Just weeks after securing substantial backing from Silicon Valley venture capital firm Benchmark, the Chinese startup behind the viral AI agent Manus is reportedly weighing significant strategic shifts to navigate tricky US-China relations.
According to people familiar with the discussions cited by The Information, leaders at Butterfly Effect are contemplating setting up new headquarters outside China and potentially separating its global operations, centered on Manus, from its domestic business.
This internal deliberation follows a recently closed $75 million funding round led by Benchmark, which quintupled the startup’s valuation to $500 million. The move highlights the tightrope walk for Chinese tech companies attracting US investment while facing mounting scrutiny from Washington.
Previous investors in Butterfly Effect, who contributed to over $10 million in earlier funding and also participated in the latest round, included China-based heavyweights Tencent, ZhenFund, and HSG (formerly Sequoia China). The new capital injection, overseen by Benchmark partner Miles Grimshaw, is intended to fuel expansion into the US, Japan, and the Middle East, including a potential Tokyo office.
Manus: The Autonomous Agent Sparking Buzz and Bans
Manus AI made waves upon its public introduction around March 6, 2025, marketed as a different breed of AI. Unlike assistants requiring step-by-step instructions, Manus is designed as an autonomous agent—a system capable of understanding a goal and then planning and executing the necessary digital tasks without constant human oversight.
This might involve using techniques like LLM chaining, where multiple AI model interactions are sequenced, or multi-signature control systems, potentially for internal validation before acting. Co-founder and Chief Scientist Yichao “Peak” Ji, who previously developed the Magi search engine, described Manus as potentially “a glimpse into AGI” focused on delivering “results” rather than just assisting ideation.
The system, developed by Butterfly Effect (Hong Kong) Limited—also behind the AI browser tool Monica—apparently acts as an orchestrator, utilizing models like Anthropic’s Claude and Alibaba’s Qwen within a cloud environment.
This approach led to rapid scaling challenges and costs, with early reports suggesting tasks consumed about $2 of Claude compute each, quickly accumulating over $1 million in expenses.
Initial access was limited to invite codes, creating a frenzy where codes were reportedly resold for as much as ¥50,000 (around $7,000). Zhang Tao, a partner at the firm, stated they had “completely underestimated the level of enthusiasm” but denied the company was involved in selling access, claiming Manus “has never opened any paid channels for invitation codes” nor “allocated any marketing budget”.
Formal $39 and $199 monthly plans launched on March 31. While the company pointed to strong self-reported scores on the GAIA benchmark—a test measuring real-world task completion, reasoning, and tool use by AI agents—independent reviews noted performance quirks, including generating simulated data or getting stuck in loops.
The Geopolitical Tightrope
The agent’s autonomous nature immediately raised red flags beyond technical reliability. The lack of human-in-the-loop verification sparked warnings about potential misuse for fraud or disinformation. These concerns translated into swift government action: Tennesse Governor Bill Lee banned Manus from state networks on March 6citing risks of “censorship, propaganda, and bias,” and Alabama’s Governor Kay Ivey followed with a similar prohibition due to security vulnerabilities.
Butterfly Effect’s current strategic considerations are deeply intertwined with this climate of suspicion, particularly regarding AI developed in China. The situation echoes the challenges faced by DeepSeek AI, another high-profile Chinese lab. A recent report from the U.S. House Select Committee on the CCP branded DeepSeek a national security threat, alleging data siphoning, CCP-aligned censorship, and intellectual property theft.
OpenAI confirmed to the committee that DeepSeek employees “circumvented guardrails in OpenAI’s models to extract reasoning outputs,” a technique known as model distillation used to transfer capabilities. The report also probed DeepSeek’s alleged use of restricted Nvidia chips, further entwining AI development with hardware access controls.
The US government has already considered bills aiming to ban DeepSeek from federal devices, creating a palpable precedent for Butterfly Effect.
Broader Industry and Hardware Context
Butterfly Effect’s strategic planning occurs amid intense domestic competition and external hardware constraints affecting China’s AI sector. Players like Baidu are aggressively cutting prices on models like ERNIE Turbo to compete domestically. Simultaneously, tightening US export controls, such as the effective restriction on Nvidia’s H20 chip sales to China, are intensifying China’s push for technological self-reliance.
This has spurred companies like Huawei to ready mass shipments of domestic alternatives like the Ascend 910C chip, though achieving performance parity often involves trade-offs in power efficiency or reliance on complex supply chain maneuvers.
Other Chinese AI startups like Zhipu AI, also facing US restrictions, are pursuing global expansion through partnerships with firms like Alibaba Cloud while planning domestic IPOs. Butterfly Effect’s reported discussions about separating its international arm, potentially relocating its HQ, and maybe using Alibaba infrastructure domestically, represent a calculated strategy.
Backed by Benchmark’s capital, the startup aims to find a structure that allows Manus to thrive globally while navigating the complex web of technology competition, national security concerns, and shifting geopolitical alliances.
