In recent times, the hottest topic within China’s AI large model sector has shifted from simply launching new models to a widespread surge in company valuations.

Kimi’s parent company, Moonshot AI, is rumored to have secured approximately $2 billion in funding, with its post-investment valuation rocketing to around $20 billion. StepFun is said to have raised close to $2.5 billion and is now being considered for a Hong Kong IPO by the market. Meanwhile, DeepSeek, despite fewer public capital moves, finds itself immediately cast in discussions as the “next AI unicorn” whenever financing rumors arise.

Looking back further, Zhipu AI and MiniMax have successfully listed on the Hong Kong Stock Exchange, setting a crucial precedent for the market: China’s AI large model firms don’t necessarily need to remain hidden within tech conglomerates but can also be independently valued as “pure AI assets.”

In my view, this development carries far greater significance than the valuation of any single company.

It signals that China’s AI large model industry is transitioning from technical hype to gaining a foothold in the capital markets. Previously, discussions about large models often centered on who had the largest parameters, the highest rankings, or the most human-like responses.

Now, institutional investors are focusing on a different set of criteria: Do you have financing capabilities? Do you generate commercial revenue? Is your order pipeline visible? How long can your cash flow sustain operations? Can you provide a tradable valuation benchmark for the secondary market post-listing?

In essence, capital is no longer solely evaluating whether you “can build models” but is beginning to assess whether you “can become an asset.”

I. The Hong Kong Stock Market Provides the First Valuation Benchmark: AI Companies Finally Have a Tradable Price

I’ve always believed that many industries truly begin to be re-evaluated not due to a technological breakthrough but because the capital market first assigns a price.

After Zhipu AI and MiniMax went public, the market finally saw what China’s foundational AI large model companies look like in the secondary market. Previously, the high valuations of these companies were largely driven by bidding wars among primary market investors. One might claim a company is worth $10 billion, another $20 billion, based on model capabilities, founding teams, tech giant shareholders, financing rhythms, and future potential.

But the Hong Kong stock market operates differently. It provides a daily price, with daily capital inflows and outflows and daily emotional swings. It transforms an abstract AI narrative into a tradable stock. A price rise indicates increased risk appetite; a fall suggests the market is starting to scrutinize; increased trading volume signals capital inflow; valuation recovery indicates a willingness to re-examine such assets.

Thus, the greatest significance of Zhipu and MiniMax lies not just in their own listings but in providing an external reference point for subsequent companies like Kimi, StepFun, and DeepSeek.

The capital market is pragmatic. Since Zhipu and MiniMax can already be traded on the Hong Kong stock market, the remaining unlisted companies with stronger market recognition, larger user bases, and more distinct technological labels will be preemptively placed within the framework of the “next batch of AI IPOs” for revaluation.

This is why Kimi’s financing, StepFun’s financing, and DeepSeek’s rumors have been amplified into a new round of AI asset revaluation. On the surface, it appears that AI large model companies have suddenly become more expensive. In capital market terms, it’s actually the market re-searching for valuation benchmarks for China’s AI assets.

In the past, Chinese AI companies faced an awkward situation: the primary market was hot, but the secondary market lacked targets. If you wanted to invest in Chinese AI, your options were limited to internet giants, computing power chains, or application companies with less pure AI concepts. However, none of these assets were “pure” enough. Tencent has games and advertising, Alibaba has e-commerce and cloud, Baidu has search and Robotaxi. While they all incorporate AI, AI is only a part of their business.

Companies like Zhipu and MiniMax are different. Their labels are straightforward: we are large models, we are AI applications, we are the new infrastructure of the AI era. Once this asset attribute is accepted by the Hong Kong stock market, the primary market will naturally adjust valuations upward.

In my view, the primary driver of this round of large model valuation surges is not profit realization but rather the valuation transition brought about by “tradability.” The market finally has a set of referenceable pure AI assets, and leading companies will naturally be re-priced accordingly.

II. Model Companies Have Shifted Focus: Rankings Are Just Tickets, Commercialization Is the Examination

However, a problem arises: just because valuations have risen doesn’t mean the story is over. On the contrary, the higher the valuation, the harsher the market’s subsequent demands.

AI large model companies differ from traditional internet companies. Internet companies may burn cash in the early stages, but once they achieve scale, marginal costs can be very low. For social media, search engines, e-commerce, and content platforms, more users mean more ad inventory, higher commercialization efficiency, and the potential for profit elasticity to be released later.

AI applications, however, face a different reality. The more active users are, the greater the token consumption; the more frequent the calls, the higher the inference costs; the better the product, the greater the server pressure. You may see DAU (Daily Active Users) rise, making your financing PPT look impressive, but if the payment rate and average revenue per user (ARPU) don’t keep up, cash flow pressure will also increase simultaneously.

This is what I consider the most realistic contradiction for AI large model companies: they are priced by the capital market as “future infrastructure,” but the income quality of many companies has not yet fully proven they deserve infrastructure-level valuations.

Let’s break it down. Zhipu’s story leans toward government, enterprise, developer, and agent infrastructure. The advantage of this direction is more solid orders, clients more willing to pay, and easier alignment with government and enterprise digitalization budgets. The downside is heavy delivery, long project cycles, and revenue that may not scale as quickly as internet products. It resembles more of an “infrastructure + solutions” company in the AI era rather than a lightweight application company.

MiniMax’s story leans toward multimodal and global applications. Products like Hailuo AI and Talkie have reach, users, and overseas potential. The advantage is rapid growth and market sentiment that can easily command a premium; the downside is the most challenging question for C-end AI applications: are users willing to pay continuously? Can retention remain stable? Can inference costs be reduced? Will content compliance suddenly become a short-term disruption?

Kimi’s capital story revolves around long-text entry and super workflows. It once gained recognition for its long-context capabilities, but now that it’s been pushed into a high valuation range, the market will ask: can high user engagement translate into money? If it’s just high-frequency usage without sufficient payment conversion, it will become another source of pressure. AI applications are not traditional content platforms and cannot be simply explained by “more users are good.”

StepFun is even more intriguing. What the market focuses on is the possibility of models entering terminals. If future phones, cars, smart hardware, and office equipment all require a system-level intelligent agent, then model companies won’t just sell APIs but will have the opportunity to become part of the terminal industry chain. This path sounds exciting, but there’s also a question: who controls the terminal entry points? Between model companies, phone manufacturers, system vendors, and cloud providers, who takes the largest share of the profits?

DeepSeek is even more unique. Much of its market premium comes from model efficiency, technological influence, and its significance as a benchmark for domestic AI. It’s not valued entirely as an ordinary commercial company because it carries strong scarcity labels. However, scarcity is not a get-out-of-jail-free card. The higher the valuation, the more it needs to prove its commercial closed-loop later; otherwise, expectation gaps will turn into valuation suppression.

Therefore, I don’t agree with simplifying this round of large model valuation surges as “AI is hot again.”

AI is indeed hot, but what the capital market is trading now is not just hype but stratification: who merely has strong model capabilities versus who already has scenario-based revenue; who merely has rapid user growth versus who can convert users into cash flow; who merely wins on leaderboards versus who can secure positions among industry clients, developer ecosystems, and terminal entry points.

Rankings now serve merely as tickets; commercialization is the examination.

III. Valuations in the Hundreds of Billions Are Not the Finish Line but the Start of an Elimination Round

Next, China’s AI large model companies will enter a brutal phase: the leading ones will become increasingly expensive, while the mid-tier ones will face growing difficulties.

This is not alarmist but is determined by the industry’s cost structure.

AI large model companies are not ordinary startups. They require continuous training, continuous inference, purchasing computing power, hiring top talent, iterating products, and withstanding price wars from tech giants. Financing capabilities themselves will become a core competitiveness (competitive advantage). The more money you have, the more time you can buy; the longer the time, the closer you get to commercialization maturity; the closer commercialization is to realization, the smoother the next round of financing and IPO will be.

Conversely, if a model company lacks a clear ecological niche, stable revenue, or strong scenarios, relying on financing to survive will become increasingly difficult. Because the market has already seen Zhipu and MiniMax and will soon see more prominent players like Kimi, StepFun, and DeepSeek. Once capital begins to concentrate, mid-tier companies will quickly be marginalized.

In my view, the future valuation differentiation of AI large model companies will revolve around three key factors.

First, look at whether the income is of high quality.

API revenue is not difficult to generate, but price wars will compress margins; enterprise projects can bring revenue, but delivery costs are high; C-end subscriptions are imaginative, but payment rates and retention fluctuate greatly; terminal collaborations sound high-end, but profit distribution may not be controlled by model companies. Which companies will the market ultimately value highly? It will be those with high growth, high margins, low delivery costs, and strong retention.

Second, look at whether the scenarios are deep enough.

If an AI company only offers a generic chatbox, it will inevitably be outcompeted by tech giants. Alibaba, Tencent, Baidu, and ByteDance all have traffic, cloud services, data, ecosystems, and money. Independent AI large model companies that want to survive must either secure government and enterprise scenarios, office workflows, content production chains, terminal systems, or developer ecosystems.

Third, look at whether there is sustained catalyst after listing.

The capital market doesn’t fear companies losing money; it fears companies losing money without marginal improvements. As long as order visibility increases, revenue sustained growth (continues to grow), inference costs decline, and user payment rates improve, the market can continue to assign valuations. However, if there are no new products, no new clients, and no revenue realization after financing, only more stories, valuations will quickly be slashed.

Therefore, I believe this round of valuations in the hundreds of billions should not be viewed merely as a trend.

It’s more like an entry ticket. Achieving a high valuation means you have a seat at the table. But once seated, the market will start scrutinizing your financial reports, costs, orders, margins, cash flow, and client structure. Previously, you were seen as a tech star; now, you’re seen as a business.

This is also a crucial shift in China’s AI industry. Over the past two years, the market has been asking: does China have its own OpenAI? Now the question has changed: who can become China’s true infrastructure company in the AI era? Who can integrate model capabilities into enterprise processes, content production, office systems, smart terminals, and industry chains? Who can transform tokens from mere consumption into revenue and ultimately into profits?

In my view, this is the true bellwether behind AI large model companies’ valuations surpassing $10 billion. Not all model companies will become great companies. In fact, it’s likely that most model companies will eventually prove to be transitional species. The ones that truly survive cycles may not be the loudest or the most aggressively funded but rather those that earliest integrate AI into real businesses, real scenarios, and real cash flows.

Model rankings determine visibility, capital markets determine prices, and cash flow determines survival.

The first round of asset revaluation for China’s AI large model companies has begun. In the next round, the market will no longer pay solely for “being able to build models.” Only those who can convert tokens into revenue, solidify revenue into profits, and transform profit expectations into long-term valuation anchors will have the right to stay at the table.