Tokens are booming! A comprehensive guide to the morphology economy industry chain

“Tokens are the new bulk commodity.” At NVIDIA’s 2026 GTC (GPU Technology Conference), NVIDIA founder and CEO Jensen Huang first introduced the token economy.

Huang proposed a formula: revenue = tokens per watt × available gigawatts. He explained that data centers have now become 24/7 “token factories,” taking in electricity and data and outputting tokens. And the revenue of a “factory” depends on the product of the efficiency and scale of token production.

Recently, Liu Liehong, head of China’s National Data Administration, said that by this March, China’s daily average token usage had exceeded 140 trillion. Compared with 100 billion at the beginning of 2024, it was up by more than 1,000 times.

The token economy— a new industrial chain is emerging.

What Is the Token Economy

Tokens are the basic unit used by large models to process information. When a user asks an AI model a question, the model first cuts the user’s words into tokens, then after calculating, it assembles the resulting tokens back into sentences. For each token generated, what is essentially happening is a call to the data center’s GPU computing power, accompanied by power consumption.

Therefore, tokens naturally function as a unit of measurement. API access for large model vendors is priced per token, cloud service providers price computing power per token, and tokens are to AI what “kilowatt-hours” are to electricity.

However, for quite a long time, tokens were only a concept of cost. From 2023 to 2024, each model competed on parameter scale and training data volume. Tokens were seen as costs, and no one had treated them as “products.”

The change came after AI moved into the inference stage. Over the past two years, AI has been rolling out at scale into commercial scenarios; every user conversation and task execution continuously consumes tokens. Under usage-based pricing models, many AI vendors charge users per token— the more consumed, the more they sell. At that point, tokens become a commodity that can be mass-produced, tiered for pricing, and traded at scale.

At GTC 2026, Huang first put forward the token economy, saying, “Tokens are the new bulk commodity.” In his description, data centers are like 24/7 token factories, with raw materials being data and electricity, and the product being tokens.

He proposed a new metric, “tokens per watt,” believing it will measure the future revenue capacity of data centers. This is because, “under a fixed power cap, whoever has the highest tokens-per-watt throughput has the lowest production cost.” NVIDIA’s technology iterations have always focused on improving token production efficiency.

In short, the essence of the token economy is to measure, price, and trade AI’s intelligent output just like industrial goods.

The token economy is happening. OpenAI CEO Sam Altman, in a speech earlier this year, said, “At a fundamental level, our business—and every AI model provider’s business—will essentially become the business of selling tokens.”

Liu Liehong, head of China’s National Data Administration, recently said that by this March, China’s daily average token usage had exceeded 140 trillion, up by more than 1,000 times from 100 billion at the beginning of 2024. Compared with 100 trillion at the end of 2025, it further increased by more than 40% in just three months.

Liu believes that tokens are not only a value anchor in the intelligent era, but also a “settlement unit” connecting technology supply with commercial demand—providing quantifiable possibilities for business model implementation.

The “Token Factory” Industrial Chain

“A new industrial revolution is underway: the inputs into the factory (data center) are data and electricity, and the output is tokens.” Huang said.

Just like a manufacturing factory, a “token factory” needs facilities, equipment, logistics, sales, and other links. Based on this logic, and considering research reports from multiple brokerages, the token economy can be broken down into four stages.

#1

Production Stage

Involved segments: AI chips and servers, AIDC (AI data center) infrastructure, liquid-cooling heat dissipation, power supply systems

The process of token production is the same as the inference process: converting electricity and data into tokens. And what determines the upper limit of a data center’s production capacity is its physical hardware, including AIDC rooms, AI chips and servers, liquid-cooling systems, and power infrastructure. Together, they determine power utilization efficiency—how many tokens each watt of electricity can be converted into.

Huang mentioned, “A 1-gigawatt factory will never become 2 gigawatts; that’s a law of physics.” This means that competition in the production stage is fundamentally a fight for efficiency: with the same amount of electricity, whoever can produce more tokens gains more advantages.

#2

Optimization Stage

Involved segments: inference optimization algorithms, scheduling systems, optical modules, etc.

After a data center is built, total power is fixed. With hardware unchanged, the core way to increase revenue is to make each watt of electrical power output more billable tokens.

At GTC 2026, Huang cited an example: Fireworks AI and Lynn. Without replacing any hardware, by relying only on NVIDIA updating the software stack and inference algorithms, their token generation speed increased from about 700 tokens per second to nearly 5,000. This means that technologies such as scheduling algorithms and inference optimization can significantly increase factory output without adding hardware.

#3

Distribution Stage

Involved segments: CDN (content delivery network), cross-border private networks, undersea fiber-optic cables

After tokens are produced, they must be delivered to end users with extremely low latency. Unlike physical goods, token production and delivery often happen at the same time.

The edge nodes of the CDN (content delivery network) take on the “last-mile” delivery role, and when tokens need to be delivered across countries, cross-border private networks and undersea fiber-optic cables form international logistics channels.

“Token going overseas” also occurs in this stage. Domestic models, leveraging their clear inference cost advantage, are massively outputting tokens via overseas API platforms, supporting the network infrastructure that makes cross-border flows possible—forming the fundamental pipeline for going overseas.

#4

Application Stage

Involved segments: large model vendors, agent applications, vertical industry SaaS, multimodal generation platforms

The application stage is also where the token economy ultimately realizes its value. At GTC 2026, Huang predicted that in the future, every SaaS company will become an Agent-as-a-Service company (intelligent agent services), and every engineer will have an annual token budget.

As AI applications continue to roll out, token consumption scenarios will go far beyond today’s conversational AI. They will expand into areas such as intelligent agents, multimodal content generation, and financial analysis. The larger the consumption, the more it in turn pulls demand for capacity expansion in the upstream production stage, creating a virtuous cycle. This also serves as the underlying flywheel that keeps the entire industrial chain running continuously.

Focus on investment directions such as compute infrastructure

In a research report, Great Wall Securities believes that OpenClaw represents a new and strong acceleration point for AI, and the speed at which tokens “burn” will rise sharply. In this kind of model, token consumption can grow by multiples, even by dozens of times.

From an investment perspective, the first beneficiaries of the rapid development of the token economy are the production stage of token factories, including AI chips, data centers, liquid cooling, power supply, and other compute infrastructure—this is also the direction with the highest degree of consensus among institutions today.

A research report from CICC (China International Capital Corporation) shows that ByteDance’s token consumption roughly doubles every three months. When domestic large cloud vendors reach an average daily consumption of 60 trillion tokens, they will face a clear compute capacity shortage. Therefore, it is expected that when domestic large cloud vendors reach an average daily token consumption of 30 trillion, they will feel power tightness, and when they reach 60 trillion, they will start to see a certain compute capacity shortfall.

Jiang Ying, chief analyst of the communications industry at Open Source Securities, believes that tokens = AI chips (domestic compute + compute leasing) = AIDC. A research report by Guojin Securities states that in 2026, the compute industry chain will enter a “full-chain inflation” cycle, and the improvement in business conditions will propagate from chips to AIDC, cloud services, and power equipment.

In addition, compute leasing and token going overseas are also hot areas that benefit from the token economy.

Great Wall Securities believes that the essence of token going overseas is that China’s local AI models provide inference services to the world through API interfaces, charging based on processing volume—thereby enabling a “digital export” of compute and electricity. The core reason why China’s large models can quickly capture global market share is their highly competitive cost control, especially in the power segment.

According to calculations by the Shenwan Hongyuan computer team, the comprehensive inference cost of domestic AI models is only one-sixth to one-tenth of overseas.

“Under the token industrial chain is fundamentally a transformation that turns electricity from the physical world into the intelligence of the digital world,” Great Wall Securities believes. The pricing logic of this industrial chain follows the path: “explosion in overseas demand → shortage of compute-and-storage hardware → bottlenecks in energy/infrastructure → re-estimation of costs across the full chain.” A foundation is built from green electricity and ultra-high-voltage power transmission that have a cost advantage in the upstream, locking in the lower bound of gross margin; the midstream compute and storage layer is the core capacity bottleneck constraining supply; the sub-midstream model and scheduling layer obtain technological value-added through algorithmic optimization; and the downstream applications and overseas layer open up an upper limit on profits by leveraging a globally high willingness to pay.

Great Wall Securities believes that from an investment perspective, the prioritization of attention can be divided into several stages. The first stage is storage and GPU memory, capturing the maximum pricing upside from short-term supply-demand mismatches; the second stage is compute chips and servers, locking in mid-term performance; the third stage is power equipment and green electricity operations, which have long-term barriers; the fourth stage is leading companies that have real-world scenario deployment capabilities and the ability to monetize overseas at high premiums.

（Source: China Securities Journal）