Inside Meta's Multi-Billion Manus Deal: Why This AI Agent Is Different
Meta’s multi-billion dollar acquisition of Manus on December 30, 2025, marks a rare milestone: a Chinese-founded startup acquired outright by a U.S. tech giant, not just for talent or patents, but as a going concern with real revenue and users.Yet the deal has raised questions, particularly in tech circles. On Hacker News, engineers noted that Manus relies on publicly available models like Claude, questioning its technical differentiation. The valuation timeline seemed unusually fast. Manus launched in March 2025, raised $75 million at a $500 million valuation just a month later, and sold to Meta before year-end. Some commenters wondered whether the deal was driven more by relationships than revolutionary technology. Users who’d tried the product had mixed reactions, with several reporting it wasn’t significantly better than existing alternatives like ChatGPT. The prevailing view suggested Meta might have acquired strong go-to-market capabilities rather than breakthrough technology. “Meta acquired a great marketing team,” one commenter observed, noting that “their marketing skills are far superior to their technical skills.”
But to understand what Meta actually bought, and whether these questions are fair, we need to look at Manus from a Chinese perspective, where the company’s journey reveals a fundamentally different approach to building AI products.Manus isn’t a foundation model company. As founder Xiao Hong repeatedly emphasized, Manus never tried to become the next OpenAI or Google DeepMind. Instead, through in-depth conversations with Xiao Hong, his technical co-founder Ji Yichao (Peak), and early investor Liu Yuan from ZhenFund, we can see the unique niche this company established from Day One: If large models are chips, then Manus aims to be the consumer electronics company that packages chips into an iPhone.
I. Core Philosophy: Don’t Be Intel, Be Apple
In early 2023, as the AI boom exploded and capital chased “China’s OpenAI,” Xiao Hong proposed what seemed like a contrarian view: “Large model companies are very much like chip companies, like Intel or Qualcomm. We’re like a consumer electronics company, like Apple, Xiaomi, or DJI.”
Xiao Hong’s logic was built on a deep understanding of industry value chains. He believed model-layer competition follows Moore’s Law and cost-performance dynamics. The core is producing stronger intelligence at lower cost. It’s an extremely capital-intensive, commoditized battlefield that would ultimately resemble the chip industry, with only a few giants remaining.
In contrast, application-layer companies build moats through brand, distribution, supply chain management, and differentiated user experience. Manus’s strategy was “don’t own models,” treating them as generic industrial raw materials instead. Xiao Hong was blunt in interviews: “We don’t build foundation models ourselves... because model capabilities drive this wave of technological dividends, and model companies deserve to be rewarded. But from Day One, we felt we didn’t have those resources. It was a very pragmatic choice.”
This “pragmatism” allowed Manus to flexibly integrate the world’s most advanced models (whether GPT-4, Claude 3.5, or DeepSeek) while focusing on solving the “last mile” interaction problem. Xiao Hong called it the “new era’s Andy and Bill’s Law”: hardware (model) performance improvements must ultimately be consumed by software (applications) and converted into user-perceivable experiences.
II. Redefining Agents: From “Chat Box” to “Async Intern”
When Manus went viral in China in early 2025, tech forums dismissed it as a “shell company” with no real technology. Just another wrapper. But real users were burning hundreds, even thousands of dollars in tokens. Mark Zuckerberg became a customer before acquiring the company. What they recognized was that Manus had done something technically clever on the engineering side.
What Meta likely values in Manus is its radical redefinition of the General Purpose Agent form factor. Before Manus, most AI products (including Xiao Hong’s earlier product Monica) resembled “feature phone” era stacking. Adding search plugins and image generation tools to chatbots created what Xiao Hong mockingly called “Frankenstein products.” But the Manus team realized true agents shouldn’t be synchronous chatbots responding in lockstep, but interns with asynchronous planning capabilities.
“Human communication isn’t waterfall-style (you say something, I respond),” Xiao Hong explained. “A real intern, when you give them a task, might spend time researching and thinking, discover they made a mistake and restart, then deliver results.”
To realize this vision, Manus built a unique “universal container” architecture, earning it the tech media label “terminator of various AI applications”:
Virtual Server + Full-Featured Browser: Manus’s agent doesn’t just generate text in a chat box. It has a cloud-based virtual computer and browser. It can open webpages, click buttons, and even use keyboard shortcuts like a human.
Async Planning & Self-Correction (Yolo Mode): It has long-range planning capabilities. In internal testing, to answer questions about specific YouTube video details, Manus’s agent not only autonomously opened a browser but learned to use YouTube’s keyboard shortcuts to precisely control video playback and capture specific information from frames. More importantly, it introduced a “Yolo” (You Only Look Once) mode: encountering errors, it automatically feeds error messages back to the model for correction rather than throwing problems back to users.
Code Execution Capability: It can autonomously write and run code to solve complex data processing tasks. Xiao Hong mentioned watching his agent download and run code repositories from GitHub to complete tasks, a moment that felt like “being struck by lightning.”
This architecture solves large models’ “hallucination” and “all talk, no action” problems. It no longer just “says” answers but “does” them by using tools.
III. Team DNA: From “Rejected” to “Survivor”
Manus’s story is a classic underdog reversal narrative. The company was rejected by nearly every mainstream Chinese VC in its early days.
Xiao Hong isn’t a CS PhD AI scientist from a prestigious university, but an obsessive product manager. His early investor, Liu Yuan from ZhenFund, recalled that Xiao Hong’s ventures were dismissed as “tool products without moats.” At his lowest point, Xiao Hong reportedly cried on a Shanghai street after coming closest to securing top-tier VC funding, only to fail.
But this “outsider” status forced the team to develop extreme survival skills and business instincts.
Ruthlessly Pragmatic Commercialization: When ChatGPT released its API in late 2022, Xiao Hong didn’t blindly fundraise. Instead, he set an extremely pragmatic goal: “First make 500,000 RMB monthly (about $70K) to sustain the team.”
Traffic Operations & M&A: To quickly validate product-market fit, he flew to Shanghai during 2023’s Spring Festival and decisively acquired a browser plugin called “Chat for Google.” This move brought critical early traffic and user feedback to Manus (then called Monica), validating that “browser plugins” were an efficient form factor for AI deployment.
Global Vision: Though the team was based in China, Xiao Hong insisted on “only targeting overseas markets” because he judged as early as 2022 that overseas SaaS users had 5x the payment willingness of domestic users, with a more favorable competitive environment for “small hill” survival strategies.
This “profit-first,” “traffic-is-king” grassroots wisdom let the Manus team establish a foothold in overseas markets without massive funding, ultimately building a breakthrough product like Manus.
IV. Marketing & Controversy: A Perfect “Hunger Marketing” Campaign?
In March 2024, Manus’s launch triggered viral-level propagation, dubbed a “Silicon Valley all-nighter” moment by media.
Investor Liu Yuan recalled that the morning after launch, his WeChat messages hit “three dots” (over 999 unread). That week felt like living in a movie. Manus invitation codes were scalped on secondary markets for up to 100,000 RMB (about $14K) each.
Though outsiders suspected orchestrated marketing, even “riding DeepSeek’s wave,” Liu Yuan and Xiao Hong denied this in interviews. They noted the launch was actually accidental, and strict invitation code control stemmed from extremely high costs (single query costs reached $6-8), making large-scale opening impossible. This created the “hunger marketing” effect objectively.
The viral spread proved one point: the market desperately craves “real agents,” that Jarvis-like AI that works for you.
V. The Singapore Pivot: Geopolitical Headwinds
However, Manus then entered deeper controversy. After launch, because it used Claude’s model, Manus didn’t directly serve mainland Chinese users. They subsequently announced partnering with Alibaba’s Qwen to build a Manus for Chinese users.
But after accepting Benchmark investment, Manus relocated headquarters to Singapore, laid off most mainland Chinese employees, and wiped Chinese social media content. Conservative Chinese media outlet Guancha.cn even commented, “If the US market isn’t friendly to Manus, or even won’t open doors, where will Manus go?”
That question now has a clear answer. Meta has powerful distribution capabilities but lags competitors in AI products. Compared to Silicon Valley’s popular “acqui-hires,” the Meta-Manus deal is clearly a traditional acquisition. Manus’s acquisition by Meta provides a new paradigm for Chinese AI entrepreneurs. In the current environment, this outcome is especially precious.
Note: Details in this article are based on public podcast interviews with Xiao Hong (Manus founder), Ji Yichao (Manus co-founder), and Liu Yuan (ZhenFund partner) from 2023-2025.