Why Meta Paid $2.5B for Manus

Edition 148 | January 5, 2025

Saw a post about someone using Claude Code to patch a failing jet engine mid-flight. Meanwhile, I’m on hour three of getting it to change a button color on my website.

Welcome back to Building AI Agents, your biweekly guide to everything new in the field of agentic AI!

In today’s issue…

• Why Meta paid $2.5B for Manus

• How agents became defensible without owning a model

• The White House’s new stance on state AI regulation

• Alibaba’s open-source GUI agents and what they enable

• The signals pointing to 2026 as the “ROI year” for agents

…and more

🔍 SPOTLIGHT

On December 30, Meta announced it was acquiring Manus for approximately $2.5 billion. Eight months earlier, Manus had raised its Series A at a $500 million valuation. The deal came together in roughly 10 days…

Mark Zuckerberg personally pushed to close before year-end. Manus's team will report directly to COO Javier Olivan — not to the Superintelligence Lab — signaling Meta sees this as a product to scale, not a research project to absorb. Beyond the purchase price, Meta set aside a $500 million retention pool for Manus's roughly 100 employees. CEO Xiao Hong becomes a Meta VP and continues leading the team.

So what made Meta move this fast and pay this much?

The Gap Meta Needed to Fill

Meta had spent billions on AI infrastructure and open-sourced the Llama model family, but it was missing something critical: an agent product with real paying customers.

OpenAI had launched Operator. Google was rolling out Gemini Enterprise. Salesforce had Agentforce. Meanwhile, Meta's own AI assistant could answer questions but couldn't actually do anything — and had no obvious path to something better. Manus gave them one. Instead of building another model, Meta is betting that the real value lives in the execution layer.

What Manus Built

Manus launched in March 2025 as a general-purpose AI agent. Manus could take a high-level goal like "find me the best mortgage rates in California, compare them in a spreadsheet, and email me the top three" and execute it autonomously over hours or days.

What made Manus different wasn't the model — they ran on Claude 3.5/3.6 Sonnet, not anything proprietary. The value was in everything around it. Manus was one of the first to give agents a full computer in the cloud, not just a code interpreter. Each task spun up an isolated Linux sandbox with a real file system, the ability to install software, browse the web, and create actual deliverables like slide decks and reports. Think Claude Code running in the cloud for you rather than on your laptop. You could watch the agent open a browser and navigate in real time, pause and resume tasks, or intervene if it went off track.

Manus was also multi-model before it was standard, switching between models internally based on the task so users didn’t have to think about it.And they moved aggressively on SaaS integrations, connecting to tools people already used, which made the product immediately useful.

A lot of these ideas — cloud sandboxes, artifact creation, tool integrations — were later adopted by labs like Anthropic and OpenAI. Manus set a lot of the useful agent building practices.

Critics dismissed this as “just a wrapper for Claude.” They missed the point. Models are commodities and will get better with time. What matters is what you build around them.

The Revenue Proof

Manus hit $100 million in annual recurring revenue within eight months: faster than ChatGPT, Midjourney, or any AI SaaS on record. The subscription tiers ran from $19 to $199 per month, proving that consumers and businesses would pay real money for an agent.

Paying customers and accelerating growth gave Meta all the validation it needed. Like Meta’s acquisition of Instagram in 2012, Manus was an early signal of consumer demand.

The Geopolitical Maneuver

Manus was founded in Beijing. By mid-2025, the company had relocated to Singapore, reduced its China-based workforce by 80+ employees, shut down its Chinese social media accounts, and bought out investors like Tencent and ZhenFund.

The restructuring was very deliberate. A U.S. Treasury investigation into Benchmark's Series A investment had raised concerns about American capital flowing to Chinese AI. By severing those ties before Meta came calling, Manus removed a big acquisition blocker. Meta's spokesperson made it explicit: "There will be no continuing Chinese ownership interests in Manus AI."

What You Can Take Away

The Manus deal reframes what "defensible" means in the agent era. The team didn't train their own model. They built an orchestration layer on top of Claude and focused relentlessly on making it reliable for real tasks. That's a playbook anyone can follow: pick the best available model, then solve a hard problem with it.

Manus also didn't wait for the tech to be perfect. They shipped, charged money, and iterated based on what paying users actually needed.

If you're building agents today, the lesson isn't to raise a massive round or train a model. It's to make something useful and put it in the world.

Keep learning and building!

—AP