Update September 2026: In April 2026, China's National Development and Reform Commission ordered Meta to unwind the Manus acquisition. Meta separated the two businesses in June, and Manus has said it will operate independently again. In September, Meta launched Muse, a persistent agent it built in-house. The race for the intent layer this piece describes has continued without Manus.
While the industry wound down for the year, Meta announced on Monday (29 December 2025) its acquisition of Manus, the Singapore-based AI agent company for a reported $2 billion. In Meta’s words, the deal is about “accelerating AI innovation for businesses“ bringing Manus’s team and technology into Meta’s ecosystem to “help businesses get more done.” The deal was surgically structured to eliminate Chinese ownership (Tencent, ZhenFund, and HSG were fully bought out), with the 100-person team reporting directly to Meta COO Javier Olivan. The price represents a 4x premium over Manus’s $500 million valuation from just eight months ago.
The deal comes just a couple of days after Nvidia’s $20 billion non-acquisition acquisition of inference chip specialist Groq, and reports that it is in talks to purchase AI startup AI21 Labs for as much as $3 billion. Manus is fundamentally different from both, but it’s emblematic of how big players are racing to consolidate their position as they try to understand where the new moats will be built in the Agentic Era.
In this case, Meta isn’t just buying model capabilities or research talent. They’re buying an operational orchestration platform with $125 million in annual recurring revenue from enterprises paying for agentic workflows. These are not pilots, not research projects, but production deployments generating real cash flows. As I documented in my earlier analysis of Manus, the company had already processed 147 trillion tokens across 80 million virtual machine deployments, in one of the earliest agentic AI deployments.
As we step into the new year, the “AI Trade” is entering a phase of maturation. We are moving past the era of infrastructure speculation, where the focus was primarily on who owns the chips and the models, into a period defined by value capture resulting from the AI build-out. The defining question of 2026 is no longer only about which model is smartest, but rather: Who captures what this infrastructure enables, and how is that value defended? I called this Orchestration Economics.
To navigate this shift, I am excited to introduce a new short-form feature for paid subscribers titled “Orchestration Economics.” I’m offering this one as a special preview. These pieces will be published between our deeper weekly publications, providing tactical analysis on the AI trade.
As always, I welcome your feedback as we build this framework together.
Understanding why Meta might want to acquire Manus offers an important case study into the emerging Orchestration Economics.
Why Did Meta Buy Manus?
With Manus, Meta is taking a step toward acquiring a key orchestration feature it doesn’t have and desperately needs: Intent, a place where users express what they want to accomplish rather than what they want to see. When Meta buys Manus, they’re attempting to capture an entirely different position in the value chain: orchestration revenue.
So much of the focus around competition remains on model power. But as we’ve seen in recent months, the difference between frontier models is becoming negligible, almost a commodity.
The real moats will be created around the orchestration layer, the ability to direct multi-agent systems, and capture the value as it migrates to this layer over time. The opportunity here will be massive.
So, what does Meta need to succeed here?
Here’s what Meta has today:
AI capability with Llama
Compute
Attention through Instagram and Facebook. Meta owns communication through WhatsApp and Messenger
Valuable assets for sure. But in the agentic future, these are also elements that could be left open to be orchestrated by others who might capture more value down the line.
The deal is both a testament to Manus’ early success and an admission that Meta has a strategic hole in its $1.7 trillion empire.
As the agentic economy materializes, part of the new paradigm that will emerge will include orchestration revenue and fees for completed workflows. That revenue will dwarf the infrastructure costs underneath. A financial analysis workflow might cost $10 in compute but generate $100+ in orchestration fees because it replaces hours of analyst time.
Meta is betting $2 billion that they can capture those orchestration margins through WhatsApp Business, even though they’re late to the intent capture race and face formidable competition from ChatGPT, Claude, Salesforce Agentforce, and Microsoft Copilot.
Let’s look more closely at the role intent plays in Orchestration Economics.
The Intent Problem
When you open Instagram, Meta knows what you want: entertainment, connection, distraction. The algorithm serves content. You scroll. Meta sells ads against your attention. When you open WhatsApp, Meta knows what you want: to message someone. You type. They reply. Meta sells ads (theoretically, since WhatsApp monetization remains mostly aspirational).
But when you want to accomplish something, such as “analyze our Q4 sales pipeline,” “plan the product launch,” “research competitor pricing strategies,” you don’t open a Meta product. You open ChatGPT. Or Claude. Or Gemini. Or increasingly, you talk to an AI agent embedded in your CRM, your project management tool, or your data warehouse.
This is the intent capture problem, and it’s existential in the agentic era. The architecture of value is reorganizing around a simple principle: whoever captures user goals at their point of origin controls which agents execute those goals and how value flows.
Where Does Value Migrate When Agents Take Over?
The system where intent originates becomes the orchestration layer. It decides which agents get called, in what sequence, with what context, under what constraints, and how results are synthesized. The underlying systems, such as your CRM, your data warehouse, your email, become API calls. They’re valuable, but they no longer control the workflow. The orchestrator does.
Traditional software made users come to it. You opened Salesforce to update a deal. You opened Tableau to build a dashboard. You opened Jira to file a ticket. Each application was a destination, a discrete stop in your workflow where you manually translated intention into action.
Agents invert this model entirely. You express a goal in natural language: “Close the Q4 pipeline.” Agents decompose that intent into a plan, route tasks to appropriate systems, execute the workflow, and synthesize results. You never leave the interface where you expressed the goal.
That’s why, a few weeks ago, I surprised many readers by claiming that Salesforce’s move to block access to Slack was “strategic genius”. My goal, then, was not to acclaim Salesforce, but to highlight that it had well understood one of the fundamental laws of value in the Agentic Era: proximity to user intent.
Meta is demonstrating that now.
What Did Meta Buy From Manus?
Manus isn’t a chatbot. It’s an orchestration platform that coordinates multi-step workflows across systems based on natural language instructions. Examples of tasks Manus agents complete include resume screening (parsing hundreds of applications, extracting key qualifications, ranking candidates against job requirements, flagging top prospects), trip planning (researching destinations, comparing flights and hotels, creating detailed itineraries with reservations and backup options), and investment analysis (pulling financial data, analyzing trends, generating reports with specific insights on risks and opportunities).
Manus had $125 million in annual recurring revenue - real money from businesses actually paying for agent orchestration, not vaporware or pilot programs. The company is still young (founded recently enough that it’s in early scaling mode), but the revenue proves enterprises will pay subscription fees for coordinated AI workflows that produce tangible business outcomes.
Perhaps the most critical technical detail is that Manus is model-agnostic. While early iterations of Manus leaned heavily on Anthropic’s Claude 3.5/3.7 for reasoning, its architecture is designed as a universal orchestration wrapper. Manus uses a “CodeAct” paradigm that turns natural language into executable Python within a secure sandbox to coordinate tasks. This means the system can treat foundation models (GPT, Claude, or Llama) as interchangeable “engines” while Manus remains the “transmission” that converts that power into real-world action.
For Meta, this is the ultimate strategic hedge:
Llama Integration: Meta can immediately swap out high-cost proprietary APIs for their own fine-tuned Llama 4/5 models, instantly improving margins on Manus’s $125M ARR.
Future-Proofing: Meta is no longer tethered to a single model’s performance. If a specialized open-source model emerges that is superior for “financial analysis” or “web-scraping,” Manus can integrate it without re-architecting the entire platform.