Meta launched its personal AI agent Muse on September 8. The product quickly gained consumer traction—topping the US Apple App Store’s free app chart within two weeks—and triggered a sharp market reaction: Meta’s stock jumped more than 11% on September 21, lifting its market value by nearly $200 billion and producing notable moves across chipmakers and cloud infrastructure suppliers.
Meta’s long detour in AI
Meta spent years struggling to convert huge user reach and significant computing resources into market-leading AI products. By the first half of this year, Meta’s AI services held less than 5% share in the consumer (C-end) market and under 10% in the paid enterprise (B-end) segment—well behind OpenAI (46.4% and 35%) and Google (27.7% and 20%).
Meta made large annual AI investments but faced setbacks. The open-source Llama, launched in March 2023 and followed by Llama2 less than six months later, proved technically demanding and fell short of expectations. Internal tension over whether to prioritize product delivery or technology, repeated team reorganizations and talent departures compounded problems: from March 2023 to May 2025 Meta restructured its AI team four times. Meanwhile, the closed-source Muse Spark project encountered delays, internal security reviews and resource interruptions; its API release was postponed in April this year for additional safety alignment.
An early-June internal company video meeting also exposed internal strain, when an employee interrupted a colleague and complained about treatment; some executives reportedly questioned whether Muse would be “the best product of its generation.” Despite that fraught buildup, Muse was released.
Muse: usability over raw benchmarks
Muse’s benchmark scores are described as average, with clear strengths and weaknesses. It excels at straightforward, everyday tasks—sorting email, sending messages, booking travel, filling forms and placing online shopping orders—completing those tasks quickly and reliably. It is less consistent on complex, multi-step tasks such as cross-site logins and multi-stage electronic payments.
Both Muse and current GPT versions sit at what OpenAI described as “Level 3: Agents” (AI able to perform some system actions). Muse’s commercial success appears tied less to top-tier benchmark performance than to practical usability and deep integration with Meta’s ecosystem. Those simple, frequent conveniences—ordering food, managing messages, filling forms—proved more influential in consumer adoption than higher-end capabilities.
Muse’s market ripple effects
Muse’s rapid uptake produced immediate effects in public markets. On September 21 Meta’s stock rose more than 11%, the company’s largest single-day gain since April 2025, adding nearly $200 billion to its market value. Chipmakers benefited: AMD shares rose nearly 10%, pushing its market cap past $1 trillion for the first time; AMD and Meta had signed a five-year, $100 billion agreement in February this year. Although NVIDIA remains a core supplier to Meta, the AMD deal was read as a sign of shifting supplier dynamics.
Intel was an unexpected beneficiary, with its stock climbing more than 12%. Observers linked this to a changing hardware profile for Agent-style applications: answering questions relies heavily on GPU parallel compute, while agents that execute tasks increasingly depend on CPUs to call tools, orchestrate services and handle sequential logic. Storage components also gained, and the Philadelphia Semiconductor Index rose alongside the broader market reaction.
Pushback from Amazon and industry stakes
Not all players welcomed Muse’s browsing and task capabilities. On September 21 Amazon blocked Muse’s access to its e-commerce platform after failing to reach an agreement with Meta. Amazon cited three reasons: Meta did not inform Amazon in advance that Muse would access the site; the agent would not actively identify itself while browsing pages; and Muse “seems to collect and save user account credentials,” a claim Amazon said raised privacy and security risks.
Meta declined to publicly confirm the first two points and responded to the third by stressing Muse’s privacy policy and security protections. Amazon’s action follows earlier legal steps against third-party agents—last November Amazon sued the AI company Perplexity over an agent product, briefly securing access restrictions that were overturned on appeal in March this year. Amazon is also developing its own shopping agent and appears to be using access controls to buy time for its in-house efforts; the market has expressed skepticism about Amazon’s prospects in the AI sector.
The launch period also intersected with developments around Manus, which was introduced in March 2025 as “the world’s first general-purpose AI Agent” and initially attracted heavy demand for invitations. After turmoil related to a proposed Meta acquisition, Manus lost momentum; blocking that acquisition did not prevent Muse’s release. Manus announced a resumption of independent operation on September 1, 2026 but remains unable to enter the Chinese mainland market.
Competition in this phase of the AI industry is focused on applications and user-facing implementation: usability and integration can outweigh raw benchmark dominance. Muse’s early success highlights that dynamic and has shifted attention in the market toward the hardware and software patterns that best support agent-driven workflows.
Is the factor restricting the implementation of China’s AI applications only the “choke point” of computing-power blockade? The launch and reception of Muse underline that execution, ecosystem access and application design are central to who wins early in the Agent era.
