Inference spending to surpass training for the first time in 2026, rise of Token factories to change AI infrastructure investment logic

Science and Technology Innovation Board Daily, October 4 (Reporter Wang Nai) In 2026, the investment logic for AI infrastructure is undergoing a fundamental turn.

Consider a recent set of data. The latest Gartner forecast shows that as agentic AI raises computing intensity, the market's center of gravity is shifting from training to inference. Global inference spending is expected to reach $23.3 billion in 2026, surpassing training spending of $19 billion for the first time and accounting for 55% of total AI IaaS spending; that share will rise to 59% in 2027.

The "Token Economy Development Research Report (2026)" released by the China Academy of Information and Communications Technology (CAICT) forecasts that in 2026, inference computing power will account for 70.5% of the total AI computing load, as AI computing demand shifts further from model training to inference and agent operation. As of the end of June 2026, China's total intelligent computing power reached 2,185 EFLOPS (FP16); in 2025, China built a cumulative total of 42 10,000-card-level intelligent computing clusters.

Where is the real explosion of inference demand?

Competition among large-model companies over model capabilities and parameters—which major company bought chips, which model vendor released a new-generation model—has long dominated industry discussion. But spending money on chips and piling up computing power does not directly translate into realized commercial value. Even well-funded major companies, facing a computing business that burns hundreds of millions of dollars a month, must define their strategic approach clearly: cut where you can, spend where you must. Today, the explosion of inference demand is not an abstract concept; it is already quantifiable in frontline customer scenarios.

In July this year, the relevant business and some teams of Tencent's QClaw product center were moved to the department of the AI office agent product WorkBuddy; in August this year, Doubao released an Agent product and brand for productivity scenarios—Doubao Work; in September this year, Alibaba's Qianwen Office released a series of enterprise-level AI Agent products at the Yunqi Conference, and Qianwen Office simultaneously released the AI Agent portable assistant QwenNote A2.

GoodVision AI CEO David Wang believes that 2025 to 2026 is the market's watershed. "In the past, enterprises' AI efforts were mainly experiments and POCs (proofs of concept). In 2026, a large number of businesses will enter regular operation, and programming, biological data processing, game AI characters, and video generation will become high-consumption scenarios." The business of selling Tokens has already achieved a commercial closed loop in some scenarios.

In the interview, he also mentioned a biopharmaceutical customer case: previously 1 to 2 people spent 7 days processing experimental data; after introducing an Agent, this was compressed to 10 minutes, with Token consumption of only single-digit dollars. However, the number of Tokens does not equal business value and commercial revenue. David Wang said bluntly: "What customers buy is not the number of Tokens, but the results of completing business tasks." By reducing invalid calls and retries through routing optimization, Token consumption falls while task value actually rises.

How is the value of Tokens calculated? According to the "Token Economy Development Research Report (2026)" released by CAICT, the value density of Tokens varies enormously across application scenarios. Consumer-level scenarios are traffic-consumption type with low value density—typically chitchat, entertainment Q&A, and the like, which require only basic interaction capabilities and cost about $0.01 per million Tokens. Industry-level scenarios have higher value density—typically drug R&D, financial quant, precision code generation, and the like, where value can reach $1,000 per million Tokens, making them the high-value ground of the Token economy.

Token going global ≠ models going global: what still stands between traffic and revenue?

Based on the latest OpenRouter data, total global AI large-model calls from September 21 to September 27 were 146 trillion Tokens, up 13.18% month on month. The weekly call volume of Chinese large models has exceeded that of the United States for 22 consecutive weeks, firmly ranking first globally.

However, several industry insiders interviewed said that the market commonly conflates Hugging Face weight downloads and OpenRouter Token call volume as overseas performance, but the commercial value of the two is worlds apart.

Gou Yurui, chief media analyst at Southwest Securities, said, "Token call volume does not equal the number of requests, the number of users, or revenue. What OpenRouter counts is only traffic routed through its API, excluding model vendors' own APIs, other cloud marketplaces, enterprise private deployments, and traffic from developers running locally after downloading weights." Tokens are first of all a workload indicator, not a financial indicator."

"Hugging Face's high download volume is more like 'brand exposure' than 'real revenue.' It shows that the global attention on Chinese models is rising rapidly, but from attention to real money paid for Tokens, there is still a long conversion chain in between," said Xie Jin, an AI product expert at PPIO.

Xie Jin also distinguished between models going global and Tokens going global. He believes that models going global is essentially open-source weights going global. Its core value is building technical influence and developer awareness, similar to making the "recipe" public. But model weights themselves do not directly generate revenue. Tokens going global is essentially inference services going global: overseas users remotely call our deployed model inference services through APIs and pay according to the actual number of Tokens consumed.

"When Tokens go global, what is sold is not the model itself, but 'model-as-a-service,' a complete set of service capabilities including compute, inference optimization, stability assurance, and compliance architecture. This is the path that truly generates commercial revenue," Xie Jin said. It is understood that PPIO currently has more than 680,000 registered developers worldwide, with average daily Token calls reaching 2.2 trillion.

"Chinese AI going global will not be limited by a single factor. Technology, cost, compliance, and monetization models affect each other," said Liu Hua, deputy general manager of UCloud's Emerging Business Division. At the technical level, he believes, the capabilities of Chinese models continue to improve and already possess strong international competitiveness. Future priorities will focus on multilingual capabilities, stability during long-running operation, tool-calling capabilities, inference speed, and controllability in complex business scenarios.

Tokens going global is not the same as models going global, Liu Hua continued. He believes that models going global mainly refers to the global spread of model weights, technical architecture, and the developer ecosystem. Overseas developers can download weights and deploy, fine-tune, and run inference themselves. Tokens going global, by contrast, delivers model inference capabilities as an online service.

Compared with models going global, Tokens going global has a more direct relationship with actual revenue, but it also requires continuously bearing GPU, bandwidth, platform operations and maintenance, channel revenue-sharing, and compliance costs.

The rise of Token factories: who is making money?

In the inference era, a group of "Token factories" is rising. Their common logic is: not just selling compute, but turning bare metal into deliverable, billable model inference capabilities. Who is laying out this field? From the perspective of player types, they can roughly be divided into the following categories.

Public cloud giants are natural heavyweight players in this track. Volcano Engine Ark, Alibaba Cloud, and Baidu AI Cloud, relying on cloud ecosystems and infrastructure scale, hold first-mover and scale advantages in the Token factory track.

According to IDC data, in 2025 the call volume of large models on China's public clouds reached 1,944 trillion Tokens, a 16-fold year-on-year increase. Among them, Volcano Engine ranked first in China with a 49.5% market share, while second- and third-ranked Alibaba Cloud and Baidu AI Cloud took 28% and 10% market shares respectively. Together, the three companies accounted for nearly 90% of the share.

Also heavily bet on by capital are independent MaaS inference service providers. Companies such as SiliconFlow, Fireworks AI, Infinigence AI, and Together AI are trying to tear open a gap in a track crowded with giants by using self-developed inference engines and compute scheduling capabilities.

Recently, industry-related financing has also been frequent.

For example, on September 20, AI infrastructure company SiliconFlow announced that it had completed the second phase of its B+ round and its C round, with cumulative equity financing in 2026 approaching 2.9 billion yuan. There has also been recent news that Fireworks AI, the inference service provider ranked first by revenue scale, is also considering a new round of financing. If the financing proceeds, the target valuation could reach $30 billion, nearly double the previous round's valuation announced in July.

In addition to these newcomers, manufacturers that once held GPU chip resources are also striding into the Token industry, and building Token factories has become a hot trend in the industry. For example, UCloud has proposed building "Token factories" for the AI era and has deployed 36 availability zones in 28 regions worldwide, covering markets in Southeast Asia, Europe, North America, South America, and Africa. According to the company, its Pakistan, Kazakhstan, and Uzbekistan nodes recently went online, further entering the Central Asian market.

In addition, Runjian Co., Ltd. said in a recent investor relations activity that the company has launched the "Xingsuan Cloud Pool" product, and the Wuxiang Yungu Intelligent Computing Center will further develop into a Token factory. Tianyima issued an announcement on the record of investor relations activities, stating that the first phase of the company's Tianyima No. 1 International Token Factory plans to deploy 2,000P of computing power, and some equipment has already been put into production and operation.

In addition, the three major telecom operators are also collectively turning to "Token operations." For example, China Telecom's Tianyi Cloud launched the Xingchen TokenHub unified service platform; China Mobile's MoMA multi-model service engine was upgraded to 2.0; and the "Unicom Xingluo" computing power scheduling platform took the lead in achieving 2,000 km cross-regional mixed training across four locations and three chips.

Bottlenecks and endgame: short term depends on compliance, medium and long term on monetization

Xie Jin, an AI product expert at PPIO, believes that Token going global currently has two biggest bottlenecks. Xie Jin believes that compliance is currently the hardest threshold. The EU GDPR, Middle East data localization policies, and various countries' rules on cross-border data transmission all require user data to remain local. Once violated, the maximum fine can reach 4% of global revenue."

In addition, global low-latency distribution is another major bottleneck. "AI applications, especially consumer-facing products, have extremely high requirements for response speed, with end-to-end latency needing to be controlled within 5 seconds. But when calling APIs across continents, one-way network latency alone may exceed 200 milliseconds, and adding model inference time, the user experience will noticeably decline." In response, PPIO adopts a multi-region localized deployment solution while providing absolute tenant-level data isolation and PrivateLink dedicated line access.

Liu Hua, deputy general manager of UCloud's Emerging Business Division, listed five directions worth attention from a medium- and long-term perspective: model APIs and unified inference platforms, inference infrastructure and cost optimization, AI services for key industries, global networks and localized deployment, and security and compliance services. At the cost level, China has advantages in electricity costs, infrastructure construction, and inference engineering optimization. But Token services also need to bear costs such as GPU depreciation, network bandwidth, overseas node operations and maintenance, security protection, channel revenue-sharing, and customer service.

He particularly emphasized that general model capabilities need to be further integrated with industry processes such as e-commerce, gaming, finance, manufacturing, and security "in order to form sustained payments."

But what truly determines how far Tokens going global can go is the monetization model itself. "Currently, the main source of revenue for Tokens going global is still API calls. This is a 'selling water, electricity, and coal' model, with unit prices continuing to fall and revenue growth relying entirely on exponential growth in call volume. If call volume growth slows, or price wars intensify, the profit space of the entire business model will be severely squeezed," Xie Jin said.

STAR Market Daily reporter Huang Xinyi 2026-10-05 21:29

STAR Market Daily reporter Li Jiayi 2026-10-05 11:55

Collapse; the SSE Innovative Drug STAR Leading Index and the SSE STAR Market Software Services Index were released on the 29th