Os grandes modelos agora vão cobrar no mercado secundário

问AI · 大模型公司为何陷入资金消耗循环？

**出品｜虎嗅科技组
**

作者｜宋思杭

编辑｜苗正卿

头图｜视觉中国

Just this week, according to market news, Anthropic and 月之暗面 have successively indicated plans to enter the capital market in 2026.

The former is valued at approximately $380 billion and has progressed to Series G financing; the latter is valued at about $180 billion, making it one of the highest-valued model companies before an IPO in China. Both companies’ foundational models have long ranked at the top of various tech community lists.

At the same time, OpenAI’s IPO plans have been frequently mentioned, and 阶跃星辰 has also been reported to be in the preparation stage.

This means that in China and the United States, the two countries with the strongest capabilities in large models, the leading foundational model companies will almost all head towards the capital market around 2026.

Large model companies are collectively moving towards the same “endpoint.”

Two signals emerge from this.

First, this round of large model competition has shifted from a focus on capability to validating commercial value.

For the past two years, the industry was more concerned with who was closer to SOTA and who could lead in the next model release. But now, the focus has changed; with the popularity of Agent applications like OpenClaw, everyone is beginning to pay attention to whether the capabilities of large models can be continuously supported by investments and whether they can be transformed into verifiable commercial value.

The second signal is that once model companies reach a valuation of $100 billion, the patience of the capital market has also reached its limit.

As of now, if these large model companies claiming to IPO in 2026 successfully go public, by the end of 2026, almost all leading large model companies in China and the U.S. will have converged in the secondary market.

When leading companies begin to intensively approach an IPO, it indicates one thing: this track can no longer rely on the patience of the primary market to maintain operations. Large model companies still need money. However, capital is beginning to be unwilling to wait indefinitely.

Over the past two years, the large model track has siphoned off a significant amount of funds. Whether in China or the U.S., leading companies have easily raised tens of billions or even over a hundred billion dollars. This money was seen as a “bet on the future” at the time, but now it is becoming an account that needs explanation.

If we calculate from the end of 2022 when ChatGPT appeared, it has already been three years. In three years, the valuations of large model companies have skyrocketed to over $100 billion. This cycle is exceptionally fast compared to any previous era.

So, why do large model companies need money so badly? And why is the primary market no longer able to provide funds?

Remember, when large model companies first emerged, no one regarded them as “infrastructure”; back then, infrastructure was computing power. But now, as AI applications bloom everywhere, large models have also become “infrastructure.”

If we value them as “infrastructure,” they are no longer a light asset business.

The essence of infrastructure is continuous investment. Whether it’s model training, inference costs, or ongoing iterative upgrades to maintain performance leadership, all imply one thing: these companies cannot build barriers through one-time investments; they must continuously consume resources.

In other words, large model companies are entering a structure of long-term capital consumption.

This also brings about a more realistic problem: their revenues currently cannot support this structure. Whether it’s API calls, enterprise clients, or subscription models, although all are growing, the growth itself is not stable, and prices are continually declining.

The stronger the model, the more calls it receives, and the higher the costs; yet price competition is continuously compressing profit margins. This is a typical business scenario where “the larger the scale, the greater the pressure.”

Therefore, the demand for funds from large model companies is not temporary but continuous. The issue is that this demand has begun to exceed what the primary market can bear.

The primary market can bear risks for a company, but the premise is that the risk has a clear time boundary. The biggest problem for large model companies is that—the time boundary is not clear.

They are neither like internet platforms that can quickly create network effects through user scale; nor like traditional infrastructure, which can gradually recover costs through stable cash flow.

They are in a more awkward position: needing to invest at the scale of “infrastructure” but not yet having the return capabilities of “infrastructure.”

When a single company needs continuous investments of tens of billions or even over a hundred billion dollars, and the return cycle remains uncertain, the logic of the primary market begins to break down.

In this context, when long-term capital is no longer willing to unconditionally extend life, financing is no longer an action that can be repeated indefinitely.

In such a situation, large model companies are left with one choice: to go to a place where they can continue to talk about the “future,” which is the secondary market.

It has been proven that market sentiment towards AI is already optimistic enough.

Taking MiniMax as an example, it opened high on its first day in the capital market, surged at one point during trading, and subsequently experienced noticeable increases over several trading days, with its market value continuously rising compared to the issuance stage. Similarly, the performance of “the world’s first large model stock,” 智谱AI, also followed a comparable trajectory, experiencing a brief fluctuation in the initial listing stage, then quickly rebounding, with single-day increases exceeding 8% or even 10% on certain trading days.

If we pull the timeline further back, comparing it to the previous generation of AI companies, such a trend is not common.

Whether it’s 出门问问 or 商汤科技, both experienced a lengthy valuation digestion period post-IPO. Their stock price performance was more cautious, and market skepticism regarding their commercialization capabilities long suppressed their valuation space.

So next, for OpenAI, Anthropic, 月之暗面, and 阶跃星辰, these large model companies about to IPO, we can only wait for the capital market to provide the final answer.