Transitioning from basic reasoning to tackling more intricate tasks, the advent of DeepSeek has largely marked the completion of this phase in AI's evolution. Looking ahead, achieving self-learning and self-iteration capabilities will emerge as the next frontier for large models to conquer.

Recently, the large model sector has witnessed remarkable dynamism, with Zhipu and MINIMAX making their market debuts in succession. Through their IPOs, these entities have showcased the present value of large models, with the market responding with its stance and assessments of these nascent, yet-to-be-profitable ventures.

This prompts us to ponder: What form will future AI large models take, and what narrative logic will they follow?

From the current market vantage point, the shift from concept to commercialization has become the inevitable trajectory for large model firms. Rapid expansion, swift financing, followed by the launch of large model products, coupled with projections of future market scale, underpin the current market valuations of large model companies.

It's noteworthy that the influx of substantial capital and the rush of tech firms into the arena have led to industry overheating. This excessive investment has resulted in most companies squandering resources on redundant foundational algorithms. Consequently, the current buzzword for the large model industry has shifted from 'scale' to 'innovation'.

This signifies that the narrative of large models as mere 'shells' no longer holds water. Those endowed with absolute product prowess and innovation that remains undisturbed in the short term will carve out a niche in the market.

Of course, in the long haul, short-term safety doesn't guarantee sustained security, and going public doesn't ensure absolute safety. Market valuations are not static; only technologically proven routes can command high valuations. Companies that bet on the wrong technologies may find themselves 'worthless' as certain technological routes are phased out.

From this vantage point, some investment institutions already harbor intentions to cash out, making IPOs a crucial 'exit' route. However, it's imperative to note that large models that can successfully navigate the market will ultimately yield above-expectation returns. Thus, risks and opportunities coexist.

Overall, the listings of Zhipu and MINIMAX have provided benchmarks for valuing large model companies, with the capital market being generous in its offerings. As of the latest closing, Zhipu's market value stands at HK$80 billion, while MINIMAX's reaches HK$112.8 billion. Whether future performance can meet expectations remains to be seen, but the market has already expressed optimism. We believe that as the iteration speed of future large model products accelerates, the commercialization knockout stage will also pick up pace.

Model as Product

In the AI era, everything is 'accelerating'.

From conception to product, from product to market launch, entrepreneurial myths in the AI era continue to unfold. Behind this lies the 'explosive' growth in capital expenditures of global leading tech companies.

Google anticipates that its capital expenditures will reach between $91 billion and $93 billion for the full year of 2025, surpassing the previously estimated $85 billion. Looking ahead to 2026, Google's capital expenditures are expected to increase significantly.

Microsoft's financial reports indicate that its capital expenditures for the third quarter of 2025 were $34.9 billion, with $80.5 billion for the first three quarters, primarily invested in infrastructure supporting OpenAI models and its own Azure AI. Judging from the capital expenditure data, Microsoft's capital expenditures are in an 'accelerating' state.

Meta has raised its full-year capital expenditures for 2025 to between $70 billion and $72 billion. Meta also cautions that its capital expenditure growth in 2026 will be 'significantly greater' than in 2025.

Amazon expects its full-year capital expenditures for 2025 to reach $125 billion, surpassing the market consensus of $117.5 billion, and will continue to rise next year.

Oracle has taken it a step further, resorting to debt expansion for its AI investments.

The 'reckless' investments by leading tech giants have propelled the global AI process into a phase of rapid development and iteration. Moreover, as investments accelerate, each company's competitive edge begins to emerge, and the industry's knockout stage will also pick up pace.

Currently, the development of the large model industry has transitioned from basic reasoning to tackling more intricate tasks. Many 'shell' large model companies have been weeded out at this stage. In the future, the company whose products can achieve self-learning and self-iteration will secure entry into the next phase.

At the recently held AGI-Next Frontier Summit, Lin Junyang, head of Alibaba's Qwen, proposed a clear viewpoint: 'Model as Product'. He stated that with the development of active learning and autonomous decision-making capabilities, future AI Agents will be able to undertake 'managed tasks'. AI Agents can not only execute tasks but also self-evolve and dynamically plan paths during the process. This capability places unprecedented demands on the model's stability, generalization, and reasoning depth. Therefore, he believes that when training a foundational model, you are essentially creating a product.

Yao Shunyu, Tencent's newly appointed Chief AI Scientist, argues that as the technological focus shifts to Agents, commercialization paths have also significantly diverged. He states that the logic of toC (consumer-facing) and toB (business-facing) will increasingly diverge. In the toC market, enhanced user experience does not necessarily lead to increased retention; however, in the toB market, what enterprises fear most is not slowness but 'being wrong and uncontrollable'.

Of course, from a global perspective, the tech industry has historically witnessed the emergence of dominant players, such as Amazon in e-commerce, Google in search, and Meta in social media...

Kan Jian Finance anticipates that with sufficient competition and consolidation in the large model industry, future global leading AI companies will also exhibit a trend of concentration. This implies that the massive capital expenditures of some leading tech companies may ultimately benefit others. Currently, at the application level, both toC and toB remain in a diverse and vibrant state, but how long this will last, we believe, will not be long.

Why Zhipu and MINIMAX

From the perspective of capital expenditures, the disparity is evident.

Yao Shunyu, Tencent's newly appointed Chief AI Scientist, believes that China already possesses capabilities comparable to or even surpassing those of the United States in business, industrial design, and engineering. The only area where more entrepreneurs and risk-takers are needed is in the exploration of frontier paradigms.

From the perspective of capital market progress, domestic large model companies are clearly accelerating their embrace of capital, as evidenced by the listings of Zhipu and MINIMAX.

So, why Zhipu and MINIMAX?

In reality, this represents a divergence and validation of two routes. Zhipu represents the toB business model, while MINIMAX represents the toC business model.

From a revenue perspective, Zhipu's revenue grew from 125 million yuan to 312 million yuan from 2023 to 2024; MINIMAX's revenue surged from 24.51 million yuan to 219 million yuan during the same period.

In the first half of 2025, Zhipu's revenue increased to 191 million yuan, but its net loss also expanded to 2.351 billion yuan. In contrast, MINIMAX's revenue grew to 379.7 million yuan in the first three quarters of 2025, with its loss widening to 3.638 billion yuan.

Reflecting on the capital side, the market values MINIMAX at HK$112.8 billion and Zhipu at HK$80 billion. From a market preference standpoint, MINIMAX commands a higher premium than Zhipu.

So, why does this phenomenon occur?

We believe that from a historical market logic perspective, profitability in the toC segment generally outperforms that in the toB segment. In the tech and internet industry, the toC market is generally more prone to explosive growth.

Relevant data shows that MiniMax was founded more than two years later than Zhipu, but by the end of the third quarter of 2025, MiniMax's products had covered over 200 countries and regions, with a cumulative personal user base exceeding 212 million. In the first nine months of 2025, the company's revenue grew by over 170% year-on-year, with overseas markets contributing over 70% of the revenue.

In other words, in just four years, the company has become one of the few global large model companies to achieve full modality capabilities and enter the first tier. Therefore, the current market valuation is within a reasonable range. However, it's important to note that this valuation is not long-term valid. In a rapidly changing industry, model iteration can significantly impact valuations. Moreover, as companies go public, future product considerations must also factor in changes in revenue expectations.

Relevant data indicates that Yan Junjie, the core founder of MiniMax, previously worked at SenseTime. In November 2021, Yan left SenseTime and founded MiniMax in Shanghai. From its inception, the company positioned itself in the research and development of foundational general AI technologies, focusing on foundational architecture, multimodal capabilities, and engineering system construction.

According to the prospectus, since its establishment in 2022, MiniMax has accumulated approximately $500 million in research and operational investments. Over four years of sustained investment, MiniMax has gradually formed a complete technological stack. However, this does not mean that this advantage will be maintained indefinitely.

The prospectus reveals that as of September 30, 2025, MiniMax held approximately $1.046 billion in cash and cash equivalents, providing relatively sufficient financial support for its continued investment in model research and computational resources. However, from a global perspective, domestic AI companies' capital investments are significantly lower than those of Google, Microsoft, Meta, and other enterprises. This means that for companies like MiniMax and Zhipu, commercialization must proceed at a faster pace.

From MiniMax's current route, AI company overseas expansion has become a super consensus. Whether the company's products can become global hits in the future is crucial. Additionally, sustained massive investments pose future risks for MiniMax. If its product investments cannot quickly translate into revenue, and future revenue growth slows down, the company's valuation will face significant challenges. At that point, it may encounter the same predicament as SenseTime currently faces.