The future of domestic independent large model companies is likely to be going global + vertical-specific B2B.

Why Are Domestic Large Models Going Global Favored in Developing Markets?

The financial reports of Hong Kong-listed tech companies have been released, and the market’s focus is undoubtedly on two large model companies—MiniMax and Zhipu. From these reports, it’s clear that the revenue distributions of the two companies are almost completely opposite:

• Zhipu’s nearly all revenue comes from domestic sources, while MiniMax derives 73% of its income from overseas;
• The vast majority of Zhipu’s revenue can be defined as To B, whereas about two-thirds of MiniMax’s revenue is from To C (note: AI-native product revenue is not entirely, but mostly To C);
• 73.7% of Zhipu’s income comes from localized large model deployments, which are often highly customized based on industry and enterprise needs.

It is also rumored that Moon of Darkness is planning an IPO. We have yet to see its prospectus, but media reports indicate that its overseas revenue grew strongly in the first quarter of this year. Perhaps we will see Moon of Darkness similar to MiniMax, with most revenue coming from overseas? Regardless, it won’t be mainly relying on the domestic market like Zhipu.

I believe this undoubtedly represents two feasible development paths for independent domestic large model companies (“independent” meaning outside the big internet tech giants): going global and focusing on domestic vertical To B. It’s not necessarily a binary choice; they can do both, but in reality, there will always be a focus. As for other businesses, including domestic To C and “general-purpose” To B, they will definitely continue, but I don’t think independent large model companies have much competitive advantage there.

Why is that? Because domestic To C is clearly the territory of internet giants. In general AI applications, ByteDance’s Doubao, Alibaba’s Quark and Qianwen are very strong, and Tencent’s Yuanbao still hopes to enter the top tier. In vertical AI applications, early success stories like Ant A-Fu also show that as long as internet giants want to do it, their resource advantages and synergy effects are very obvious. Not only do they have infrastructure and traffic advantages, but they are also willing to provide ample free services, making it very difficult for independent large model companies to earn much from domestic C-end users.

Take DeepSeek as an example: it remains one of the top four or five AI applications domestically, but because it is open source, competitors are happy to offer “large volume, unlimited” free services to C-end users, so DeepSeek’s ability to charge domestic C users is very limited, mainly focusing on B-end. For other independent large model companies, unless they can launch models significantly better than big tech and keep them closed source, it’s unlikely they will earn high income from domestic C-end—this possibility, while not zero, is very slim.

As for the domestic “general AI To B” business, it’s even more a battleground for big tech: Alibaba Cloud, Volcengine, Baidu Cloud, Tencent Cloud are all fighting for this lifeline. Independent large model companies can leverage their agility to achieve results in some niche areas, such as the recent “Lobster Farming for All” trend, which has greatly increased demand for tokens among many large model vendors. But don’t forget, with the overall rise in computing power costs today, independent large model companies are price takers(Price Taker), while big companies with massive infrastructure and cloud divisions are price setters(Price Maker). In the long run, in terms of token costs, independent large model companies will find it hard to compete with big tech. Even if they develop techniques to reduce inference costs, big companies will clearly imitate and iterate in their next versions.

Historically, many reshuffles in China’s internet industry have originated from “dominant players neglecting or underestimating the importance of emerging competitors,” including ByteDance, Pinduoduo, and Kuaishou. However, in the AI large model field, there is no such timing window; big tech companies generally recognize the importance immediately, with the only debate being how much to invest and how to do it. In domestic C-end applications, the emergence of a viral DeepSeek was already an unexpected surprise; even the strongest independent vendors’ applications are likely to only secure a leading position in the second tier, difficult to rival the heavily armed big tech applications.

To B, on the other hand, is a different story: this market is extremely complex, with industry verticals and customer relationships being crucial. Many vertical clients find it difficult to get substantial benefits from big tech, and the difficulty of winning their business is high. Although big companies also want to develop MaaS (Model-as-a-Service) and AI-based SaaS, regardless of who provides these services, the underlying computing power will mostly come from big tech, and their IaaS businesses will continue to benefit. From both the customer’s and the big companies’ perspectives, it’s natural for some B-end, especially vertical B-end, to be handled by independent large model vendors—no need to seek change.

As for going global, I need not say more: over the past three years, one of the hottest startup tracks in China has been AI application exports. The overseas market is huge and highly segmented, including developed markets like North America and Western Europe, as well as Southeast Asia, South Asia—markets China tech companies are very familiar with—and increasingly important regions like the Middle East and Latin America. Whether in C-end or B-end, Chinese AI vendors (including large models and application developers) have two distinct comparative advantages:

First is application development and iteration capability—if domestic vendors claim to be second, no one dares to claim first. Early in the mobile internet era, many domestic companies’ motivation to go abroad was driven by the intense competition at home; those who couldn’t keep up domestically often found strong competitors overseas. Moreover, overseas markets are diverse, with different user preferences and regulations, so there are always places where they can compete.

Second is the cost-effectiveness of token costs. Recent media reports indicate that several domestic large model vendors have made a lot of money selling tokens in overseas, especially developing markets, because of low inference costs and high cost-performance, making them popular in some countries. This is indeed true (though specific figures are hard to verify). The underlying reason? Some mention infrastructure factors like power supply—yet the real reason is likely more complex.

The low-priced tokens offered by domestic vendors mainly come from models with smaller parameter scales. While their upper limits are limited, they are more than enough to meet the needs of many developing countries. This “cost-performance competition” strategy is effective; top Silicon Valley giants may not have the ability or willingness to imitate it, just as they find it difficult to compete on “cost-performance” in many other application fields. Localized operation and promotion are also very important, and Chinese vendors’ localization capabilities are well recognized, proven repeatedly even before the AI era.

I believe that in the foreseeable future, the development path for domestic independent large model companies will boil down to choosing between “going global” and “domestic vertical To B,” or perhaps doing both. Personally, I prefer the overseas route because domestic To B is a very tough, highly demanding business—my experience studying the domestic software industry confirms this. But the B-end market does exist, and enterprises across various industries do have AI transformation needs. Perhaps some independent large model companies will find sustained success in this field?

This article has not received funding or endorsement from any of the large model companies or their competitors mentioned herein.

The author currently does not hold shares in the large model companies mentioned, but cannot guarantee indirect holdings through funds.