Meta's largest AI acquisition is stuck between two superpowers, and the product it bought runs on a competitor's brain.
On December 29, 2025, Meta announced it was acquiring Manus — a Singapore-based AI agent startup — for roughly $2.5 billion, plus a $500 million employee retention pool. The deal closed within days. The consequences are still arriving.
Manus launched in March 2025. Its demo video hit 200,000 views in 20 hours. Within eight months, it reached $125 million in annual revenue with 78 employees, making it the fastest startup in history to cross the $100 million ARR threshold. Give it a zip file of resumes and it doesn't just rank candidates — it reviews each document, cross-references skills against market trends, and produces hiring recommendations with supporting materials. Autonomously. While you sleep.
The word "manus" is Latin for hand. The product is built to be the hand that moves while you think.
Meta wanted that hand badly. Mark Zuckerberg has spent two years talking about "personal superintelligence" — AI that knows you and acts on your behalf. Manus was the closest thing on the market to that vision. The $3 billion price tag represented a 4x premium over Manus's valuation from months earlier.
But there's a problem with buying hands. Sometimes they come attached to a body.
Two Governments, One Deal
Manus was built by Butterfly Effect, a company founded in China in 2022 by three developers: Red Xiao, Peak Ji, and Tao Zhang. In June 2025, the entire core team relocated to Singapore, shut down Chinese operations, laid off domestic staff, and made the product unavailable in China.
The industry has a name for this: Singapore washing.
US investment bans prohibit American firms from backing sensitive Chinese AI ventures. Moving to Singapore sidesteps the restriction — technically. Benchmark, the Silicon Valley VC that invested in Manus before the relocation, is now under US Treasury Department investigation for potentially breaching those regulations. Senator John Cornyn raised the alarm publicly.
On the other side, China's Ministry of Commerce announced a formal review to determine whether Manus needed an export license when it moved its team, its code, and its technology out of the country. Beijing is probing whether a Chinese-born company can simply walk its intellectual property across a border and sell it to an American tech giant for billions.
Meta spokesperson Andy Stone confirmed that once the deal closes, no Chinese ownership stakes will remain. Manus will shut down all services and operations in China.
One deal. Two investigations. Both governments treating the same acquisition as a potential national security threat — for opposite reasons.
The Architecture Irony
Here's the part nobody at Meta wants to talk about: Manus is built on Anthropic's Claude Sonnet.
The agent's multi-step reasoning — the thing that makes it worth $3 billion — runs on a competitor's model. Meta spent $135 billion on AI infrastructure in 2026. It built Llama, one of the most capable open-source model families in the world. And it bought an agent powered by Claude.
This isn't unusual. Most successful AI agents are model-agnostic — they route to whatever LLM handles a given task best. But it creates a dependency that looks strange on a balance sheet: Meta's flagship agent product relies on technology from a company valued at $380 billion that directly competes with Meta AI.
Meanwhile, Meta banned OpenClaw from its corporate networks over security concerns. It acquired a Chinese-founded agent startup instead.
The Customer Problem
Some of Manus's customers didn't celebrate the acquisition. They left.
Seth Dobrin, CEO of Arya Labs, called Manus his favorite agentic AI platform. He stopped using it. "I do not agree with a lot of Meta's practices around data and how they essentially weaponize people's personal data against them," he told CNBC. "I don't want to engage with a company who I don't feel comfortable with how they're going to use data."
Karl Yeh, co-founder of consulting firm 0260.AI, dropped Manus entirely and recommended his clients do the same.
The pattern is familiar. Meta has tried and failed to crack enterprise repeatedly. It shuttered Workplace, its corporate communication platform, in 2024. It killed Portal, its video calling device, in 2022. Each time, the same trust gap: companies that tolerate Meta for consumer advertising refuse to hand it their operational data.
Manus was different — a product people trusted enough to feed resumes, financial documents, and business workflows. Meta's name on the door changed the calculus overnight.
What Meta Actually Bought
Strip away the geopolitics and the trust problem, and Meta bought something real: a proven agent execution layer.
Manus is already integrated into Ads Manager, where it automates campaign analysis and report generation. The plan extends further — agentic shopping tools across Facebook and Instagram, autonomous customer service on WhatsApp Business, inventory management, order processing. Zuckerberg's phrase: "A lot of what makes agents valuable is the unique context that they can see."
Meta has 3.3 billion monthly users across its platforms. If even a fraction of them start delegating tasks to Manus-powered agents, the data flywheel becomes self-reinforcing. More context means better agents. Better agents mean more delegation. More delegation means more context.
That's the bet. It's also why the customer exodus matters. The companies leaving aren't random skeptics — they're the early adopters who proved the product works. Every one who walks away takes institutional knowledge about how autonomous agents behave in production with them.
The Bigger Picture
The Manus deal is a preview of how the agent market consolidates. The acquisition price sets a floor: $3 billion for a 78-person team with $125 million ARR and eight months of operating history. Every VC-backed agent startup in the world just repriced itself.
But the deal also exposes the fracture lines. AI agents are too valuable to ignore and too sensitive to trust casually. They touch internal documents, financial data, hiring decisions, customer communications. The question isn't whether they'll be adopted — it's who gets to own the execution layer.
Right now, that question runs through Singapore, Washington, and Beijing simultaneously. The hand Meta bought is still attached to more bodies than anyone counted.
Originally published on Substack. Follow for daily AI analysis.
Top comments (0)