Deep Analysis of AI Large Model Unicorn IPO Prospectus: MiniMax to C, Zhipu to B

Deep Analysis of AI Large Model Unicorn IPO Prospectus: MiniMax to C, Zhipu to B

Within a 48-hour window, two “unicorns” in China’s large model sector successively submitted IPO prospectuses to the Hong Kong Stock Exchange.

On December 19th and December 21st, Beijing Zhipu Huazhang Technology Co., Ltd. (“Zhipu”) and Shanghai Xiyu Technology Co., Ltd. (“MiniMax”) each disclosed their Hong Kong prospectuses.

These nearly thousand-page documents, with unprecedented granularity, not only reveal the true accounts of China’s AI large model enterprises, but also highlight the fierce clash between two distinctly different commercialization paths: ToC and ToB. Under the spotlight of the capital market, although both companies are on the same large model track, they represent completely different worlds in terms of technical background, revenue structure, customer profile, and cost control.

Business Model Endgame: ToC VS ToB

MiniMax: Consumer-driven, betting on "super applications"

MiniMax’s business logic is closer to an internet product company. The company clearly upholds the banner of “AI Native Products” with core offerings including the all-modality interactive platform Talkie (overseas version) / Xingye (domestic version).

MiniMax’s prospectus shows that “AI Native Products” are its absolute pillar. In 2023, this sector’s revenue was only $758,000, but soared to $21.8 million by 2024, and reached $38.02 million in the first nine months of 2025, accounting for 71.1% of total revenue.

This growth is driven by an explosive increase in users. MiniMax’s products like Talkie/Xingye saw average monthly active users (MAU) grow from 3.1 million in 2023 to 27.6 million by September 2025. More importantly, the number of paying users reached 1.7716 million in the first nine months of 2025, and the average revenue per paying user (ARPPU) increased from $6 in 2023 to $15. This indicates that MiniMax has successfully achieved a closed loop for consumer products: user acquisition-retention-payment.

  • User Data: As of September 30, 2025, its AI native products have served over 200 million users cumulatively, with average MAU reaching 27.64 million.
  • Monetization Model: Mainly through subscriptions, virtual goods purchases, and online marketing services. Talkie/Xingye’s deep interaction between users and virtual characters creates strong user stickiness.

It is worth noting that MiniMax’s “specialization” is not limited to consumer apps; its audio modality B-end output has become a hidden giant.

The prospectus reveals its Speech-02 voice model is recognized as among the world’s best. Data shows that in 2023, its largest client (contributing 37.2% of revenue) is described as a “Shanghai-based digital reading and IP development giant” (e.g., China Literature, etc.). This means a large number of audiobooks, podcasts, and content platform AI reading/dubbing functions in the market are powered by MiniMax. This “consumer products earn traffic, B-end audio earns tech premium” model explains why its open platform maintains nearly 70% ultra-high gross margin.

Zhipu AI: Enterprise-oriented, focused on “local deployment”

Zhipu AI has taken a typical enterprise-level service path, acting more as a technology enabler.

As China’s first company to release a trillion-parameter large model GLM-130B, Zhipu AI has built its business map around its MaaS (Model as a Service) platform.

As of June 30, 2025, its localized deployment (private deployment) revenue reached RMB 162 million, accounting for 84.8% of total revenue.

This data reflects Zhipu’s current client structure: primarily serving national enterprises and large institutions with rigid demands for data security and private deployments. As of June 30, 2025, Zhipu has served more than 8,000 institutional clients spanning technology, finance, public service, and more. Its GLM series models have surpassed 45 million downloads in the open-source community.

Technology Foundation: Multimodal vs Pretrained Frameworks

In terms of technical approaches, both ultimately aim at AGI, but their priorities differ.

  • MiniMax: Emphasizes multimodal capabilities. The prospectus shows its model matrix covers text (MiniMax M series), speech (Speech-02), video (Hailuo-02), and music. Its video generation model Hailuo-02 ranks among the world’s top models on independent benchmarks. The company also highlights its proprietary MoE (Mixture of Experts) architecture and linear attention mechanisms, designed to lower inference costs, which is crucial for consumer-scale applications.
  • Zhipu AI: Emphasizes academic depth and GLM framework. Originating from Tsinghua University's Knowledge Engineering Lab (KEG), Zhipu AI began developing the GLM pretrained framework as early as 2020. Its technical highlights include the GLM-4.5 base model and AutoGLM agent, which can simulate human operations on phones and computers, achieving a leap from “conversation” to “execution.” Zhipu AI frequently emphasizes “safety” and “control” in the prospectus, aligning with its focus on large enterprise and public sector clients.

Financial Breakdown: The Crucial Test—Gross Margin

Gross margin is a key metric for assessing an AI company’s technical premium. In this disclosure, MiniMax’s API business performance surprised the market, while Zhipu AI’s cloud-based business showed weakness.

MiniMax: Rapid consumer monetization, gross margin turns positive

MiniMax’s gross margin underwent a rollercoaster-style recovery: -24.7% in 2023, turned positive at 12.2% in 2024, and further increased to 23.3% in the first nine months of 2025.

Most noteworthy is its B-end business—“open platform and other enterprise services”. While its revenue share declined, in the first nine months of 2025 the gross margin reached 69.4%. This far exceeds the industry average, indicating MiniMax has achieved significant efficiency in model inference cost control, API pricing, or technical architecture (such as MoE, linear attention), enabling it to support commercial services at lower computing costs.

The prospectus shows its open platform serves “digital reading” and “internet service” giants, meaning MiniMax is not just building super apps but is also the “AI audio creator” behind content giants like Himalaya and China Literature.

Meanwhile, according to MiniMax’s prospectus, company revenue is growing quickly: $34.60 million in 2023, and $53.44 million in the first three quarters of 2025 (as of September 30), a sharp annual increase.

Zhipu AI: Stable B-end growth, gross margin at 50%

The prospectus shows that from 2022 to 2024, Zhipu AI's income was RMB 57.41 million, RMB 125 million, and RMB 312 million, respectively—a compound annual growth rate over 130%. In the six months ending June 30, 2025, Zhipu AI’s revenue was RMB 191 million.

Zhipu AI’s overall gross margin remains at a high level of 50%, mainly due to its localized deployment business's high gross margin of 59.1% (2025 H1 data). Local deployment typically includes software licensing and technical services, offering high premium potential.

However, cloud deployment (MaaS) business has become a red flag. The prospectus shows the gross margin for this segment dropped from 76.1% in 2022 to -0.4% in the first half of 2025, resulting in gross margin loss. Zhipu admits in the prospectus this is due to “adjusting service prices to market trends.” This reflects the fierce price war in China’s large model API market, and the pure MaaS model is currently difficult to profit from.

R&D Investment “Arms Race”: Who’s Not Counting Costs?

The large model industry is still a “cash burner.” Both companies dedicate a high intensity of R&D investment, but Zhipu AI is even more aggressive.

Zhipu AI: Super-saturated Investment

In the first half of 2025, Zhipu AI’s revenue was RMB 191 million, but R&D expenses reached RMB 1.595 billion. This means for every 1 yuan earned, over 8 yuan was spent on R&D, with the R&D expense ratio at 835.4%. This “super-saturation” investment is mainly in computing service fees (RMB 1.145 billion) and R&D personnel salaries, showing Zhipu’s firm determination to chase SOTA (state-of-the-art) model capabilities.

MiniMax: Improved Efficiency

MiniMax is also burning cash, but thanks to fast revenue growth, its R&D expense ratio is quickly dropping. In 2023, its ratio exceeded 2000%, but fell to 337.4% in the first nine months of 2025 (R&D spend $180 million vs revenue $53 million). As volume grows for consumer products, MiniMax is gradually diluting massive R&D costs, and operational leverage effects are becoming evident.

Globalization and Geographic Risk: Different Battlefields

The two companies have chosen different geographical battlefields.

MiniMax: International expansion as focus

MiniMax is highly globalized. The prospectus shows that in the first nine months of 2025, income from mainland China accounted for only 26.9%, while income from the U.S. market accounted for 20.4%, with other overseas regions also having significant proportions. Its hit product Talkie targets overseas markets.

Zhipu AI focuses mainly on the domestic market. While the prospectus shows some overseas income (Southeast Asia accounted for 11.1% in H1 2025), its core business still serves government and enterprise customers in mainland China.

Capital Structure: A Shareholder List Full of Giants

Behind the two unicorns stand half of China’s top tech capital.

  • MiniMax: Founder Dr. Yan Junjie holds shares and controls the company via voting rights structure. Major shareholders include Alibaba (Alisoft China) (approximately 13.66%), and Tencent (about 2.58%). Notably, gaming giant miHoYo is also a key shareholder, owning over 5%.
  • Zhipu AI: Shareholding is more diverse, with both market-driven and state-owned backgrounds. In addition to internet giants (Ant Group, Tencent, Meituan, Xiaomi), there are state-owned platforms (Social Security Fund Zhongguancun Innovation Fund, Beijing AI Industry Investment Fund), as well as renowned VCs like Hillhouse, Legend Capital.

Summary and Outlook

Currently, MiniMax has ample cash and financial assets (about $1 billion), while Zhipu AI’s cash and equivalents reach RMB 2.55 billion. In the short term, both companies seem to have enough “ammo” for the long technical race.

This Hong Kong IPO submission essentially presents two stories to investors:

  • MiniMax’s narrative: This is a story about “AI native super applications.” Investors must assess whether its consumer products can maintain high retention and leading model inference efficiency, ultimately achieving the internet logic of “traffic monetization covers computing cost.”
  • Zhipu AI’s narrative: This is a story of “AI new infrastructure.” Investors need to evaluate the depth of its moat in the enterprise market, and whether high gross margins from local deployment can offset large basic model R&D expenses in the future.

Who will be the “world’s first large model IPO” remains undecided, but these two prospectuses unambiguously announce that China's AI industry has moved from the noisy “hundred-model war” into the deep-water area of commercial deployment.

Risk Warning and Disclaimer ClauseMarkets pose risks, and investment should be cautious. This article does not constitute personal investment advice, nor does it consider the specific investment goals, financial situations, or needs of individual users. Users should consider whether any opinions, views or conclusions in this article are appropriate for their specific circumstances. Investment decisions made based on this article are at your own risk.