Source | YuanSight

A Chinese AI startup leveraging an overseas model to attract global capital finds itself caught between policy and market dilemmas.

After nearly two months of silence, Manus, which gained significant attention in March, suddenly became the subject of headlines in foreign media. Recently, foreign media disclosed that Butterfly Effect, Manus's parent company, has completed a new round of financing worth $75 million, led by Benchmark, a renowned Silicon Valley venture capital firm, with participation from several existing investors. The company's latest valuation has reached nearly $500 million.

Under these circumstances, Butterfly Effect's valuation has quintupled. However, media outlets quoted insiders as saying that the company did not raise funds solely through Manus but through overall external financing, which also includes Monica, the company's first AI product that has successfully achieved commercialization.

Monica is an intelligent assistant claiming to be the "world's first universal AI agent." It is not satisfied with mere dialogue but tries to perform complex tasks akin to humans.

AI Concept Diagram | Created by YuanSight

Behind Butterfly Effect's funding halo lies the increasingly fierce competition in the domestic large model market. Major players like ByteDance and Baidu are advancing aggressively, while policy red lines have delineated forbidden zones for technological routes. Manus's "Achilles' heel" - the Claude model - happens to step on this red line.

The fact that Manus has attracted heavy investment from American venture capital firms is an unconventional script in the capital market. Perhaps it is an extreme tug-of-war between technology, policy, and survival.

01

Money from Silicon Valley

Butterfly Effect's sudden overseas financing has stirred complex emotions in the domestic AI circle. On the one hand, there is relief that Chinese AI technology is being recognized overseas. On the other hand, there is more reflection on the underlying issues of the domestic general AI startup environment.

Venture capital firm Benchmark is renowned in the field of technology investment, with a portfolio that includes star companies like eBay, Twitter, Snapchat, Instagram, and Uber. Benchmark's investment in Manus's parent company is seen as a high recognition of the company's specific technical direction and potential business model.

However, behind this affirmation lies a crucial but often overlooked fact: Manus's core model is not rooted in Chinese local technology. Manus has always insisted on using the Claude model, which, amid the current wave of domestic large models emphasizing independent controllability and "domestic AI," appears somewhat different or even incompatible.

"When Manus secured its previous round of funding, domestic large models still lagged behind international mainstream models, leaving little room for choice. But the situation has now changed dramatically," an investor told YuanSight. With the rise of DeepSeek and domestic large models already possessing the strength of first-tier models, capital is more eager to see a grand narrative about domestic AI.

However, Manus is still insisting on using overseas models. On April 16, at the Amazon Web Services Summit, Zhang Tao, co-founder of Manus, said that Claude is the best large model for creating intelligent agents.

Using the Claude model allows Manus to quickly launch AI agents and other products, rapidly validate concepts, and enter the market in the short term. However, it also brings some limitations. For example, data security and cross-border compliance risks will always loom over it, and reliance on overseas models brings potential policy uncertainties and operational risks.

So far, Manus's related layout for localization is still limited. It only announced in March that it would formally establish a strategic partnership with Alibaba's Tongyi Qianwen team. The two parties will realize all functions of Manus on domestic models and computing power platforms based on the Tongyi Qianwen series of open-source models.

"If Manus insists on using overseas models, it needs to take some corresponding measures for data security and localization," said the investor. Simply forming a strategic partnership with Alibaba is not enough to demonstrate the company's systematic layout in data security and localization.

Perhaps Manus's overseas financing is, to some extent, an endorsement of its existing technical route but may also be a reluctant choice due to its domestic market dilemma.

02

General AI 'Meat Grinder'

Another challenge that Manus faces, which cannot be ignored, is direct competition and business model misalignment with internet giants in the field of general AI agents in the domestic market.

Currently, domestic internet giants such as ByteDance, Alibaba, and Baidu have clearly identified general AI agents as important strategic focuses for their future. Leveraging their vast user base, rich data resources, substantial capital strength, and powerful technical teams, these giants are actively planning or have already launched their AI agent products.

More than a week ago, ByteDance launched the internal beta of its first AI agent product, "Kouzi Space." Positioned as the "best place for collaborative office work," the product adopts an invitation code access system, allowing users to choose "general interns" or expert-type agents in various fields as needed. Official demonstration scenarios cover diverse areas such as market research report analysis, hot finance data interpretation, college entrance examination volunteer planning, and travel itinerary customization.

Almost simultaneously, Baidu's AI agent "Xinxiang" App landed on the Android app market. Unlike ByteDance's invitation system, this product adopts a completely free strategy with no registration threshold. Its core functions focus on four major scenarios: complex task coordination and management, immersive travel planning, one-click generation of visual charts, and automation of routine work processes.

In the view of Zhu Xiaohu, managing partner of GSR Ventures, compared to vertical intelligent agents, general intelligent agents represent opportunities for large companies. This is because general intelligent agents require entry points, and the products of these internet giants are often integrated within their own ecosystems, possessing natural distribution channels and user stickiness.

In this competitive landscape, Manus's product strength has not demonstrated an overwhelming advantage that is "one dimension" higher than these giants. From the trial experiences of mainstream intelligent agents in the market reported by various media outlets, although Manus has some innovations in specific functions or user experience, overall, it has not yet created a generational gap with the products developed by giants that have mobilized elite forces and invested huge sums of money.

This means that with their absolute advantages in capital and technology, domestic giants' huge investments in the field of general AI can easily form a crushing posture in the market. For startups like Manus, lacking core technologies or product barriers, it is easy to be overwhelmed by the capital and technological advantages of giants in this "meat grinder"-style competition, losing their living space.

In addition, there is a certain degree of uncertainty in the cooperative relationship between Manus and domestic leading companies, which further weakens its position in the domestic industrial chain.

Screenshot from Manus's official website

Although Manus previously formed a strategic partnership with Alibaba, in Alibaba's already launched AI agent "Xinliu," it is only marked as a "Manus task" in some push notifications, suggesting the cooperative relationship between the two parties. However, Alibaba has not clearly stated whether the core technology of "Xinliu" is fully based on Manus, nor has it disclosed the specific depth and boundaries of the cooperation in detail.

This lack of transparency in the cooperative relationship has raised questions about Manus's dominant position and bargaining power in the entire AI enabling industrial chain. It is still unclear whether Manus is playing the role of a technology or service provider or a technology leader of the core platform.

03

Domestic Solutions

More deeply and potentially more decisively, Manus's strategy of insisting on using the overseas Claude model conflicts with the current national strategy and mainstream narrative logic of China's AI industry development.

The policy level is continuously increasing support for independent and controllable technology, and domestic large models are accelerating their catch-up and demonstrating capabilities comparable to international models in some areas. In the government, state-owned enterprises, and industries with high requirements for data security and supply chain security, solutions based on foreign models face strict scrutiny and restrictions.

An investor who wished to remain anonymous said bluntly, "If Manus specifically targets the overseas market, it also needs some core technologies of domestic solutions." In this investor's view, the most ideal AI overseas expansion story involves having domestic AI models and technologies as a premise, perhaps incorporating overseas elements in some non-core aspects to meet local needs.

Even purely from a commercial perspective, Manus's insistence on the Claude model also faces severe challenges. At present, domestic mainstream large models have demonstrated relative advantages in terms of token costs, providing application developers with more cost-effective options.

Executives from Butterfly Effect have stated on LinkedIn that despite Manus's ongoing restriction on the number of users, the money spent on the Claude model exceeded $1 million within two weeks of its launch.

Taking DeepSeek's R1 model, which has the highest cost price, as an example, there is a significant gap in token costs between the Claude 3.5 Sonnet model and it. Numerically, the input cost of DeepSeek R1 is only 18% of Claude's, and the output cost is 14.6%. If we take an interaction of 1 million tokens as an example (100,000 inputs + 900,000 outputs), Claude would need to pay $138 (100,000 × $3 + 900,000 × $15), while DeepSeek R1 would only need to pay $20.21 (100,000 × $0.55 + 900,000 × $2.19), a cost difference of 6.8 times.

Even so, some in the industry have raised objections to DeepSeek's costs. On April 25, Baidu founder Robin Li publicly commented on DeepSeek, stating that the problems with DeepSeek are slowness and high cost. The majority of large model API call prices in the Chinese market are lower than DeepSeek's full-blooded version and faster.

Therefore, as the internal competition among domestic models intensifies, it is only a matter of time before token and related costs further decline. If Manus insists on using the Claude model in China, it means facing enormous cost pressures.

In enterprise-level application scenarios, the invocation volume of AI services is tens of millions of times that of ordinary users. If the service cost of Manus is significantly higher than that of competitors based on domestic models, customers will naturally tend to choose the lower-cost option. Price disadvantages will severely restrict its market expansion and profitability.

AI Concept Diagram | Created by YuanSight

Changing models on the spot places immense pressure on the entire technical team. The underlying operating logic, API interfaces, and model characteristics of each large model are different. Switching from one model to another is not as simple as modifying a few lines of code. Especially for products like general AI agents that rely on the capabilities of underlying models to build core logic, changing models is almost equivalent to rebuilding from scratch.

This not only requires significant investment in research and development resources and time costs but may also lead to issues in product performance, stability, functional compatibility, etc., posing enormous technical risks.

A domestic technical staff developing AI humanoid writing told YuanSight that for an intelligent agent of Manus's scale, changing the large model is a complex system engineering. Changing the model requires large-scale adjustments and reconstructions to the entire technical architecture, algorithm logic, and engineering implementation. "Most importantly, there are few cases in China of changing large models on the spot, and everyone lacks experience, with too many unpredictable variables."

This huge investment in technology and time costs makes Manus, which relies on overseas models, appear somewhat awkward and passive in the domestic market. Overseas financing may provide it with a breathing space and the possibility of turning to external markets.

After Butterfly Effect completed the aforementioned financing, the market also heard news that Manus may analyze its international and domestic businesses separately. This also seems to be an attempt by the company to circumvent the aforementioned risks.

On April 27, media reported that Manus was considering relocating its headquarters overseas. The report pointed out that in recent years, tensions between China and the United States have affected the operations of many Chinese companies' global products. Since Manus's primary target market is the United States, the company is planning to establish a new headquarters overseas. It is reported that Manus has considered Singapore as the location for its new headquarters and has registered an entity there, but the specific plan has not yet been determined.

In addition, the company's founders have also discussed with some investors the possibility of completely separating international and domestic businesses, planning to let Manus focus on the global market outside China.

Manus is currently standing at a crossroads. If it can establish a true technological barrier overseas (such as deep modification of the Claude model), it may create a myth of counterattack; but if it remains trapped in the "Chinese team doing overseas markets" sandwich, it may ultimately become a victim of geopolitical technological competition.

Some images have been sourced from the internet. Should there be any infringement, kindly inform us for prompt deletion.