The Life-and-Death Game of Large Models: From the 'Six Dragons' to the Dual Giants Going Public — The Bubble, Breakthrough, and Endgame of AI Entrepreneurship

marsbit发布于2026-04-03更新于2026-04-03

文章摘要

The Chinese AI large model startup landscape has undergone a drastic reshuffle in just two years. The initial "AI Six Dragons" quickly narrowed to the "Four Strong," and by early 2026, only Zhipu AI and MiniMax had successfully listed on the Hong Kong Stock Exchange, becoming the first independent large model companies to go public. The industry has shifted from a technology and capital-driven frenzy to a focus on commercial viability and sustainable business models. Zhipu AI and MiniMax, though now publicly traded, face immense pressure with significant losses, high valuations, and challenges in achieving profitability. Zhipu relies heavily on enterprise customization projects, while MiniMax depends on overseas consumer products with limited monetization. In contrast, non-listed companies like DeepSeek and Kimi have thrived by focusing on technical excellence and niche markets. DeepSeek targets global users with cost-efficient operations, and Kimi dominates long-text processing for professional use cases. Meanwhile, former contenders like Baichuan AI and 01.AI have shifted to vertical sectors, struggling against tech giants and thinner margins. The industry is governed by three key realities: only a few players can compete in the general-purpose large model space; public listings bring heightened scrutiny and inevitable valuation corrections; and vertical markets are highly competitive, not a safe retreat. The sector is expected to consolidate within one to two years, with...

Large model entrepreneurship was never an opportunity for everyone to get a piece of the pie, but a battlefield where the strong survive and the weak are eliminated.

Over the past two years, the domestic AI large model entrepreneurship market has undergone a ruthless industry shakeout. The "AI Six Dragons" that emerged from the 2023 hundred-model battle were soon reduced to the stronger "Four Little Giants." By early 2026, Zhipu AI and MiniMax took the lead in listing on the Hong Kong stock market, becoming the first truly independent large model companies to go public in China. In just two short years, the entire independent AI large model industry underwent a dramatic transformation: some fought desperately to break into the capital market, others gradually fell behind and lost competitiveness, some rose against the odds with solid technology, and others were forced to pivot and struggled to survive in niche sectors.

At the core of this industry upheaval is the shift of the entire track from relying on technological storytelling to speaking with commercial strength; from frenzied capital investment and追捧 (pursuit)回归 (returning) to rational valuation; from everyone crowding into developing general-purpose large models to making precise layouts in细分 (segmented) vertical fields. And going public was never the end goal for these AI companies, but a critical juncture determining their survival—companies that successfully listed seemingly obtained a capital access pass, while those that didn't faced direct survival pressure. The entire industry has entered the final stage where the strong remain and the weak are eliminated.

01

From Six Dragons to Dual Giants Going Public:

Rapid Industry Shakeout Driven by Capital and Computing Power

The companies dubbed the AI Six Dragons from 2023 to 2024 were once seen as China's hope to match top-tier international large models. Their founders were top talents in the industry, fundraising was effortless, valuations soared, and all eyes were on the general-purpose large model track, aiming to create products that could international顶尖 (top-tier)水平 (level). But this heated bubble lasted only a year before彻底 (completely) bursting. The pace of industry淘汰 (elimination) was faster than anyone anticipated.

The first to be淘汰 (eliminated) were the players who couldn't keep up. The Six Dragons directly lost two members, with Baichuan AI and 01.AI falling behind first. Building AI large models was never a low-cost endeavor—it requires continuous capital investment, massive computing power, and top technical talent. Each model upgrade costs hundreds of millions, and without tens of billions in annual R&D investment, it's impossible to maintain technological competitiveness. When DeepSeek broke through with solid technical strength and a more pragmatic business model, those companies lacking sufficient strength instantly lost their footing.

After becoming unable to compete in general-purpose large models, Baichuan AI had to completely pivot to the medical vertical field, focusing on specialized models for the healthcare industry. While it achieved some success in this professional domain, it彻底 (completely) withdrew from first-tier large model competition. 01.AI收缩 (contracted) its business even more, abandoning the large model R&D race and turning to lightweight, customized industry models. Core talent continuously流失 (bled away), fundraising became unsustainable, and it gradually faded from public view. Thus, the Six Dragons shrunk to four: Zhipu AI, MiniMax, Kimi, and StepFun (阶跃星辰), known in the industry as the Four Little Giants. But the competition didn't stop there.

Subsequently, the Hong Kong Stock Exchange relaxed listing rules for unprofitable tech companies, offering a new出路 (way out) for these cash-burning AI companies. Zhipu AI and MiniMax抢先一步 (took the lead) in listing on the Hong Kong stock market, becoming the first batch of listed large model companies. This看似 (seemingly) an industry breakthrough was actually an inevitable choice under capital and survival pressure: early investment institutions needed exit channels, and companies' massive annual losses could only be sustained by financing—going public became the only lifeline.

But after listing, all problems were laid bare. Two financial reports directly戳破 (punctured) the AI industry's bubble. Although Zhipu AI's 2025 revenue achieved significant growth, its losses approached 5 billion yuan, with revenue barely covering a quarter of R&D investment. Most revenue still came from customized projects, making it difficult to achieve scalable profitability. MiniMax同样 (also) suffered severe losses, yet its market valuation was pushed to extremely high levels. Revenue primarily relied on overseas consumer-end products, user growth gradually peaked, and high computing costs continuously吞噬 (devoured) revenue, making profitability始终 (consistently) unattainable.

For these two companies, going public was not a victory but moving forward under greater pressure. Extremely high valuations, continuously expanding losses, and immense pressure for performance growth weighed on them like three mountains. Seemingly successful in reaching shore, every step was taken as if treading on thin ice.一旦 (Once) performance fell short of expectations, market value would immediately shrink significantly.

02

Listing as the Tipping Point:

Three Divergent Fates for Independent AI Companies

The listings of Zhipu AI and MiniMax directly set the industry rules. 2026 became the life-and-death year for independent large model companies. Whether they went public or not directly determined their completely different developmental destinies.

The two companies that have already listed, seemingly possessing channels for continuous financing, are firmly shackled by high valuations. Zhipu AI relies on government-enterprise cooperation and corporate custom orders to stabilize its foundation. Its open-source models also have some market recognition, but the business is overly reliant on project delivery, making rapid scalable growth difficult. Profitability is still years away, and the future依旧 (still) requires continuous cash burning. MiniMax focuses on multimodal products, with a significant proportion of overseas users. Its consumer-end products have some user stickiness, but the profit model is very fragile (fragile). User payment rates are low, market competition is fierce, and computing costs are rising ever higher. Revenue growth can never catch up with cost increases.消化 (Digesting) the ultra-high valuation is almost an impossible task.

In contrast, the unlisted DeepSeek and Kimi have taken a different path of逆袭 (counterattack), becoming benchmarks among independent AI companies. DeepSeek, as a latecomer, has model performance approaching international top-tier levels while excellently controlling R&D and operational costs. It focuses on overseas markets to avoid domestic competition with giants, experiences rapid user growth, and maintains very healthy cash flow. Not needing to rush to list allows it to focus on refining technology, optimizing products, and gradually building its competitive moat.

Kimi accurately targeted the niche track of long-text processing, perfecting a single function to the extreme. It precisely captures the needs of professional groups like workplaces, research, and law, with extremely strong user stickiness and high willingness to pay. Once唱衰 (written off), it made a comeback through technological iteration, securing smooth financing and rapid revenue growth. Not being in a hurry to list allows it to maintain team control over the company, focus on R&D, and ultimately enjoy a more从容 (leisurely) development space.

In contrast, the companies that fell behind are not having an easy time. Companies like Baichuan AI and 01.AI, unable to keep up due to funding, technology, or talent, were forced to exit the general-purpose large model track and dive into vertical AI fields. But vertical fields are no safe haven. Lacking core technical barriers, they can only undertake customized industry projects, dealing with scattered clients, high delivery costs, meager profits, and还要 (also) facing pressure from giants like Baidu, Alibaba, and Huawei. They can only struggle in the cracks, completely losing any chance of returning to the first-tier track.

03

Industry Truths: Three Iron Laws

Determining the Final Outcome for Large Model Companies

The industry upheaval of the past two years is essentially the market规律 (laws) correcting the industry bubble after the capital frenzy subsided. Three stark industry iron laws had long destined the outcome for most AI large model companies.

The first iron law: General-purpose large models have always been a game for a few players; very few independent companies can remain. Developing general-purpose large models requires massive capital, computing power, data, and talent support. The larger the scale and the more users, the lower the costs and the stronger the competitiveness. Internet giants like ByteDance, Alibaba, and Baidu have inherent advantages in data, capital, and ecosystems that independent AI companies simply cannot compete with comprehensively. In the future, the general-purpose large model track will only have a few giants and two or three top independent companies left. The rest will either be eliminated or forced to转型 (pivot).

The second iron law: Listing is a double-edged sword. AI companies that are not yet profitable will eventually face rationalized valuations. The previous high valuations of Zhipu AI and MiniMax were merely the capital market透支 (overdrawing) the future of AI. But the capital market only looks at real performance, not illusory stories. If losses continue and profitability remains distant, valuations will inevitably fall significantly. For listed companies, the next two to three years are the profitability test—failing it means being abandoned by the market. For unlisted companies, inability to enter the capital market in the short term will lead to大幅 (significant) valuation shrinkage, resulting要么 (either) in acquisition by giants or outright closure.

The third iron law: Vertical AI is not a retreat but a more brutally competitive quagmire. Many companies think that if they can't compete in the general track, they can move into vertical fields. This is actually the biggest misconception. Vertical fields have low technical barriers, allowing giants to quickly launch similar products and capture the market with low prices. Moreover, vertical industry clients are scattered, demands are highly customized, delivery cycles are long, profits are low, and payment collection is slow. Without core technology and scale advantages, long-term survival is impossible, leading only to a困境 (predicament) of working harder for less profit.

The AI companies that can truly go the distance are never those that abandon the mainstream track for vertical fields. Instead, they leverage core large model technology to deeply cultivate specific scenarios, using technical advantages to reduce implementation costs. This preserves the technical moat while achieving commercial落地 (landing). This is the correct path for independent AI companies.

04

Future Prediction:

Industry Landscape Solidifies Within One to Two Years

Standing at this current juncture, the landscape of the domestic AI large model industry will彻底 (completely) solidify within the next one to two years.

The already-listed Zhipu AI and MiniMax will see significant differentiation. Zhipu AI can survive relying on stable B-end business, but growth will gradually slow, and its market value will return to reasonable levels. If MiniMax cannot solve its consumer-end profitability难题 (problem), its valuation will likely shrink significantly, and it might even be acquired by a giant.

Top unlisted companies like DeepSeek and Kimi will continue to lead凭借 (relying on) technological and product advantages. Their cash flow will gradually turn positive, valuations will steadily rise, and they will gain higher market recognition after their business models are fully proven and profitable before listing, becoming benchmarks for independent AI companies.

As for the fallen and pivoted companies like Baichuan AI and 01.AI, they will either be acquired (acquired) at low prices by relevant industry giants or exit the market entirely due to broken capital chains, never having another chance to return to the first tier.

The entire industry will eventually form a stable structure dominated by giants, coexisting with a few top independent companies, and small vertical companies深耕 (deeply cultivating) niches. The door to the general-purpose large model track will彻底 (completely) close, leaving no entry opportunities for latecomers. Vertical fields will also be carved up by giants and leading independent companies, and the industry will彻底 (completely)告别 (bid farewell to)野蛮生长 (wild growth).

04

Conclusion:

Listing is Not Reaching Shore,

It's the Rite of Passage for AI Companies

Over the past two years, the AI large model industry has moved from capital狂热 (frenzy) to market rationality, from a hundred schools of thought contending to残酷 (cruel) shakeout, completely leaving behind the bubble era where storytelling and burning cash could establish a foothold. It has entered the practical stage of competing on technology, products, and profitability.

Going public was never the final destination for independent AI companies but a rite of passage they must face. Either withstand the pressure, achieve the transformation from a tech company to a commercial company, and truly establish a profitable model; or fail the market's test and become a casualty of the industry shakeout.

For companies like DeepSeek and Kimi, not rushing to list has instead become their greatest advantage, giving them ample time to refine technology, accumulate users, and perfect their business models. For the already-listed Zhipu AI and MiniMax, listing is just the beginning—every financial report is a life-and-death test.

This industry shakeout also serves as a reminder to all AI entrepreneurs: large model entrepreneurship was never an opportunity for everyone to get a piece of the pie, but a battlefield where the strong survive and the weak are eliminated. Relying on capital and stories won't take you far. Only solid technology and a healthy business model can ensure a long-term foothold in the industry. This is the ultimate survival law of the AI large model industry.

This article is from the WeChat public account "竞合人工智能" (Coopetition AI), author: 竞合 (Coopetition)

相关问答

QWhat are the three core industry laws that determine the fate of AI large model companies, as mentioned in the article?

A1. General large models are a game for only a few players, with very few independent companies able to survive. 2. Listing is a double-edged sword; unprofitable AI companies will eventually face a return to rational valuations. 3. Vertical AI is not a safe fallback but a more brutal and competitive battlefield.

QWhich two AI companies were the first to go public in Hong Kong, and what challenges did they face after listing?

AZhipu AI and MiniMax were the first to go public in Hong Kong. After listing, they faced challenges such as high valuations, continuous and expanding losses, and immense pressure for performance growth. Their market value is at risk of significant shrinkage if they fail to meet expectations.

QHow did companies like DeepSeek and Kimi manage to succeed without rushing to go public?

ADeepSeek focused on overseas markets to avoid direct competition with domestic giants, controlled R&D and operational costs effectively, and experienced rapid user growth with healthy cash flow. Kimi excelled in the long-text processing niche, capturing strong user loyalty and willingness to pay among professional groups, allowing both to grow steadily without the immediate pressure to list.

QWhat happened to the companies that fell behind, such as Baichuan Intelligence and 01.AI?

ABaichuan Intelligence shifted entirely to the medical vertical field, exiting the top-tier large model competition. 01.AI scaled back its business, abandoned the large model R&D race, turned to lightweight industry-specific models, and gradually faded from public view due to talent loss and funding difficulties.

QWhat is the predicted industry structure for the AI large model sector in China within the next one to two years?

AThe industry will stabilize into a structure dominated by major tech giants, coexisting with a few top independent companies, and smaller vertical-focused firms. The door to the general large model track will close completely, and the vertical field will be divided among big companies and leading independents, ending the era of野蛮生长 (wild growth).

你可能也喜欢

Rubin Ultra大减配,英伟达也扛不住内存涨价了?

知名投研机构SemiAnalysis近日报告指出,英伟达已向主要客户预览了其顶级AI芯片Rubin Ultra的最新规格,但相比此前预期出现显著降配。核心变化包括:算力峰值保持与普通版Rubin相同的35 PFLOPs;显存容量降至8层堆叠的192GB,甚至低于普通版的288GB;显存带宽仅微升1TB/s;芯片功耗反而略有提高。Rubin Ultra的主要升级方向转向“扩展互联规模”,支持通过NVLink将最多576张GPU互联成统一计算域,远高于普通版的72张。 报告分析认为,此次调整主要源于HBM(高带宽内存)价格持续快速上涨。以HBM3为例,其价格已从2025年二季度的低点180-220美元,飙升至目前约700-850美元。HBM成本飙升使得Rubin Ultra单机架的物料成本一度从约660万美元升至800万美元。英伟达因此重新评估设计,通过减少昂贵的HBM配置(成本占比从近40%降至28%),将资源更多投入互联能力等方向,优化整体成本结构。 此消息引发市场对HBM需求见顶的担忧。受冲击影响,韩国存储股开盘大跌,SK海力士、三星股价均下挫约8%。市场解读认为,若英伟达此举成为趋势,意味着AI芯片厂商可能开始通过优化设计来降低对高容量HBM的依赖,这或将限制存储厂商未来的提价空间,标志着AI基础设施的“堆料涨价”时代可能接近尾声。

Odaily星球日报59分钟前

Rubin Ultra大减配,英伟达也扛不住内存涨价了?

Odaily星球日报59分钟前

生成模型也能端到端训练了?核心竟是一个for循环

长期以来,深度学习在图像分类、检测等任务上凭借端到端训练取得了巨大成功,但生成模型领域却始终是个例外。当前主流的自回归、扩散模型等训练时只学习预测“一小步”,推理时则需要反复展开数百上千步,这种训练与推理的不一致导致了“暴露偏差”等问题,使得生成模型一直未能实现真正的端到端训练。 最近,一篇来自UIUC与哈佛大学的论文提出了“探索式建模(Explorative Modeling,XM)”新范式,其核心是一个简单到近乎朴素的for循环:在每个训练步骤中,模型生成K个候选样本,然后仅选取其中最接近真实数据的一个样本来计算损失和回传梯度。这种做法巧妙地绕开了传统生成模型中因使用重构损失(如平方误差)而导致的“模态模糊”问题——即模型倾向于输出多个可能答案的平均值,而这个平均值往往不属于任何真实的数据模式。 XM通过这种方式,在不拆分生成过程的前提下,直接提升了模型的“生成表达力”,即模型捕捉数据分布中多个独立模式的能力。论文将“探索”验证为继模型参数量、数据量之后的第三根可扩展轴。实验表明,随着模型规模、数据量和计算量的增大,探索带来的性能收益越发显著,在图像、视频、语言等多种任务上均能带来效率与质量的提升。 更重要的是,当探索程度足够时,XM可以实现真正的端到端生成。在机器人控制等任务中,仅需一次网络前向传播的XM策略,其性能可媲美甚至超过需要上百次前向的扩散模型策略,实现了推理效率的飞跃。 尽管“最佳K样本选择”的思想并非全新,但该研究的贡献在于清晰地揭示了这一简单机制如何直接解决生成模型的核心瓶颈,并为实现高效、端到端的生成模型开辟了新的可能性。随着模型规模的持续扩大,探索这一新维度可能将发挥越来越关键的作用。

marsbit1小时前

生成模型也能端到端训练了?核心竟是一个for循环

marsbit1小时前

交易

现货
活动图片