revenue was 55.33 million yuan, a year‑on‑year increase of 653%, but its operating costs reached 68.63 million yuan, resulting in a gross margin of -24% and a net loss widening to 345 million yuan. For token factories to become truly profitable, further breakthroughs are needed in compute scheduling, cost control, and economies of scale.

4. Excessive supply‑side bargaining power. A Mu, Vice President of Lenovo Group, noted that the AI industry chain suffers from excessive supply‑side power, leading to structural imbalances. To fully unleash the value of tokens, a more balanced development between supply and demand must be promoted.

IV. Conclusion

Tokens are fundamentally reshaping the computing world and are becoming the core "currency" of China's AI industry. The shift from "selling cards" to "selling tokens" is not just a change in billing methods—it represents a wholesale reconfiguration of value distribution across the entire AI industry chain. Whoever can produce and distribute tokens more efficiently will hold the profit hub of the AI era.

As demonstrated by a certain AI data centre: a capacity of 200 billion tokens per hour proves that when AI capabilities are broken down into standardised, clearly settled token units, the technology leaves the laboratory and becomes true social infrastructure. The commercial inflection point of the token economy has arrived—it is transforming from a "future prophecy" into "today's reality."