China's AI Battle: The 'Hundred-Model War' Has Concluded, with the Largest Profit Pool Going to Major Players—How Will Zhipu and MiniMax Break Through?

When models are no longer scarce, what truly becomes rare is the ability to turn models into cash flow.

According to Zhui Feng Trading Desk, on February 9, JPMorgan Securities (China) released a research report titled 'China's Artificial Intelligence Industry: Global Expansion and Model Innovation Driving the New Generation of Leaders,' providing initial coverage of China's independent large model vendors Zhipu and MiniMax.

The report begins by stating that China’s artificial intelligence industry is transitioning from the 'hundred-model battle' phase to a new stage where the ability to commercialize, innovate in model development, and achieve global expansion will determine success. The Chinese AI market is rapidly consolidating, with the number of capable and well-funded model developers shrinking from over 200 to less than 10.

JPMorgan sharply pointed out that the largest profit pool in China’s AI industry may flow to platform giants who control distribution. Independent vendors, however, can break through by finding survival niches through 'structural neutrality'—Zhipu focuses on high compliance in localized deployment, while MiniMax expands into premium global markets.

The reasoning behind this judgment is not complicated. The report noted that as model training costs, computing power access thresholds, and commercialization difficulties continue to rise, hard constraints from capital and computing power have begun to dominate the industry structure. In other words, the industry no longer rewards 'whether you can build a model,' but instead rewards 'whether you can survive in the long term.'

In JPMorgan’s view, the core change in this phase lies in the gradual convergence of model capabilities, exponential increases in capital consumption, and customers beginning to focus more on 'delivery capability, stability, and sustainability.'

This means that the main competition among large models is shifting from a technology race to the ability to build commercial systems.

The most 'provocative' conclusion: the profit pool may not lie with model companies.

Among all sections of the report, the part most likely to spark market discussion is not the financial forecasts for Zhipu or MiniMax, but JPMorgan’s assessment of profit ownership.

The report explicitly states in the section 'Long-term Profit Pool of China's Generative AI Industry' that the enduring profit pool of generative AI may be highly concentrated among large internet platforms.

We still believe that the enduring profit pools of domestic generative AI will be highly concentrated among large internet platforms, particularly Tencent and Alibaba, as they control nationwide distribution, monetization channels, and high-frequency consumer and merchant transaction flows.

JPMorgan's reasoning is very straightforward.

First, platforms control distribution. The report points out that large internet companies naturally possess high-frequency user touchpoints and mature application scenarios, making it easier for AI capabilities to be 'internalized as features' rather than sold as standalone products.

Second, platforms control monetization pathways. Whether through advertising, e-commerce, gaming, content subscriptions, or enterprise services, platforms already have well-established revenue mechanisms, with AI serving more as a tool to enhance ARPU and conversion rates.

Third, platforms control high-frequency transactions and consumption flows. The report emphasizes, 'High-frequency usage scenarios determine inference call volumes and whether economies of scale can be achieved.'

The report provides examples of platform reach:

• WeChat is at the center of daily consumption activities, with a combined monthly active user base of approximately 1.4 billion.” Tencent has embedded the chatbot 'Yuanbao' into WeChat, “allowing users to add it as a contact for interaction without downloading a separate application.

• Alibaba, on the other hand, has integrated AI into its sales funnel: 'Alibaba has upgraded its Qwen AI application... fully integrating it into Taobao, Alipay, Fliggy, and AutoNavi... AI compresses the funnel from browsing to payment, thereby supporting higher conversion rates and potentially increasing ad yield and commission rates.'

Within this framework, model capability itself does not necessarily equate to profit potential. In the Chinese market, delivering AI capabilities to users and monetizing them may often be more important than the model itself.

This is also a statement repeatedly emphasized by JPMorgan.

The capability of a model does not necessarily translate into profitability; distribution and monetization pathways are particularly crucial in the Chinese market.

Is there still an opportunity for independent model companies?

Given the dominance of platform giants, where does the space for survival lie for independent model manufacturers such as Zhipu and MiniMax?

JPMorgan has not denied the value of independent model companies, but its assessment is notably more pragmatic. The report categorizes industry players into a dual-track competitive landscape: one group consists of comprehensive technology giants with full-stack ecosystems, while the other comprises independent model developers excelling in specific dimensions.

In JPMorgan's view, the opportunity for independent model companies lies not in direct competition with platforms, but in offering a 'structurally neutral' option.

The report mentions that the incentive structures of independent model developers differ fundamentally from those of platform-based companies; their goal is to empower client applications rather than compete with clients.

Independent providers typically monetize their models through APIs, enterprise licensing, or private deployment... These channels serve the same fundamental objective—maximizing model adoption and utilization—without requiring clients to be tied to a single infrastructure or software ecosystem.

For large enterprises, adopting platform models often entails potential strategic dependency risks; independent model providers, on the other hand, are more easily perceived as 'tool-oriented partners.' JPMorgan emphasizes:

Independent model providers mitigate these concerns through structural neutrality. Their business incentives focus on empowering client applications rather than competing with clients, thereby reducing perceived strategic and execution risks.

Zhipu: Securing cash flow through private deployment

Within JPMorgan's analytical framework, Zhipu is defined as a quintessential example of 'a structurally enduring localized business serving as an anchor, while reaching an inflection point for capability-driven API operations.'

1. Financial Reality: Localized Deployment Is the Current Profit Pillar

Zhipu’s business model is clearly divided into two segments: on-premise deployment and cloud-based deployment.

Data shows that Zhipu’s current revenue focus lies in high compliance demands: “In the first half of fiscal year 2025, 85% of the company’s total revenue came from localized deployment, with this segment achieving a robust gross margin of 59.1%, compared to a negative gross margin of -0.4% for cloud-based deployment.”

JPMorgan’s analysis notes that in China’s regulated industries (such as government, finance, and central state-owned enterprises), localized deployment is typically required.

This is not merely a one-off transaction. The report highlights: “As foundational models iterate, this installed base has the potential to evolve into upgrade-driven, recurring economic benefits.” This is because once a model is embedded in critical workflows, switching costs are substantial, and continuous model iteration can transform localized deployment into a quasi-SaaS-like recurring economic benefit.

2. Growth Inflection Point: Cloud-based APIs Poised for Expansion

While localized deployment generates high gross margins, the future of scalability lies in cloud-based APIs. JPMorgan believes that Zhipu is at a pivotal inflection point.

With the release of GLM-4.7, Zhipu’s strategic focus has clearly shifted toward intelligent agent systems and tool-augmented reasoning. The report notes: “We expect that as GLM-4.7 gains recognition within the global developer community (particularly in programming workflows characterized by high willingness to pay and usage intensity), its adoption rate will accelerate significantly.”

JPMorgan forecasts that as economies of scale take effect, “we anticipate both revenue and profit margins from cloud-based deployment will climb starting in the second half of 2025.”

3. Valuation and Forecast

Based on its solid localized foundation and high-growth API potential, JPMorgan has assigned Zhipu a 'Buy' rating with a target price of HKD 400.

• Growth Forecast: Revenue is projected to achieve a compound annual growth rate (CAGR) of up to 127% from 2026 to 2030.

• Profit Timeline: The company is expected to become profitable by 2029, with a normalized adjusted net profit margin reaching 20% by 2030.

• Financing Needs: The company may require external financing in 2026 and 2027, with an estimated annual funding amount of RMB 5 billion.

MiniMax: Expanding Capabilities Through Global B2C Growth

While Zhipu exemplifies deep cultivation of the domestic B2B market, MiniMax is described by JPMorgan as a 'full-spectrum AI enterprise with a scalable growth engine,' characterized by its core attributes of being 'inherently global' and 'multi-modal.'

1. Revenue Structure: Overseas revenue exceeds 70%, with business divided into three key segments.

MiniMax exhibits a revenue profile that starkly contrasts with other domestic manufacturers.

The report disclosed a striking figure: 'For the first nine months of 2025, 73% of the company’s total revenue was generated from markets outside China, with deployments spanning over 200 countries and regions.'

This globalized layout has brought significant economic flexibility. JPMorgan noted: 'In the context of high inference costs and intense domestic competition within the industry, entering international markets, diversifying the customer base, and accessing differentiated pricing environments provide the company with structural advantages.'

In terms of business composition, MiniMax has achieved an excellent balance of risks: 'For the first three quarters of 2025, the revenue contributions from the open platform, generative media, and AI companion businesses each accounted for approximately one-third.'

• AI Companion (Talkie/Starry): Contributes 35% of revenue. JPMorgan predicts that by 2030, the paid rate for this business will reach 18% (benchmarking Tencent Music's 2023 level), with an annual ARPU of USD 31.

• Generative Media (Conch AI): Contributes 33% of revenue. Provides video tools to content creators, achieving an annual ARPU as high as USD 75.

• Open Platform (API): Contributes 29% of revenue. Serves 132,000 developers, with an annualized ARPU of USD 8,200 for paying users.

2. Technology Strategy: Technology as a Product

MiniMax’s technology strategy is summarized as 'full-stack' and 'multi-modal.' The report highlights that MiniMax employs a Mixture of Experts (MoE) architecture, iterating at an extremely rapid pace: 'Model iterations occur every two months (faster than the industry average of 3-4 months).'

This speed is attributed to its unique 'dual-engine' strategy: using consumer applications as validators for technology.

'Unlike many AI labs that first build models and then search for use cases, MiniMax develops both models and consumer products simultaneously... With millions of users interacting with Talkie daily, MiniMax receives real-time feedback... This proprietary interaction data is fed back into the R&D process to fine-tune the models.'

3. Valuation and Forecast

In view of its scarce global capabilities, JPMorgan has assigned MiniMax an 'Overweight' rating with a target price of 700 Hong Kong dollars.

• Growth forecast: Revenue CAGR for 2026-2030 is projected to reach as high as 138%.

• Profit timeline: The company is expected to become profitable starting in 2029, with the adjusted net profit margin normalizing to 24% by 2030.

• Financing needs: The group is projected to require external financing in 2026 and 2027, with an estimated annual amount of 700 million US dollars.

A decisive variable: Inference costs

In its in-depth analysis of the two companies, JPMorgan revealed a shared financial inflection point within the industry, which is critical for understanding the long-term value of AI companies: the computational cost structure will completely shift from 'training-driven' to 'inference-driven'.

The report noted that while total computational consumption will continue to expand, the growth curves and cost drivers of 'training versus inference' will show significant differences compared to the expansion phase of 2022-2025.

1. Training costs: Moving towards 'normalization'

With the establishment of foundational model architectures, pre-training for cutting-edge expansion will become more selective. JPMorgan predicts:

• Zhipu: The percentage of training costs in total computational costs will decrease significantly from 93% in 2025 to 32% in 2030.

• MiniMax: The proportion during the same period will decrease from 80% to 28%.

This implies that the previous 'arms race'-style expenditure, which pursued parameter scale without regard for cost, will come to an end, and R&D spending will enter a more rational 'normalization phase'.

2. Inference costs: Becoming the dominant expense

Future competition will focus on the efficiency of inference. JPMorgan predicts:

• Zhipu: The proportion of computation costs related to inference will surge from 7% in 2025 to 68% in 2030.

• MiniMax: The proportion during the same period will increase sharply from 20% to 72%.

This shift will have profound implications for financial models: computing power expenditures will gradually shift from 'R&D expenses' (R&D) to 'cost of goods sold' (COGS). This also explains why JPMorgan places such emphasis on API pricing, inference efficiency (single GPU invocation duration), and utilization rates as key determinants of gross margin.

This means that the core of future competition will no longer be about 'who can train larger models,' but rather: who has cheaper inference; who achieves higher utilization; and who can control pricing power.

In JPMorgan's view, the value of Zhipu and MiniMax does not lie in challenging platforms, but in occupying positions outside the platform that are nonetheless indispensable.

Editor/Lambor