An internal beta invitation code once soared to 100,000 RMB (approx. USD 13,800 / KRW 18.9 million). Four months later, the official website was completely shut down to Chinese users. In 2025, Chinese AI star startup Manus staged a shocking “roller coaster” rise and fall. It burst onto the scene as a disruptor with the title of “world’s first general-purpose AI agent,” attracting over 2 million people to its waitlist in just three days, topping global benchmarks, and earning the title of “China’s AI beacon.” But just 127 days later, it exited quietly, with its core team hastily relocating to Singapore, leaving behind unanswered questions and a trail of chaos.
Did the fall of Manus expose the illusion of general-purpose AI agents? Or does it reveal deeper structural issues within the industry?
From a 100,000 RMB Invite to Shutdown: The Fall of a Chinese AI Star
On March 6, 2025, at dawn, a demo video lit up China’s tech community: on the left side of the screen, commands were entered into a chat box; on the right, real-time actions were executed. A so-called AI agent named Manus completed a full-cycle task—from price comparison on e-commerce platforms to generating an investment report—within 15 minutes. The “world’s first general-purpose AI Agent” had arrived.
Within three days, the waitlist on its website exceeded 2 million users, and the servers crashed due to overload. On China’s secondhand market platform Xianyu, beta invite codes were being sold for as high as 100,000 RMB (approx. USD 13,800 / KRW 18.9 million). In its launch week, it even overtook OpenAI on the global GAIA benchmark leaderboard.
However, by July, just four months later, the Manus website became inaccessible to Chinese IP addresses, and its social media accounts were wiped clean. Out of a 120-person team, only about 40 moved to Singapore, with the rest laid off. From peak to collapse—only 127 days.
Capital Frenzy: The 100,000 RMB Invite and the Myth of Chinese AI
In early 2025, China’s AI sector was booming. DeepSeek had just proven the strength of Chinese foundational models, and the AI field was in a “hundred-model war.” The market was hungry for the next disruptor.
Manus satisfied that hunger on March 6. Unlike most AI products, it claimed to have an “autonomous decision-making closed-loop system”—seen as a critical leap from “assisted decision-making” to “autonomous execution.”
Its defining feature was the ability to execute tasks independently through a multi-agent architecture—planner, executor, and verifier agents working together to handle everything from task decomposition and tool invocation to results validation. In a New York real estate example, Manus could break down user needs, call property APIs, write budget calculation programs, and output an interactive home-buying report.
This power stemmed from the Monica team’s deep application of scaling laws: when model size (175B-level GPT-4o), training data (tens of millions of user logs), and compute (distributed cloud infrastructure) form a golden triangle, AI exhibits emergent behaviors in tool use and system planning.
Even more impressive was its asynchronous execution engine. When prompted with “analyze the past three years of Tesla stock volatility,” Manus would spin up a cloud VM, autonomously complete data crawling, cross-checking, code writing, visualization, and more—17 steps—without human intervention.
This “digital intern” model redefined productivity. GAIA benchmark results showed that Manus outperformed entry-level human analysts by 9.3% in complex problem-solving accuracy.
In the demo, Manus resembled a true digital employee: receiving commands, decomposing steps, invoking tools, and delivering complete results. This harmony between “knowing” and “doing” drove tech enthusiasts wild. Media dubbed it “China’s AI beacon,” and state broadcaster CCTV called it a “new paradigm of human-machine collaboration.”
But Manus’ explosive popularity wasn’t just about technical innovation. Its invite-only model drove extreme scarcity. Invite codes reached prices where people joked “you could hire a real assistant for six months instead.” Amid the craze, co-founder Zhang Tao had to apologize publicly, citing limited server capacity and assuring waitlist users that expansion was underway.
Capital followed closely. In April, Silicon Valley VC giant Benchmark led a $75 million Series B round, valuing Manus at $500 million—setting a record for a Chinese AI application-layer startup.
At the funding press event, founder Xiao Hong declared, “Manus will redefine human-machine collaboration, evolving from ‘AI assistant’ to ‘AI colleague.’” At the time, no one questioned the vision.
Yet hidden risks were already surfacing. Even as invitation prices climbed into six-figure territory, actual users reported issues: high task failure rates, infinite loops with complex commands, and hallucinations.
Collapse: From Tech Disputes to Capital Pressure
The downfall began when Manus opened public registration. On May 12, it launched a three-tiered subscription model: Basic ($19/month), Plus ($39/month), and Pro ($199/month). These prices were far beyond what Chinese consumers could accept—especially when ChatGPT Plus cost only $20/month.
The market reaction was brutal. Monthly active users fell from a March peak of 20 million to 10 million by May. Paid conversion was under 3%, far below the industry average of 8–10%. Enterprise subscription renewal was only 45%.
Worse problems surfaced. Tech insiders revealed that Manus mainly relied on Anthropic’s Claude and Alibaba’s Tongyi Qianwen—making it a case of “assembly-style innovation.”
Associate Professor Tan Jian of Beijing University of Posts and Telecommunications commented: “Its ‘autonomous execution’ is essentially automated workflows built via API calls.”
User tests exposed fatal flaws. In March, only 4 of 8 tasks succeeded (50%). Its travel itinerary feature recommended nonexistent deals; its code often had redundancy.
The real turning point came from capital regulation. Under the U.S. Outbound Investment Security Program, American investments in China’s critical tech—including AI—are restricted. To avoid scrutiny, Benchmark forced Manus to move its HQ to Singapore.
That decision triggered a chain reaction: two-thirds of Manus’ China staff were laid off; its partnership with Alibaba’s Tongyi Qianwen ended; Chinese-language development stalled. By June, Manus ads began appearing in Singapore subway stations, while its Chinese social accounts went silent.
“It’s a textbook case of Silicon Valley capital sidestepping U.S. regulation,” one industry insider noted. “Just like how AI video startup HeyGen moved from Shenzhen to Los Angeles after raising funds.”
Double Squeeze: Compute Bottlenecks and a Broken Business Model
Moving to Singapore didn’t save Manus—it accelerated the collapse.
U.S. export controls on GPUs to China meant Manus couldn’t access NVIDIA’s latest H100 chips. Insiders said its ability to handle complex tasks dropped by 30% compared to launch.
Cost structure was another fatal flaw. As a multi-model AI agent, Manus cost about $2 per task. Its “diseconomies of scale” were clear: more users meant bigger losses.
Its flawed business model became more obvious in contrast. Competitor Genspark, focused on web page analysis, reached $36 million in ARR (annual recurring revenue) in 45 days. Manus’ “do-everything” approach looked clumsy.
“Consumer-facing AI agents are money pits unless they find a killer vertical use case,” one investor bluntly said.
On July 19, co-founder Ji Yichao published a technical retrospective, indirectly admitting strategic mistakes: “Manus bet on context engineering instead of foundational models.”
He explained that during his last venture, their self-trained model became obsolete overnight.
But it was too late. In just four months, the $75 million round was nearly burned through. Its technical foundation collapsed amid turmoil, and user trust in China hit zero.
After the General AI Dream: Where Should Chinese AI Go Next?
The collapse of Manus pierced the bubble of general-purpose AI agents.
Fang Han, CEO of Kunlun Wanwei, pointedly remarked: “General-purpose often means lack of specialization.”
Industry data shows a new path: In 2025, China’s core AI industry is projected to reach 1.1 trillion RMB (approx. USD 152 billion / KRW 208 trillion), with 65% from vertical markets.
Enterprise AI applications have a 32.6% paid adoption rate—far higher than the consumer market’s 15.8%. Vertical large models in healthcare and finance now reach 93.6% accuracy, approaching human expert levels.
Success stories are already emerging.
- Genspark, focused on web analysis, hit $36M ARR in 45 days.
- Devin, a coding assistant, charges $500/month and still has a waitlist.
- New unicorns like Hippocratic AI (healthcare) and Cyberhaven (data security) are rising fast.
Experts point out a broader issue: “AI is plagued by tech exhibitionism—flashy demos don’t equal real user value.” This event offers three lessons:
- Beware of general-AI worship
- Localization is critical
- Technical pragmatism cannot replace foundational innovation
China’s AI industry is also becoming more pragmatic.
- Zhipu AI has shifted entirely to government tech projects.
- MiniMax moved from emotional consumer apps to B2B API services.
- Stepfun shut down its roleplay app “Bubble Duck” to focus on smart hardware integration.
As GSR Ventures’ Zhu Xiaohu predicted: “Large models will eat 90% of agents.”
With giants like DeepSeek, Alibaba, and ByteDance controlling the foundation model layer, startups can only survive in narrow application niches.
After Manus relocated to Singapore, core team salaries tripled—AI engineers now earn $16,000/month. But without the support of China’s vast market, model optimization has stagnated.
In his retrospective blog, Ji Yichao focused on technical details but avoided addressing why the company withdrew.
Manus ads remain prominent in Singapore’s subways, while in Hangzhou, its former office has quietly been taken over by a new medical AI startup.
Their product makes no bold claims of general intelligence. It simply achieves 98% accuracy in pathology analysis—and already has contracts with three top-tier hospitals. As DeepSeek founder Zhou Siyuan put it: “The ultimate goal of AI may be general intelligence. But the path to that future must be paved by solving one real-world industry pain point at a time.”
[Disclaimer]: The above content reflects analysis of publicly available information, expert insights, and BCC research. It does not constitute investment advice. BCC is not responsible for any losses resulting from reliance on the views expressed herein. Investors should exercise caution.