The 400 Billion Yuan Reversal: Twilight Falls on China's AI Four Dragons

The old AI story has reached a dead end. The new AI story is just turning its opening page. The so-called "Four Dragons" that once represented the pinnacle of Chinese AI — SenseTime, Megvii, Yitu Technology, and CloudWalk Technology — are experiencing a collective twilight. Meanwhile, a new class of large-model upstarts represented by Moonshot AI, Zhipu AI, MiniMax, and Baichuan Intelligence is writing an entirely different narrative for a new era.

The Rise and Fall of the Four Dragons

Rewind to 2018 — the golden age of the AI Four Dragons. SenseTime's valuation once surpassed $10 billion, Megvii secured the world's largest single AI funding round, and both Yitu and CloudWalk dominated their respective niches. Together, the four companies commanded a combined valuation approaching 400 billion yuan, virtually monopolizing the top-tier narrative in China's AI sector.

Their core technology was heavily concentrated in computer vision — facial recognition, image analysis, and security surveillance. At that stage, these capabilities genuinely represented the cutting edge of AI commercialization. Capital poured in relentlessly, government contracts flowed without interruption, and the "AI unicorn" halo shone brilliantly.

However, the cracks beneath the surface had already formed. From 2022 onward, the Four Dragons' fortunes took a sharp turn for the worse:

• SenseTime: After listing in Hong Kong, its share price declined steadily, falling more than 80% from its peak. The company has sustained continuous losses, undergone multiple rounds of layoffs and restructuring, and seen hundreds of billions of Hong Kong dollars in market capitalization evaporate.
• Megvii: Its IPO journey has been fraught with setbacks, and the company has still not successfully listed on any capital market. Revenue growth has stalled.
• Yitu Technology: The company voluntarily withdrew its STAR Market IPO application. A co-founder departed, operations contracted dramatically, and Yitu has all but disappeared from public view.
• CloudWalk Technology: Although it managed to go public, its share price has been lackluster, losses have continued year after year, and the market has persistently questioned its business model.

The combined market capitalization and valuations of the four companies have shrunk by more than 70% from their peak. What were once hailed as "China's AI calling cards" now read more like a sobering cautionary tale.

The Wrong Path: Why a Head Start Led to Being Left Behind

On the surface, the Four Dragons' predicament looks like a valuation correction following capital withdrawal. The deeper cause, however, is a fundamental miscalculation on technology direction.

First, the ceiling of their chosen track was too low. Computer vision, and facial recognition in particular, is essentially a tool-level technology rather than a platform-level capability. It can solve specific problems in specific scenarios but struggles to build a sufficiently broad commercial moat. When growth in the security market decelerated and government contract cycles fluctuated, the Four Dragons found they had virtually no "second growth curve" to speak of.

Second, their business model was overly reliant on project-based delivery. The Four Dragons' revenue structures were heavily dependent on customized projects for government (ToG) and enterprise (ToB) clients. While gross margins appeared respectable on paper, delivery costs were extremely high, scalable replication was difficult, and cash flow was under sustained pressure. In essence, they more closely resembled systems integrators wearing an AI veneer than genuine technology platform companies.

Third, they strategically neglected general-purpose intelligence. When GPT-3 was released in 2020, the global AI industry was already making a conspicuous pivot — from specialized AI to general-purpose AI, from perceptual intelligence to cognitive intelligence. Yet driven by inertia, the Four Dragons continued channeling their primary resources into deeper vertical exploration of visual recognition. SenseTime later made an emergency pivot toward large models, launching its "SenseNova" large-model ecosystem, but this looked more like a hasty remedial strategy than a forward-looking initiative.

The Four Dragons had a head start, but they were running on a track that kept getting narrower. When the large-model wave came crashing in, they found that not only had they lost their lead, but their legacy baggage had left them even further behind.

The Large-Model Upstarts: An Entirely Different Species

In stark contrast to the Four Dragons' project-driven, fragmented-scenario model, the new generation of AI companies has aimed at a far more ambitious target from day one — artificial general intelligence.

Moonshot AI broke into mainstream consciousness with its Kimi intelligent assistant, cracking open the consumer market with its long-context processing capabilities. Its valuation soared to over $3 billion in little more than a year. Founder Yang Zhilin's thesis is clear: large models represent an infrastructure-level opportunity, not a vertical application.

Zhipu AI, backed by Tsinghua University's deep technical heritage, launched the GLM series of large models and pursues a dual strategy of open source and commercialization. The company has completed over 4 billion yuan in funding, with its valuation surpassing 20 billion yuan.

MiniMax is pushing forward simultaneously on multimodal capabilities and the application layer. Its consumer-facing products, including Hailuo AI, have gained strong traction, and its valuation has likewise exceeded several billion dollars.

Baichuan Intelligence, founded by former Sogou CEO Wang Xiaochuan, focuses on deploying large models in specific industries, rapidly penetrating verticals such as healthcare and finance.

The fundamental differences between these upstarts and the Four Dragons are stark:

Put simply, the Four Dragons sold "solutions," while the upstarts sell "intelligence itself." The former's value ceiling is determined by the market size of specific scenarios; the latter's potential is virtually boundless.

The Deeper Industrial Logic: AI's Generational Revolution

The decline of the Four Dragons and the rise of the large-model upstarts represent, at their core, a profound generational revolution within the AI industry.

The previous generation of AI (2015–2022) was built on a core paradigm of "perceptual intelligence" — enabling machines to see, hear, and recognize. While important, this capability is fundamentally "pattern matching" — statistical learning over existing data that lacks genuine understanding and reasoning.

The new generation of AI (2022–present) is built on a core paradigm of "cognitive intelligence" — enabling machines to understand, reason, generate, and converse. Through massive-scale text training, large language models have exhibited astonishing emergent general capabilities, representing a qualitative leap forward.

This generational gap has produced a harsh reality: the technology assets, customer relationships, and industry experience accumulated by the previous generation of AI companies are virtually non-transferable under the new paradigm. No matter how precise SenseTime's facial recognition algorithms are, they cannot provide an advantage in the large-model race. The moats of the old era have become the prison walls of the new.

Even more noteworthy is that large models are launching a "dimensional reduction attack" on the previous generation's core territory. Multimodal large models such as GPT-4V and Gemini have already demonstrated formidable visual comprehension capabilities, even surpassing specialized computer vision models on certain tasks. The technical barriers the Four Dragons relied on for survival are being dismantled by large models in an almost incidental fashion.

A Sober Reflection: Are the Upstarts Guaranteed to Win?

Of course, while applauding the large-model upstarts, we must also stay clear-eyed.

The Four Dragons were similarly elevated to the pantheon during a capital frenzy and similarly endowed with a grand narrative of "changing the world." Do today's large-model companies also harbor valuation bubbles? Can commercial deployment truly sustain multi-billion-dollar valuations? Is the cash-burning model of training ever-larger models sustainable?

In fact, some concerning signals have already emerged in the large-model space:

• Training costs remain stubbornly high, with a single training run easily costing tens of millions of dollars.