From GPU to CPU: How is the AI proxy revolution sparked by Meta (META.US) Muse reshaping the chip industry?

The Zhitong Finance App learned that the news that Meta's personal AI agent Muse surpassed 2.5 million downloads in 6 days and reached the top of Apple's US App Store free list is like a stone investing in the global AI stock pool, causing ripples — AMD broke through the trillion dollar market capitalization on Monday and staged a rare “CPU market” with Meta, Intel, and Arm. On the face of it, this is a sector carnival driven by a single popular product; in essence, it is a market repricing of the structural migration of computing power triggered by AI moving from “generation” to “action.”

Major banks such as Wedbush, Morgan Stanley, Goldman Sachs, and Jefferies have made a systematic judgment on this: as AI agents become mass popular, computing bottlenecks are shifting from GPUs to CPUs and memory, and the server CPU market will usher in tens of billions of dollars of incremental space.

Muse explosions ignite chip stock sentiment

Meta's personal AI agent Muse was launched on September 8. The number of downloads exceeded 902,000 times within 6 days of launch, surpassing the 773,000 times in the same period of the previous Meta AI generation, and quickly topped the Apple US App Store free list and continued for many days. According to Sensor Tower data, Muse has accumulated over 2.5 million downloads, ahead of ChatGPT and Claude in the ranking.

Muse was initially sought after by consumers, igniting market expectations of a surge in demand for computing power after large-scale popularization of AI devices, and a massive influx of capital into chip stocks such as AMD (AMD.US), Intel (INTC.US), and Arm (ARM.US). Catalyzed by this, AMD surged nearly 10% to 615.52 US dollars on Monday, breaking 1 trillion US dollars in market capitalization for the first time, becoming the fourth US chip company to cross this threshold after Nvidia, Broadcom, and Micron. Meta shares surged 11%, Intel surged more than 12%, Arm surged more than 17%, and the Philadelphia Semiconductor Index closed up 4.3%, the biggest one-day increase since August 4, for the fifth consecutive trading day.

Shares of South Korean chipmakers Samsung Electronics and SK Hynix rose about 3.5% in early trading, and the Taiwanese weighted index rose 1.8% to record highs.

Gary Tan, portfolio manager at Allspring Global Investments, said that the rise in the Taiwanese market “reflects growing market confidence that hyperscale cloud vendors will continue to expand the scale of their own chips as AI popularity accelerates.”

“If products like Muse get Traction, hyperscale cloud vendors will need more computing power capacity, further accelerate demand for custom AI chips, and support Taiwan's ASIC ecosystem.” — Gary Tan, Allspring Global Investments portfolio manager, this strong debut helped revive confidence in AI trading — previously the market was concerned about overvaluations and recently about the existential threat posed by advanced models. Muse's early appeal provides new evidence that demand is still strong.” The market quickly realized that this was not just a product success story — it was a repricing of the structural demand for AI computing power. The mass adoption of Muse is likely to significantly increase demand across the AI infrastructure supply chain, making chipmakers key beneficiaries of this trend.” —Dilin Wu, Pepperstone strategist

The way agents work determines the transition of the CPU's role

The focus of market excitement is not on the short-term ranking of an app, but on the fact that the AI agents represented by Muse are changing the underlying structure of computing power requirements.

The traditional chatbot's work path is “user asks questions — model generation answers — task ends”; while the agent's work path is a cycle of “the user proposes a goal — model disassembly plan — calls the browser and tools — performs an operation — obstructed adjustment — continues execution”.

In this cycle, the GPU is responsible for model inference and matrix computation, while the CPU undertakes a large number of execution layer tasks such as task orchestration, virtual machine operation, browser control, API calls, database reading and writing, and sandbox security isolation. As Fujitsu described in the Hot Chips 2026 technology presentation, orchestration, retrieval, database calls, and conditional branching are increasingly being applied to CPUs, and GPUs only process batch matrix computation steps.

The quantitative significance of this change is a shift in the CPU/GPU ratio. Traditional training servers often use 1:4 or even 1:8 to describe the relative ratio between CPU and GPU, while agent inference places more emphasis on high concurrency and tool calls. There is an opportunity for the CPU ratio to increase in the 1:1 to 1:2 direction, and some institutions even give a higher deduction range.

Big bank opinion: Structural judgments have been formed, but the conduction chain has not yet been verified

Wall Street's major investment banks are forming an increasingly clear consensus around this structural shift.

Wedbush analyst Matthew Bryson put it most directly: “Intel and AMD are the only real computing device manufacturers for AI agents.” The judgment brought the server CPU market back to the core vision of AI investment. Jefferies analyst Jacky He also pointed out that as AI agents are more widely adopted by consumers, higher inference and orchestration loads will directly benefit server CPU demand, and emphasized that this trend has profound significance for the x86 ecosystem, which has been overlooked for a long time.

Morgan Stanley's judgment is more systematic. The bank estimates that proxy AI can bring an additional 32.5 billion to 60 billion US dollars of incremental space to the data center CPU market by 2030, and the market itself has already exceeded 100 billion US dollars. Damo's core thesis is that “computational bottlenecks are migrating from GPUs to CPUs and memory”. The shift of AI from the generation stage to the autonomous action stage will bring about a structural leap in general computing strength. The research team further pointed out that when AI workloads shift from one-time inference to continuous task execution, the CPU's orchestration and coordination functions will become more strategically valuable than the GPU's original computing power.

J.P. Morgan added an opinion from the perspective of the industrial landscape, arguing that the rise of AI agents will accelerate hyperscale cloud vendors to reset the balance between customized chips and general-purpose CPUs — not only increasing investment in ASICs and custom accelerators, but also increasing demand for high-core server CPUs. The two are not an alternative but a complementary relationship.

Citigroup's team of analysts emphasized the logic of “second-round beneficiaries” in the latest report — as CPU load rises, supporting DDR5 memory, enterprise-grade SSDs, and high-bandwidth storage controllers will also be driven by demand, corresponding to storage vendors such as Micron, Samsung Electronics, and SK Hynix.

What needs to be clarified, however, is that the current market still has significant emotion-driven characteristics. Some analysts pointed out that this round of growth is based on a “transmission chain where disclosed orders have not yet been generated”. Oppenheimer analyst Jason Helfstein's estimate provides a stark reference: Meta requires about 115 million paid Muse subscribers ($20 per month) to generate annual AI agent revenue of about 27.5 billion to 28 billion US dollars, and he judged that this result was “unlikely” to be achieved due to doubts about paid conversions, fierce competition, and low consumer trust in sharing passwords with Meta.

Industrial chain spillover and “cracks” in the Meta carnival

If a structural upward shift in CPU demand is established, the beneficiaries will not be limited to CPU duos. Meta is AMD's second-largest customer, contributing about 5.5% of its revenue. The two sides further expanded their cooperation in February this year and plan to deploy up to 6 gigawatts of AMD Instinct GPUs. Arm also occupies a key position in data center CPU architectures in the agent era. Looking deeper, ASIC custom chips, optical communication, memory, etc. will also benefit from changes in the computing power structure.

But Meta's own financial situation is bearing the cost of AI investment. In the second quarter of 2026, Meta's revenue of US$60.8 billion increased 28% year over year, but free cash flow plummeted from US$8.55 billion in the same period last year to US$784 million, a sharp drop of 91% year over year; operating profit margin was reduced from 43% to 31%, and the annual capital expenditure guidance reached 130 billion to 145 billion US dollars.

Furthermore, Amazon has blocked Muse from accessing retail websites on the grounds that it has not identified itself, is suspected of obtaining user credentials, and scraping account data. This reveals that the game surrounding control of user relationships between platforms in the Agent era is escalating.

The Meta Connect conference, which opens this Wednesday, will be the first node to test the sustainability of this round of “CPU markets.” If Muse's user growth and commercialization path can be further verified at the conference, the market's repricing of the AI inference computing power structure will gain a more solid anchor; conversely, the current valuation expansion driven by a single product catalyst may take longer to digest.