I'm LongbridgeAI, I can summarize articles.Meta Platforms' AI agent Muse saw its daily active users (DAU) grow tenfold to approximately 700,000 in 11 days, but service issues have begun to surface: actual resource consumption is about 10 times higher than internal tests, SaaSHub has flagged it as "degraded," and only 33 out of 120 sub-agents succeeded in a stress test. Muse adopts an independent virtual machine architecture, meaning user growth will continue to drive up CPU, memory, and storage demands, bringing compute scaling pressures to the fore
Meta Platforms' personal AI agent Muse has seen rapid user growth, but server-side pressure has emerged ahead of schedule. Over the past 11 days, Muse's daily active users (DAU) grew approximately tenfold to 700,000, but issues such as service degradation and Agent task failures have also begun to appear, with compute bottlenecks gradually coming to light.
Muse has recently experienced several service anomalies, including resource consumption far exceeding internal test expectations, degradation of certain features, and frequent failures when Agents execute complex tasks. In one stress test, Muse attempted to create 120 sub-Agents simultaneously but succeeded with only 33, indicating that its actual operational load is already straining service stability.
According to data from web analytics firm Similarweb, Muse currently has approximately 700,000 DAU. Notably, this growth has not yet reached all potential users: Muse is currently only available in the United States and Canada, and registration portals on WhatsApp and Instagram have not yet been fully opened. If the release scope is further expanded in the future, the computational load on the server side will continue to rise.
The core issue with Muse lies in its operational architecture, which consumes high computing resources. Meta Platforms allocates an independent cloud virtual machine for each user. As the user base expands, demand for CPU, memory, and storage will grow in tandem, making infrastructure scaling a problem that Muse must confront to achieve scale.
Before Breaking One Million DAU, Muse Already Shows Frequent Service Anomalies
There are already multiple concrete signs of service pressure on Muse. On September 9, Meta Platforms' Chief AI Officer Alexander Wang revealed that early actual usage of Muse "far exceeded expectations," with user consumption about 10 times that of the internal test group. Meta Platforms subsequently adjusted its token usage policy and increased quotas.
Meanwhile, the third-party status monitoring platform SaaSHub currently marks Muse as "Degraded," meaning the service is still accessible, but some users continue to report issues, with "unable to search" being the most common feedback.
Agent execution capabilities are also being tested by concurrent loads. An AI researcher recently conducted a stress test, asking Muse to create 120 sub-Agents simultaneously; only 33 succeeded, while the other 87 creation attempts failed. The test showed that under high-concurrency tasks, Muse's Agent orchestration capabilities have certain limitations.
These phenomena do not directly prove that all problems are caused by insufficient compute power, but they reflect that Muse's actual operational load is significantly higher than early test expectations.
Independent Virtual Machine Architecture Drives Up Compute Demand
The compute pressure on Muse largely stems from its underlying architecture. To reduce privacy and security risks, Meta Platforms assigns each user an isolated cloud virtual machine, configured with 2 virtual CPUs, 8GB of memory, and a 100GB solid-state drive, along with a dedicated browser for storing user data and credentials.
Based on the above configuration, in an ideal scenario without load sharing, if the user base reaches 100 million, Meta Platforms would need approximately 200 million CPU cores, equivalent to about 1.58 million AMD EPYC processors with 126-core configurations. It would also require 800PB of memory and 10,000PB of solid-state storage.
Muse is also equipped with a security mechanism called "Sentinel," which continuously monitors the agent's operating environment. Before an agent completes a transaction, it requires approval from Sentinel and explicit confirmation from the user. This multi-layered security mechanism also increases the computational overhead of system operations.
User Expansion and Commercialization Proceed in Tandem
Muse is positioned as a personal AI agent capable of executing multi-step tasks. It can operate cloud browsers on behalf of users, browse web pages, and complete related transactions, continuing to run in the background even if the user closes the application.
Meta Platforms is further expanding Muse's interaction entry points. In addition to Meta smart glasses, the company is developing a standalone keychain-shaped device called Muse Charm, expected to ship in December this year. WhatsApp and Instagram have not yet opened Muse registration, leaving room for further expansion of user coverage in the future.
Commercialization has also begun to roll out. Zuckerberg stated this week that Meta Platforms will charge a small commission on transactions facilitated by Muse. Currently, Muse provides users with up to 100 million free tokens per week, with paid subscriptions for heavy users starting at approximately $20 per month.
As Muse continues to expand from its current scale to a larger user base, Meta Platforms needs to address not only the demand brought by new users but also how to maintain Agent execution capabilities and service stability. The independent virtual machine architecture means that CPU, memory, and storage demands will expand in sync with the user scale, which will also place higher demands on Meta Platforms' data center investments and infrastructure procurement.