AI News

Ai Analysis releases China Token Factory Competitiveness Ranking: 28 vendors divided into four quadrants, with Volcano Engine, Alibaba, Tencent and Baidu in the Leaders quadrant

Smart City Industry AnalysisUpdated Nov 16 21:22

AI summary

Ai Analysis has released the China Token Factory Competitiveness Ranking. The 28 vendors fall into four quadrants based on production capacity and commercialization capability. The four major cloud providers — Volcano Engine, Alibaba, Tencent and Baidu — occupy the best positions in the Leaders quadrant, joined there by leading startups such as Shishi Technology, Paratera, Infinigence AI, SiliconFlow and Jiuzhang Cloud. China's average daily Token call volume climbed from 140 trillion in March to 500 trillion in June, and is now estimated to have exceeded 1,000 trillion. The report also flags two major risks: declining prices of computing power assets and low gross margins.

Body · AI translation

English text is being prepared.

As the supply side, token factories (hereinafter referred to as Token factories) have become a hot track that governments and industries are racing to enter. Ulanqab, the "token capital," hosted an inspection by the General Secretary, setting the tone for the long-term development of Token factories; Token factory construction is in full swing across the country; and nearly 10 Token factories, including Jiuzhang Yunji, Infinigence AI, and SiliconFlow, are competing for the title of first IPO.

So how strong are the Token factories really? What hidden concerns lie beneath an industry that looks so promising? To this end, Aifenxi has produced a competitiveness ranking of China's Token factories, to clarify the capacity, efficiency, cost, and commercialization capabilities of different factories.

In this report, a Token factory is defined as an operating entity that independently organizes AI model inference production, provides customers with inference services billed in token units, and is responsible for delivery quality. A Token factory needs to support multiple mainstream models in the industry; entities that provide inference services for a single model, such as DeepSeek and Zhipu, are not included in the evaluation scope.

The main body of this report is divided into three parts. The first part presents the ranking and analyzes the strengths and weaknesses of each factory; the second part reveals hidden risks in the industry; and the third part explains the ranking methodology. 1. Ranking

The token market has exploded, attracting players from all sides.

Cloud giants regard tokens as their most core revenue growth engine; Volcano Engine and Alibaba are fighting over who ranks first in market share, bidding fiercely for airport advertising space. Mid- and long-tail cloud vendors, who could never achieve profitability in the CPU cloud era, see transforming into NeoClouds selling tokens as their path to self-redemption.

The three major telecom operators, after bloodying the AIDC market with low prices, are also eyeing the larger Token factory market. IT vendors find the AI transformation painful, and building intelligent computing and selling tokens is instead an easier road to rebirth.

The Institute of High Performance Computing Technology at Tsinghua University has become an incubator for Token factories; successive directors, deputy directors, and graduating students have founded batch after batch of Token factory startups. The most famous is Academician Zheng Weimin, the second director, who also serves as an advisor to several companies such as Qujing Technology and Qingcheng Jizhi.

This Token factory ranking includes 28 vendors in total, covering typical representatives of all the above groups. Aifenxi scores the shortlisted vendors based on the production capacity and commercialization capability of Token factories, placing them into four quadrants.

Figure 1: 2026 China Token Factory Competitiveness Ranking

1.1 Leader quadrant

Cloud giants and leading startups, 9 vendors in total, gather in the leader quadrant. The four cloud giants Volcano Engine, Alibaba, Tencent, and Baidu occupy the best positions in the quadrant, while leading startups such as Shishi Technology, Paratera, Infinigence AI, SiliconFlow, and Jiuzhang Yunji have also entered the leader quadrant.

The advantages of cloud giants are fully demonstrated at the production capacity level; their token production efficiency and cost are in an absolutely leading position. First, cloud giants have extremely strong self-developed inference optimization capabilities. Because they own self-developed models, their inference optimization is white-box optimization at the model source code level, and the optimization methods they accumulate are more effective; Token factories without self-developed models mostly do black-box optimization, and the iteration efficiency and speed of optimization strategies are limited.

Second, cloud giants can push inference costs to the extreme. On the one hand, cloud giants have a large number of internal businesses and overseas businesses, and GPU utilization remains at peak levels day and night; on the other hand, cloud giants have strong financial strength and low costs, and can lock in GPU prices through long-term agreements, smoothing out the cost increases brought by price rises.

Leading startups, meanwhile, make it into the leader quadrant mostly by relying on past technological accumulation. Paratera and Shishi Technology have a foundation in CPU supercomputing acceleration optimization; Professor Wang Yu of Infinigence AI comes from a background in neural network acceleration at the chip level; and SiliconFlow's technical team originates from operator optimization and acceleration.

Moreover, these five startups all belong to the Tsinghua circle and have considerable ties among them.

Wang Yu, chairman of Infinigence AI, holds bachelor's, master's, and doctoral degrees from Tsinghua's Department of Electronic Engineering. After receiving his doctorate, he stayed on to teach and eventually became department chair, during which time he founded Infinigence AI; co-founder Xia Lixue is also Wang Yu's student.

Fang Lei, chairman of Jiuzhang Yunji, is Wang Yu's senior fellow in the same department and graduated with a bachelor's degree from Tsinghua's Department of Electronic Engineering. Jiuzhang Yunji was founded back in 2013; its first 10 years were spent building an AI platform, and in its second 10 years it transformed into a Token factory.

Chen Jian, chairman of Paratera, holds a doctorate from Tsinghua's School of Aerospace Engineering and was among the first experimental students in Tsinghua's combined bachelor's-master's-doctoral program. After graduation, he worked as a high-performance computing architect at Intel China, and from there founded Paratera in the supercomputing field.

Yan Bowen, chairman of Shishi Technology, is Chen Jian's junior fellow in the same department, with a doctorate from Tsinghua's School of Aerospace Engineering and postdoctoral work in the Department of Computer Science. During his doctoral studies, Yan Bowen participated in the development of the domestic supercomputer "Sunway TaihuLight," and later led a team to incubate Shishi Technology from the National Supercomputing Center in Wuxi.

Yuan Jinhui, chairman of SiliconFlow, received his doctorate from Tsinghua's Department of Computer Science and is Yan Bowen's senior fellow in the same department. SiliconFlow co-founder Hu Jian also holds a doctorate from the Department of Computer Science.

These Tsinghua-circle startups have outstanding technical capabilities, but their commercialization capability is still at an early stage; in 2026, the highest expected revenue for a single company is just over 2 billion yuan.

Aifenxi believes that the leading startups' entry into the first tier is not short-term luck, but reflects long-term structural opportunities.

In the short term, Token factories are still in the land-grabbing phase, with market demand exceeding supply; the growth of both major cloud providers and startups is extremely rapid, and head-on competition has not yet begun. In the second quarter of 2026, Alibaba Cloud's single-quarter revenue grew 45%, a new high in nearly six years. Startups grew even faster: Paratera's revenue growth in the first three quarters exceeded 110%, and profit growth exceeded 8 times.

In the long term, major cloud providers generally have self-developed models and are competitors of independent model vendors such as Zhipu, Kimi, and DeepSeek. These independent model vendors will certainly cultivate a group of Token factories over which they have full control; for example, SiliconFlow and Infinigence AI both have Zhipu behind them. With the rapid growth of independent model vendors, opportunities for Token factory startups will exist over the long term.

Among the major cloud providers, Baidu Token Factory's progress is not ideal. Aifenxi believes that rather than persisting with the unimpressive ERNIE Bot, Baidu would be better off completely abandoning self-developed models, transforming into a neutral cloud provider, and focusing on Token factories.

In this regard, AWS has already set a benchmark. Recently, AWS shut down its AGI lab, abandoned its self-developed model Nova, and focused on Token infrastructure. Without the constraints of a self-developed model, AWS's cooperation with major model vendors has become smoother, and its revenue growth in the third quarter hit a new high, exceeding 39%. Anthropic also signed a ten-year long-term agreement with AWS, committing to a minimum consumption of US$110 billion.

In building Token factories, Baidu and AWS also have similar resource endowments. Baidu has Kunlun chips, and AWS has Trainium; in the current period of tight computing power supply, this is a core manifestation of competitiveness for Token factories.

1.2 Practitioner Quadrant

The Practitioner quadrant contains 6 vendors in total, mainly telecom operators and NeoClouds. The three major operators China Telecom, China Mobile, and China Unicom are all included, while PPIO, Wanjie, and Kingsoft Cloud are included as NeoCloud representatives.

Vendors in this quadrant have obvious strengths and weaknesses. Their strength lies in strong resource capabilities, while their weakness is the lack of self-developed technology capabilities.

The resource capabilities of the three major operators go without saying. Holding distribution channel resources for government and enterprise customers as well as consumers, all three major operators have launched their own monthly Token Plan packages, covering personal and enterprise editions respectively. At the same time, the operators also control the shipment resources of domestic computing power chips. Domestic computing power chips mainly go to governments, operators, and major cloud providers, and the government's procurement lever is often also the operators; therefore, operators frequently disclose tens-of-billions-level GPU centralized procurement. In early November, China Mobile had just disclosed a 11.6 billion yuan centralized procurement of 20,000 GPU cards.

NeoClouds' advantage lies in possessing mature computing center resources and customer resources. Wanjie Ark is backed by Wanjie Data, owns more than 60,000 GPU cards, and has more than 16 intelligent computing centers under construction or completed. Wanjie Data has operated for 12 years and accumulated a large amount of customer resources that can be directly converted into target customers for Token factories.

The shortcomings of the three major operators and NeoClouds all lie on the inference capability side. As resource-driven vendors, they all have obvious misalignment in self-developed technology capabilities. The Token factories built by the three major operators in various regions are all subcontracted to different service providers and basically do not touch inference optimization technology. The inference optimization technology capabilities of NeoClouds are still in the early stages of development and are basically on par with open-source levels such as vLLM and SGLang.

1.3 Innovator Quadrant

The Innovator quadrant mostly consists of vendors that are wary and skeptical about touching computing-power-heavy assets, including QingCloud, UCloud, Tsinghua Leap Intelligent, Qingcheng Jizhi, and Gongji Technology.

These vendors are not weak in inference technology and each has its own strengths. QingCloud and UCloud came from the CPU cloud background and are deeply versed in heterogeneous adaptation technology for one cloud with multiple cores; Tsinghua Leap Intelligent's self-developed inference engine emphasizes automatically completing model adaptation and operator optimization, and is known for speed amid rapid model iteration; the Qingcheng Jizhi team comes from the Tsinghua University High Performance Computing Institute mentioned at the beginning, with current director Professor Zhai Jidong serving as chief scientist; Gongji Technology has unique advantages in the fault-tolerance layer of computing power scheduling.

In a period when computing power resources are scarce, without computing power there is no Token production capacity. However, most of these vendors operate with asset-light models, are hesitant about holding computing power themselves, and their financing progress has not been smooth. Listed QingCloud has never released a private placement plan; UCloud originally planned to raise 1.5 billion yuan but terminated the plan in September; other unlisted vendors have financing amounts in the tens of millions, currently only enough to buy 1-2 NVIDIA DGX B300s.

In this quadrant, only Qujing Technology has completed large-scale financing and is laying out regional infrastructure construction. The company focuses on regions such as Henan and Hainan and is building Token factories together with local state-owned capital.

However, cooperation with these provinces carries significant hidden risks: state-owned capital will impose local production and local consumption constraints on Token factories. On the production side, most of the models needed by local industries are not open-source flagship models, and whether Token factories can be built and operated locally according to local conditions is highly uncertain; on the consumption side, the Token consumption capacity of local industries remains to be verified, and once overcapacity occurs, the remaining Token production capacity will also find it difficult to gain a cost advantage in cross-provincial output.

1.4 Explorer Quadrant

The Explorer quadrant includes 7 vendors in total: BONC, iSoftStone, Teamsun, Intellifusion, Qiniu Cloud, Jiliu Technology, and Moxing Intelligence. These vendors have just emerged in the Token factory field, and most have transformed from other fields, especially traditional IT.

iSoftStone is struggling painfully in its original IT domain. Over the past five years, revenue climbed steadily, surpassing 35 billion yuan in 2025, but total profit plunged in a straight line from over 1 billion yuan in 2021 to 5.58 million yuan in 2025, with the profit margin approaching zero.

Transforming to AI is the only way out. In 2026, iSoftStone began building Token factories across various locations under the brand of Factory No. 1. The first Token factory landed in Yizhuang, and the next step is to expand to computing nodes such as Zhangjiakou, Ulanqab, and Chengdu, forming a nationwide Token production network.

Vendors in the Explorer quadrant need more time to prove themselves on both the production technology side and the commercialization side. 2. Industry risks

Token factories achieved rapid development in 2026, and as the year-end approaches, all practitioners surveyed by iResearch are extremely optimistic about development in 2027.

However, all this optimism rests on one basic condition: computing power prices will continue to rise. Once computing power prices turn downward, Token factories will be the first to blow up.

2.1 Asset price risk

In 2026, Token factories with self-owned computing power fully enjoyed the gains brought by rising prices.

The price of NVIDIA DGX B300 rose from 4 million at the end of 2025 to 7 million in April 2026, 9 million in June, 14 million in August, and then 20 million in early November. In less than a year, the price rose 4 times cumulatively.

During the survey, the founder of a Token factory said that the N cards the company purchased over the past few years had appreciated to several billion yuan, and by selling old cards at high prices, the company obtained one-time asset disposal gains exceeding its full-year revenue.

However, once asset prices fluctuate downward in 2027, the risk will quickly transmit along multiple links.

First, the asset side will first recognize impairment, causing profit performance to decline sharply. The GPU cards held by Token factories are generally recorded as fixed assets, and when market prices fall, impairment must be provisioned, thereby significantly affecting current-period profit performance.

Second, the loan default risk of Token factories rises sharply. Because Token factories generally pledge GPU cards to banks for financing loans, after prices fall, the appraised value of collateral declines, and banks will cut credit lines, require additional collateral, or demand early repayment of loans, which will deal a severe blow to the cash flow of Token factories, with default risk running high.

If it is a Token factory jointly built with the government, because of government endorsement its financing leverage can be as high as 10 times, and it will be the hardest-hit area for blowups.

Third, the computing power locked in advance by long-term leases of Token factories will lead to years of operating losses. To reduce inference costs, the GPU computing power leased by Token factories is locked into 3-5 year long-term leases, and falling computing power prices will accelerate the decline in Token unit prices; by then Token factories will have to withstand the double blow of high procurement costs and low selling prices.

Fourth, the asset disposal capability of Token factories drops to zero, and the profit safety cushion that provides a floor becomes ineffective. In an asset upcycle, even if operations are poor, Token factories can recover funds by disposing of old cards to ensure a certain project profit; in an asset downcycle, the price at which Token factories dispose of old cards falls even faster, and the funds that can be recovered are unlikely to make up for operating losses.

According to the iResearch survey, the industry is not only hoarding H-series and B-series large cards, but also 50-series small cards; domestic cards are also firmly held in the hands of the government and operators. When everyone is hoarding cards, the risk of price declines will be infinitely magnified, and the computing power assets purchased at high prices in 2026 constitute the main risk for the next round of competition.

2.2 Gross margin risk

iResearch previously conducted a comparative study of overseas and Chinese Token factories, and carried out systematic survey analysis on differences in pricing power (click to view: 2026 iResearch · Token Market Overseas Benchmarking Research Report).

The largest customer of overseas Token factories is end enterprises, while the top customers of Chinese Token factories are basically locked in as model vendors.

The underlying business logic of Token factories is the cloud, so overseas they are more often called NeoCloud. In the long run, 90% of cloud vendors' revenue is contributed by their top 10 customers. Applied to Token factories, 90% of revenue will be contributed by model vendors.

Model vendors make centralized purchases from Token factories, then sell to consumers and enterprise customers. Token factories are responsible for organizing production and bearing the costs of computing hardware, scheduling, and inference optimization; if they cannot control end selling prices, gross margins will be confined to an extremely low space.

As centralized Token buyers, model vendors generally purchase Token capacity at 30-50% of the list price, then wholesale or retail it at 70-90%. The production cost of Token factories is often about 40% of the list price, so gross margins are extremely low, and inversion even occurs.

Recently, DeepSeek fully open-sourced its internal inference framework DeepInfer, which will become another kill line for the gross margins of Token factories.

English text is being prepared.

Facing the long-term risk of low gross margins, Token factories need to strengthen their pricing power, or seek end-customer enterprises with weaker bargaining power, or continue to maintain an asset-light model and give up high revenue growth. 3. List Description

3.1 Definition of Token Factory

Ai Analysis defines a Token factory as an operating entity that independently organizes AI model inference production, provides customers with inference services billed in Token units, and is responsible for delivery quality.

According to the definition, whether an entity can be shortlisted for the competitiveness list is mainly based on three criteria.

First, production autonomy. A Token factory supports inference services for multiple mainstream industry models; pure API forwarding is not included.

Second, billing unit. External delivery and settlement are measured in Tokens; entities that charge by card-hour, rack, etc. are not included.

Third, attribution of responsibility. The entity bears responsibility for availability, latency, fault handling, and compensation; channel parties that only handle sales are not included.

3.2 List Evaluation System

The list sets two major evaluation indicators: production capability and commercialization capability. Production capability covers four sub-items: inference optimization, compute scheduling and utilization, chip and model adaptation, and service stability. Commercialization capability covers four sub-items: revenue scale, compute supply chain, gross margin, and customer dispersion.

Each item uses a five-tier evaluation and is weighted and summarized into scores for the two major evaluation indicators.

For inference optimization capability, Ai Analysis mainly examines whether the entity has an independent inference stack, whether it can push optimization down to the model structure and operator layers, and whether it has publicly verifiable quantitative results.

For compute scheduling and utilization, Ai Analysis mainly examines long-term utilization, whether there is a cross-domain and cross-task peak-shaving and valley-filling mechanism, and whether the resource pool is underpinned by internal business.

For chip adaptation, Ai Analysis mainly examines the breadth and cycle of adaptation to domestic chips. For model adaptation, Ai Analysis mainly examines how soon after a new model is released it can provide external services.

For service stability, Ai Analysis mainly examines whether availability, first-token latency, and success rate meet public baselines.

For revenue scale, Ai Analysis mainly examines verifiable business-caliber revenue; consolidated group revenue must be split before use.

For the compute supply chain, Ai Analysis mainly examines the structure of self-owned and externally rented resources, who bears impairment when prices fall, and whether resources can be disposed of after amortization is completed.

For gross margin, Ai Analysis mainly examines gross margin under the Token service caliber and does not use consolidated group gross margin as a substitute.

For customer dispersion, Ai Analysis mainly examines customer tiers and concentration; those with a single customer or channel accounting for more than half are downgraded.

3.3 Data Sources

The data for this list comes from three types of sources: officially disclosed information, authoritative industry sources, and Ai Analysis first-hand research. Data production adheres to cross-validation. For the same vendor's data and information, we strive to corroborate them across multiple sources. Report Download

In the official account "Smart City Industry Analysis", reply with the keyword "Ai Analysis Report" to obtain the relevant report files. www.smartcity.team

Article tags: token factory token factory

Statement: Please indicate the source as "Smart City Industry Analysis (www.smartcity.team)" and the relevant author when reprinting. Except for original content, some information on this site comes from the Internet. If the relevant content involves infringement of rights, please contact the contact information at the bottom to delete it.

[Digital Technology Special Research] Community

In-depth peer exchange, focused on smart city and digital technology industry research and material accumulation.

Continuously update the latest industry white papers, industry research, policy developments, and implementation solutions to help industry practitioners efficiently accumulate knowledge and share experience.

Massive premium content is continuously iterated, providing frontier perspectives, connecting industry talent, carrying out case co-creation and resource exchange, and jointly building a high-quality industry exchange atmosphere.

#post-##

Recommended for you

Shanghai Data Research Institute's "National Data Group Development Blue Book (2026)": Four turns of data groups

Summary of achievements released at the 2026 Artificial Intelligence Industry Conference (with download): The Artificial Intelligence Industry Development Alliance AIIA centrally released 8 report achievements for download

2026 Embodied Intelligence Industry Innovation Report: Zhiyuan Innovation Day, jointly released by Xinhua News Agency and a subsidiary of CAICT, defining industry stages with the "XYZ curve" of embodied intelligence industry development, driving technological evolution with "three intelligences in one," and measuring productivity value with Worken

Jiazi Guangnian Think Tank's "2026 Agent Commercialization and Competition Path Research Report": Analysis from five aspects: product and demand evolution, formation of commercial value, market ecosystem and enterprise deployment, application landscape and commercialization practice, and industry evolution and competition outlook

"Embodied Intelligence Development Report (2026)": Embodied intelligence is at a transitional node toward pilot applications; the industry reveals coldness amid continued heating; the technical system composed of four elements—models, data, bodies, and compute—is evolving collaboratively; industry chain division of labor is refining, but a sustainable commercial closed loop still needs to be established

National Public Data Operation Development Report (2025-2026): There are 76 public data resource authorized operation platforms or zones that have been publicly operated or are under construction nationwide, with 202 implementation institutions and 224 operating institutions determined

Source

Smart City Industry Analysis