Muse Is America's Most Downloaded App, Meta Rose 11 Per Cent, and AMD Passed a Trillion Because Agents Need CPUs

A week ago Mark Zuckerberg was explaining why Meta would not join the industry slowdown, and offering the months it had held back its Muse agent as proof it could be trusted to pace itself. On Monday the market decided the delay had been worth it. Muse, the personal agent Meta released on 8 September, reached the top of Apple's free app chart in the United States in under two weeks, ahead of ChatGPT, and topped Google Play as well; the trade press counted more than 900,000 downloads in its first six days. Meta shares closed at 741.25 dollars, up 11.43 per cent, the biggest one-day rise since April 2025 and, Yahoo Finance notes, roughly 190 billion dollars of market value in a session. Wells Fargo's Ken Gawrelski lifted his target from 640 to 796 dollars and wrote that Meta "now has a story to tell".

What Muse is

An agent rather than a chatbot: it is given a task, a purchase, a booking, a piece of research, a message thread to manage, and it works through the steps, using the apps and services on the phone and Meta's own, and reports back. The Muse family includes a lighter "Muse Spark" model. It sits on top of the social graph and the messaging apps, which is Meta's structural advantage over OpenAI: it already knows who you talk to and what you buy. The catch, which the summer's incidents made vivid, is that an agent acting on your behalf can act wrongly on your behalf, and Zuckerberg's claim that labs have "a strong natural incentive to make their models more aligned" because "people won't want to use agents that are misaligned with them" is about to be tested by a great many people at once. Meta Connect on Wednesday, where Zuckerberg speaks on AI and wearables, is where the company will say what Muse does next and, presumably, how it does it on glasses.

Why the chipmakers moved

The second-order trade was the interesting one. Advanced Micro Devices rose 10 per cent and closed above a trillion dollars in market value for the first time. Intel gained 12 per cent, Arm 17, and the Philadelphia Semiconductor Index 4.3 per cent, its fifth consecutive rise. The logic, which Bloomberg's account of the day set out, is that agentic workloads are different from training and from chat. Training a model is massively parallel matrix arithmetic, which is what a GPU is for and why Nvidia is worth what it is worth. Running an agent is a long chain of sequential steps, each depending on the last: call a tool, read the result, decide, call another. That is CPU work, and the more agents there are, the more general-purpose processors the data centres need. AMD sells both; Intel and Arm sell the CPU side. The market had spent two years assuming the AI build-out was a GPU story. On Monday it priced in the possibility that the next phase is not.

The caveats

App-chart position measures curiosity, not retention; Threads topped the same chart in 2023 and lost most of its users within a month. Nobody outside Meta knows what Muse costs to run per user, and agents are expensive because they call the model many times per task. The antitrust suit filed against Anthropic, OpenAI, Google and SpaceXAI on Friday, which we report separately, notably does not name Meta, which is now the one large lab that said it would keep going and is being rewarded for it. And a 190 billion dollar move on a fortnight of downloads is the kind of thing that reverses on a bad Connect. But the shape of the argument has changed. A month ago the question was whether Meta had an AI story at all. Now it has the most used agent in the country and a chip rally named after it.