Has the revenue of the returned Manus quintupled?
On the night of August 11.
The line "Manus has been part of Meta" that hung on Manus' official website for nearly half a year was replaced with a new line of text: "Manus is about to resume independent operation."
Seven and a half months, a $2 billion acquisition, from official announcement to cancellation, completed a full round trip. It's back, but it's not the same company that returned.
In December 2025, Mark Zuckerberg flew to Singapore in person. It only took 10 days from the first contact between the two parties to the signing of the contract.
At a meeting attended by all members of Manus' core team, Meta showed its acquisition intention in front of a full room of people; the $2 billion deal was the third largest acquisition in Meta's history, second only to WhatsApp and Scale AI.
It was so fast that it didn't look like negotiating a business deal, it looked like poaching talents, and Xiao Hong agreed.
He moved his team to Singapore only half a year ago, downsized the 120-person domestic team to more than 40 people, blocked Chinese IPs on the official website, and cleared all Chinese social accounts.
After getting rid of all these things, the rest was packaged and sold to Silicon Valley, which gave an account to everyone. For partners who had worked with him for many years and core technical staff who had experienced two startups, selling the company was the return.
On New Year's Day 2026, he walked along the seaside in Singapore, feeling that his ten-year entrepreneurial journey had finally reached the finish line, and he would go to work at a large company next.
He didn't know that this was only the first half. After Meta got Manus, it moved at a surprisingly fast speed. In less than two months, Manus was integrated into AdsManager.
More than 10 million advertisers can directly use natural language to let Manus do competitor analysis, anomaly detection, and automatic reporting.
In the past, these tasks required hiring people to do, but now you only need to type a line of text. Then the capability was pushed to WhatsApp Business and Instagram. An independent product became a pipeline component of an advertising empire in less than seven weeks.
To be fair, the numbers don't lie:
Manus' daily revenue rose from $300,000 to nearly $1.5 million, a fivefold increase; its ARR surged from $100 million at the time of acquisition to more than $400 million, once surpassing DeepSeek.
Yes, you read that right, a company that was under regulatory investigation and was said by the outside world to be "going to collapse" saw its revenue quintuple.
Investors are the most honest. When the ban was first issued, all the venture capital institutions that used to squeeze in from any gap to grab shares fell completely silent.
Three months later, they started calling again to ask if there were still available shares. One investor said that if there was still a funding gap, they would very much like to participate.
The money part is done. The following are the things that cannot be deleted.
After the acquisition, Manus' product architecture began to integrate into Meta's ecosystem, with connectors connected to GMail, Shopify, and Slack, and the engineering rhythm was completely synchronized with Meta.
With one or two feature updates every week, the rhythm was as fast as the peak at the end of 2025. The 2.0 version had long been developed but was suppressed and not released, waiting for the transaction to be completely closed before being launched.
The team expanded from 105 to 150 people. The identity on the official website was "Manus is part of Meta". Xiao Hong's LinkedIn profile stated that he was Vice President of Meta.
You are building roads for the cities of an empire, and these things are not on the books, invisible.
150 people write code in Meta's training system every day, run Agents in Meta's pipelines, and interact with Meta's data architecture. Seven months is enough, the muscle memory has been fully formed.
No single line of code can delete these things, nor can a share repurchase buy them back.
However, on that New Year's Day when Xiao Hong was walking by the sea, he did not know these things. Seven months later, someone was going to extract all of them.
......
On April 27, the Office of the Working Mechanism for Foreign Investment Security Review issued an announcement. The wording was very brief:
In accordance with laws and regulations, a decision to prohibit investment was made for the foreign capital acquisition of the Manus project, requiring the parties to cancel the acquisition transaction.
Since the implementation of the "Foreign Investment Security Review Measures" in 2021, this is the first publicly stopped foreign acquisition case in the AI field, which belongs to the strictest category of "prohibited investment".
These four words determined everything that followed: restore the original state.
What does restoring the original state mean? Legally, it means returning to the state before the transaction occurred; Meta withdraws its shares, and Manus is restored to its original state. It sounds very clean, just like a return of goods.
Everyone has returned goods before, try returning something that has been opened.
The first thing to return is data.
From August 23 to 24, there is a 48-hour window. All data generated by affected users on the day of Meta's acquisition, that is, December 29, 2025 and thereafter, will be physically deleted.
Manus released a free backup tool, and even gave a return reward after the deletion was completed, which is more decent than most divorce agreements.
Decent as it is, a batch of users' seven-month work records, conversation history, and personalized data for Agent training will no longer be found on the server.
I have seen people in the Manus community on Reddit asking: Should websites hosted on the platform be migrated to GitHub? What about the annual subscription? How to download large-scale backups?
Some users also reported that the backup entry reported an error, and it took a lot of effort to restore it.
The second thing is channels.
In early June, Manus employees suddenly found that they could not log in to Meta's internal system, the entry to AdsManager was closed, the pipelines to WhatsApp and Instagram were cut off, and those 10 million advertisers had nothing to do with them overnight.
Meta moved faster than the ban. Two days after the ban was issued on April 27, Meta released Ads AI Connectors on April 29, integrating Claude and ChatGPT into the advertising system.
On June 3, it officially announced Meta Business Agent, covering WhatsApp, Messenger and Instagram. You can tell how reluctant Meta is to part with Manus by how quickly it arranged the alternatives.
The third thing is equity.
A Chinese consortium led by Tencent took over the stake at a valuation of $2 billion, unchanged from before. ZhenFund and HSG joined the follow-on investment, Benchmark exited, and Tencent became the largest shareholder without holding a controlling stake.
Manus will continue to operate independently in Singapore, paving the way for a Hong Kong stock IPO in the future.
Pay attention to one detail: Meta did not collect any premium. The price when it acquired is the same as the price when it sold back. It bought in at $2 billion and sold back at $2 billion, transferring at the original price.
Did it lose money? The only loss is the interest. What it gained is that Manus' Agent capabilities have been absorbed into its own product system; Ads AI Connectors and Meta Business Agent, these alternatives were trained by Manus step by step.
In this spin-off, equity is the easiest part to restore, just a matter of signing a document.
At this point, all that can be returned has been returned, the data has been cleared, the channels have been cut off, and the equity has been transferred. Presumably the restoration of the original state has been completed, and no one can find fault with it legally.
Is that all? Of course not, the real difficulties have only just begun.
After the ban was issued, none of the 150 people left their jobs. During the months of turmoil, teams in Beijing and Singapore held video calls every day to synchronize progress and arrange work.
Rumors were flying everywhere. One day it was said that Tencent would take controlling stake, the next day it was said that the founder had run away, but none of them looked for new jobs.
The product development did not stop either. On July 9, the conversation branching function was launched, on the 13th the automatic website publishing function went online, on the 14th the intelligent PPT generation function was released, and on the 22nd the Plan Mode that plans before execution was launched.
To be fair, for a company that was sentenced to death by the regulator, its internal order did not collapse. Outsiders said "the champagne was opened too early" and "2 billion dollars vanished overnight", while people inside were working quietly.
Why? To put it nicely, it is team cohesion. To be realistic, there are two accounts to calculate.
One is the skill account. For more than seven months, they ran Agents at that high intensity every day, processed the traffic of hundreds of millions of pipelines, and interacted with the world's largest advertising system.
This kind of experience cannot be bought outside. You can only get it after being trained by Meta, and there is no second place that can be used for the time being after leaving the Meta ecosystem.
The other is the money account. Tencent's repurchase is under negotiation, and the equity structure is being reorganized. Leaving at this time means giving up the share that may be realized in the future.
The turmoil during the ban is temporary, and the options after the repurchase are real. They stay and wait for the dust to settle, betting on this, this is the restoration of the original state, the contract can be voided, but the people cannot be reverted to their previous state.
......
Well, the people cannot be reverted, and then what? This body that has been trained by Meta for seven months, how will it play the next game?
First look at the track.
When Manus was acquired in December 2025, the general Agent was still a blue ocean. When it came back to operate independently, the table was already full of players.
OpenAI's Operator is running, Anthropic's Claude Cowork is running, and Google has integrated Mariner into Gemini Agent.
Elon Musk's Grok Bot was just released yesterday, positioned as an AI intelligent agent team that can independently log in to various applications ......
On the domestic side, Zhipu's AutoGLM, ByteDance's Coze, Doubao Office, Alibaba's Qwen, and Baidu's Dazi are all seizing the desktop market; Claude Code will enable Auto Mode by default next week, even saving the need for manual confirmation.
To be fair, when Manus was first released, the industry called it "the world's first general AI Agent". No one mentions this label now, because everyone has their own.
The Fresh Research Team of Tsinghua University released a report that divides the Agent track into three layers:
The product layer is dominated by three players, Manus, Genspark and Flowith each occupy a part. At the base layer, large manufacturers step back to the background to act as capability suppliers. At the vertical layer, players like Devin focus on a single scenario.
Manus is stuck in the product layer, closest to users, and also closest to the capability boundary of large manufacturers. If either side takes a step forward, it will be the first to be hit.
After more than seven months of not running independently on this track, when it returns to the starting line, it has gained some things and lost some things.
What did it gain? It got an extra card in hand.
150 people have been trained, and have processed billions of levels of advertising traffic. There is no second Chinese team with this kind of experience in the Agent track, and there is still an unreleased card in their hands.
Xiao Hong once said a word, he summed up his path with "trade, manufacturing, then technology": first find the market, then build the business, and finally go deep into the technology.
He has completed the first two steps, built products that millions of people pay for, and established a business with an ARR of hundreds of millions. The third step has never been touched.
What about the things it lost?
Manus does not build underlying models, it runs on Anthropic's Claude. Dario Amodei, the founder of Anthropic, once said a sentence:
"General large models are a powerful platform, and also a powerful competitor."
This sentence is for Manus. Every time Claude raises prices, Manus will feel the pain; if Claude enters the Agent field on its own, Manus will be the piece of meat it wants to eat.
The Meta channels have been cut off. No one can clearly separate how much of the more than 400 million ARR is organic growth and how much is poured in through Meta's advertising pipelines.
Tencent calculated two accounts when it made the acquisition:
In the optimistic scenario, assuming that the endogenous ARR can hold up at 200 million; in the pessimistic scenario, it will be cut in half and then cut in half again, falling back below 100 million. For a company valued at 2 billion dollars, the gap between these two lines is a matter of life and death.
The data has been deleted once. Although it is normal compliance, the user relationship is equivalent to a cold restart. Benchmark has exited, and the attitude of US VCs towards this company is different from when they led the Series B round in April 2025.
The Hong Kong stock IPO path has been officially mentioned, the premise is to complete all restructuring; no one can give a timetable for when it will be completed. These are all obvious at a glance.
Finally, let's talk about the choices.
Xiao Hong is re-learning mathematics in Beijing, researching large models. A founder who never touched underlying technology chose to make up for mathematics in the most turbulent time.
What is he going to do? He may not have figured it out himself. However, the direction is clear:
Either continue to be the most profitable Agent company in the product layer, survive by relying on the muscle memory of the first two steps of "trade, manufacturing, then technology"; or go down to touch the model; or turn around and sell the skills to the buyer who needs them most.
The third path is rarely mentioned, but it is the smoothest. Manus was trained by Meta for seven months, and the most valuable thing is its experience of running Agents in the 100-million-level advertising pipeline.
Tencent has WeChat, and the largest advertising inventory in the entire Chinese Internet. An Agent team that has been trained by Meta's advertising system, plus Tencent who wants to integrate AI into its advertising empire, what will happen when these two things meet?
Manus does not need to chase models or compete in the product layer, it can directly turn itself into the advertising Agent infrastructure in the Tencent ecosystem, and this path does not require Xiao Hong to learn mathematics.
On the night of August 11, Xiao Hong wrote a letter to the team and investors, with only one sentence in the title: "This is my third trip."
The first trip was starting a business. The second was being acquired, and the third is starting all over again; he didn't say where he was going this time, he only said: I don't want to watch this era from the sidelines, and I don't want to write an autobiography to recall the past. I still want to keep running.
Well, This is my third trip.
Notes:
[1]. The key data in this article comes from Manus official announcements, Financial Times, Sacra, PPC Land and public documents of the National Development and Reform Commission
