←2026-09-27 · VM Genius
分析
Muse is the first consumer-facing AI Agent capable of completing fully autonomous, closed-loop tasks in the background: sending emails, booking travel, canceling subscriptions, filling out forms, and executing price comparisons and purchases. The process runs uninterrupted even after the user closes the app.
Following its September 8 launch, Muse broke historical records:
2.8 million downloads across the US and Canada in 12 days—more than double ChatGPT’s performance during the same timeframe.
Over 500,000 users in its first week, hitting 264,000 single-day US downloads to top the App Store.
10x the traffic of internal test estimates, prompting Meta’s Chief AI Officer Alexandr Wang to admit performance vastly exceeded expectations.
In terms of monetization, Mark Zuckerberg has explicitly adopted a transaction-fee model: Meta gets paid only when the job gets done—a fundamental departure from SaaS subscriptions and digital advertising.
The most disruptive shift lies in its compute architecture. GPUs are confined to Meta’s data centers solely for "brainwork" (task decomposition and path planning), while CPUs assume full control of "execution" (opening browsers, logging into accounts, and submitting transactions). Real-world user benchmarks reveal that Muse's dedicated cloud instances run without discrete GPUs, executing the entire workflow on just 2 vCPUs alongside standard RAM and storage.
Meta boasts 3.6 billion daily active users worldwide; a limited beta test alone was enough to overwhelm compute infrastructure. The key question is no longer whether agents require compute, but rather how much CPU capacity they will consume.
While AI investments over the past three years focused exclusively on GPUs, Muse elevates the CPU from a supporting player to center stage. The training era prioritized GPUs; the agent era prioritizes CPUs. Compute dynamics are officially inverting.
1. CPU-to-GPU Ratios Shift from 1:8 to 1:1
Compute allocation varies drastically across workloads:
AI Training: 1 CPU to 7–8 GPUs (CPU functions purely as a helper component).
Standard Inference: 1 CPU to 1–3 GPUs (CPU importance increases).
AI Agents: Ratios flatten to 1:1, reaching up to 4:1 in complex execution environments.
As AMD CEO Lisa Su highlighted during an earnings call, compute ratios across the industry are rapidly converging toward 1:1. Servers running AI agents are no longer high-density GPU training clusters; they resemble traditional operational application servers where high-performance CPUs become indispensable core assets.
2. Supply-Demand Mismatch: Potential Demand Exceeds Capacity by 20x
A baseline projection illustrates the bottleneck: assuming 100 million US knowledge workers deployed with 1 personal Agent each, where every 10 Agents require 1 dedicated server CPU, demand hits 10 million chips for the US alone. Globally, 1 billion professionals deploying 10 Agents each scales demand to 1 billion CPUs. Against an annual global server CPU output of just 3.5 to 4.5 million units, the structural deficit exceeds 20x.
Manifestations of this demand wave are already evident:
Intel stated it can currently satisfy only half of customer CPU orders.
Anthropic signed a $11.6 billion, 7-year CPU compute agreement with Akamai, signaling a pivot in hyperscaler procurement toward CPUs.
AMD reported a 70% year-over-year surge in server CPU revenue last quarter.
Industry Consensus: By 2030, inference compute will surpass training to become the single largest workload in data centers, expanding at a 35% Compound Annual Growth Rate (CAGR).
3. The Consolidated Triopoly and a Fourfold Market Expansion
The global data center CPU market remains divided among three dominant architectures:
x86 Ecosystem: Intel faces severe supply constraints. AMD has expanded its market share toward 50%, with its market cap expanding from $14 billion in 2017 to over $1 trillion—a testament to this structural inflection.
Arm Ecosystem: The fastest-growing architecture segment. NVIDIA Vera, alongside proprietary silicon from AWS, Google, and Microsoft, relies on Arm architecture. As Arm enters direct chip sales, Meta has emerged as its primary customer.
AMD has revised its 2030 addressable market outlook upward three times in under a year: from $60 billion to $120 billion, and ultimately to $220 billion. Compared to a $26 billion market size in 2025, this represents a 4x expansion driven almost entirely by AI Agents. Arm similarly projects that per-watt CPU capacity demands in the agent era will exceed 4x current baseline levels.
Initial market reactions framed Muse as a net negative for Amazon, but that thesis is incomplete. Consumer agents represent a tailwind for hyperscalers like AWS. Short-term CapEx burdens are offset by long-term earnings visibility, unlocking up to $65 billion in equity upside from current valuation discounts.
1. Agents Demand Full-Stack Infrastructure
Agent workloads require unified full-stack resources: continuous inference loops require CPU power, expansive context windows demand high-density RAM, and high-frequency execution requires ultra-low latency, high-bandwidth networks. Cisco telemetry indicates that an AI Agent performing a task generates 450% of the network traffic produced by a human user. Deloitte projects inference compute will account for two-thirds of total workloads by 2026.
AWS is positioned across these vector demands:
Proprietary Silicon: Graviton CPUs deliver 30%–40% superior price-performance compared to x86 peers. Deployment of the Graviton5 generation has doubled in speed, with adoption across the top 1,000 EC2 customers driving AWS custom hardware revenue up nearly 100% quarter-over-quarter.
Heterogeneous Clusters: Pairing custom Trainium processors with NVIDIA GPUs enables workload optimization across diverse compute architectures.
2. Monetization Across the Stack: From Compute to Governance
Enterprise agent deployment requires development frameworks, orchestration, security, and lifecycle management. Beyond fundamental compute layers like SageMaker and Bedrock, AWS offers dedicated agent infrastructure including AgentCore, Quick, Transform, and Continuum, removing the need for enterprise customers to assemble custom technology stacks.
Performance metrics confirm this trajectory: AWS reported $496 billion in backlog (up 100% year-over-year), with Annual Run-Rate (ARR) for both its AI and custom silicon businesses exceeding $25 billion. Compute capacity is set to double by late 2027 relative to 2025 levels, with most 2027 and substantial 2028 capacity already pre-committed by enterprise clients.
The rise of Muse forced a strategic split across e-commerce platforms in late September:
Amazon implemented full API blockades on Muse shopping agents, citing security concerns to protect its front-end traffic funnel.
Shopify, Walmart, Best Buy, and Expedia opened full integration access to enable seamless closed-loop checkout.
Following an initial sector-wide sell-off, stock trajectories diverged sharply: Shopify erased its losses, whereas online travel agencies (EXPE, BKNG, ABNB) remained under sustained pressure. The market is pricing in a fundamental shift: in the agent era, gateway assets depreciate, while infrastructure assets appreciate.
1. Gateway Platforms: The High-Margin Decision Layer Disintermediated
For two decades, internet monetization relied on gateway monopolies (Consumer $\rightarrow$ Platform $\rightarrow$ Merchant). Platforms captured buyer intent, monetizing via search placement, sponsored ads, and traffic distribution. Amazon’s $68.6 billion in 2025 ad revenue stands as a prime example of gateway asset valuation.
In the agent era, the user flow simplifies to: Consumer $\rightarrow$ Muse $\rightarrow$ Platform $\rightarrow$ Merchant. While the transaction persists, the decision-making, price comparison, and filtering processes are captured by the Agent. Consequently, the value of sponsored search listings, recommendation algorithms, and ad bidding declines. Discovery and aggregation platforms like Expedia, Booking, Yelp, and TripAdvisor face persistent headwinds as their core value proposition is bypassed, leaving them exclusively with back-end fulfillment functions.
2. Infrastructure Platforms: Monetizing Agent Transaction Volume
Platforms like Shopify do not rely on consumer-facing traffic gateways; their value resides in merchant operating systems, inventory synchronization, order management, and payment checkout. As agents drive order volumes, infrastructure providers capture incremental subscription and software fees. Payment networks benefit similarly: regardless of front-end interface changes, clearing, fraud prevention, and settlement infrastructure remain mandatory.
3. Amazon’s Dilemma: A Two-Front Game with Intact Moats
Amazon occupies a unique dual position as both a consumer gateway and a global logistics and e-commerce infrastructure provider. Blocking agents preserves its core ad business, while opening integration captures incremental unit volume—a strategic tension that will persist over the long term.
Despite this dynamic, Amazon retains structural defensive moats. Alexa for Shopping reaches over 100 million users, leveraging proprietary transaction data to run real-time price monitoring, deal alerts, and automated reordering. This ecosystem integration has expanded US average order value (AOV) by over 40%, supporting strong customer retention.
The deployment of Muse establishes a complete Agent value chain spanning physical connectivity to high-level compute:
Tier 1: Agent Communication Infrastructure (TWLO, BAND)
Agents require automated voice calling, SMS messaging, and corporate API integration to complete operational tasks, driving exponential growth in communication bandwidth. Twilio (TWLO) offers a enterprise communication middleware platform, while Bandwidth (BAND) supplies underlying voice network connectivity with high operating leverage as a core transport layer for Voice AI.
Tier 2: Internet Infrastructure & Edge Networks (NET, AKAM)
Autonomous internet access by millions of agents necessitates real-time identity verification, bot mitigation, API security, and edge compute. Cloudflare (NET) serves as the software entry point for developer ecosystems, controlling access policy and runtime permissions for autonomous agents. Akamai (AKAM) leverages its globally distributed edge nodes to handle local inference and edge security for real-time agent workloads.
Tier 3: Identity Security & Access Control (OKTA, PANW, CRWD)
The primary enterprise risk introduced by agents is permission sprawl. Machine-driven execution on behalf of humans requires specialized security architectures. Traditional systems secured human-to-system access; modern security architectures must govern non-human, machine-to-system delegation. Okta handles identity authentication and privilege management, Palo Alto Networks (PANW) delivers cloud network security, and CrowdStrike (CRWD) secures machine endpoints and agent credentials, expanding a new TAM in machine identity security.
Tier 4: Transaction Execution & Settlement (SHOP, WMT, V, MA, PYPL)
Once an agent executes a purchase decision, order fulfillment, inventory allocation, and payment processing must occur. Shopify provides merchant transaction infrastructure, while global payment networks (Visa, Mastercard) and digital wallets (PayPal) process clearing and settlement, directly monetizing higher transaction velocity.
Tier 5: Specialized AI Compute Clouds (NBIS, CRWV, IREN)
Muse confirms that autonomous agents consume substantially more compute capacity than static conversational AI. Nebius secured up to $27 billion in next-generation infrastructure contracts from Meta. As active users and task complexity scale, specialized GPU and specialized cloud infrastructure providers capture sustained compute utilization.
Alphabet / Google: While possessing a comprehensive internal AI technology stack, Google faces structural business model friction. The majority of its net income derives from active search advertising. In an agent-dominated paradigm, users bypass traditional search engine results pages (SERPs) as Gemini completes execution natively, compressing sponsored ad unit economics. Google's core risk is not technical execution, but preserving search ad revenue while transitioning to agent-first user interfaces.
Legacy Customer Care Outsourcing (CNXC): Customer experience providers reliant on labor-intensive call centers face sustained pressure as autonomous agents automate repetitive support functions. Conversely, hybrid orchestration providers (NICE, FIVN) face a dual dynamic: legacy seat-based revenue contracts compress, but human-in-the-loop agent management platforms create new software upside.
The primary significance of Meta Muse is signaling AI's transition from model training to autonomous task execution. While GPUs remain foundational, server CPU allocation and system-level infrastructure grow increasingly critical. Gateway platform valuations face structural pressure, whereas transactional infrastructure providers capture structural tailwinds.
The market spent three years evaluating GPU demand parameters; the focus now pivots to quantifying CPU requirements, edge network capacity, and emerging agent-native monetization models.