←2026-09-27 · VM Genius
Analysis
From the body of the fallen crayfish, Meta Muse has been born, and everyone must take this seriously. This is not the concept of another hit app; it is the landmark event marking AI’s formal shift from the model-training era into the agent-deployment era. Muse completely overturns the GPU-first compute logic, tears open the value gap between the “entry point” and “infrastructure” of internet platforms, and reconstructs a new five-layer AI industry chain. The curtain is now slowly rising on an industry-wide value reconstruction.
Muse is the first consumer AI Agent that can autonomously close the loop in the background: sending emails, booking travel, canceling subscriptions, filling out forms, comparing prices and placing orders. It keeps running the workflow even after you close the app. After launching on September 8, the numbers broke records: 2.8 million downloads in the U.S. and Canada in 12 days, more than 2× ChatGPT over the same period; over 500,000 users in the first week, and a single-day U.S. download peak of 264,000 to top the App Store. Actual usage is 10× internal testing, and Meta Chief AI Officer Alexandr Wang admitted it far exceeded expectations.
On the business model, Zuck has clearly chosen a transaction take-rate route: they only charge when the task gets done, which is completely different from subscription and advertising models.
What is most disruptive to the industry is the compute structure: GPUs in Meta data centers are only responsible for “thinking” (task decomposition, path planning), while CPUs are fully responsible for “doing” (opening browsers, logging into accounts, submitting operations). In user tests, Muse’s dedicated cloud computer has no discrete GPU; just 2 virtual CPUs + memory + storage can run the entire process.
Meta has 3.6 billion global daily active users, and now even a small-scale grayscale rollout has exploded compute demand. This is no longer a question of “do Agents need compute”; it is a question of “how many CPUs do Agents need.”
For the past three years, AI investing meant counting GPUs — and nothing else. Muse has pushed CPUs straight from supporting role onto center stage. The training era competes on GPUs, the Agent era competes on CPUs — the compute logic is starting to reverse.
CPU/GPU ratios vary wildly by scenario:
AI training: 1 CPU to 7–8 GPUs — CPU is purely auxiliary;
Standard inference: 1 CPU to 1–3 GPUs — status rises;
AI Agent: the ratio flattens all the way to 1:1, and in complex scenarios can reach 4:1.
AMD's Lisa Su said on the earnings call: the industry's compute ratio is converging toward 1:1. Servers running Agents are no longer GPU-stacked training clusters — they look more like traditional operational servers, and the CPU becomes the essential core.
A rough estimate: 100 million U.S. knowledge workers, 1 Agent each, with every 10 Agents occupying 1 server CPU — that's 10 million CPUs of demand; 1 billion practitioners globally, 10 Agents each — that's 1 billion CPUs. Yet global server CPU annual capacity today is only 3.5–4.5 million units — a gap of more than 20x.
Demand is already landing: Intel says it can only meet half of customer demand; Anthropic spent $11.6 billion to sign a 7-year CPU compute deal with Akamai; the buying focus of the big players is shifting from GPU to CPU; AMD's server CPU revenue last quarter surged 70% year-over-year.
Industry consensus: by 2030, inference compute will overtake training to become the data center's largest consumption item, at a 35% CAGR.
Global data center CPUs are a three-way split among Intel, AMD, and Arm:
x86 camp: Intel's supply is the biggest bottleneck; AMD's share is already approaching half the market, and its market cap has gone from $14 billion in 2017 to trillion-dollar territory — the best proof of the inflection point.
Arm camp: the fastest-growing sub-segment — NVIDIA Vera, and AWS / Google / Microsoft custom chips are all Arm architecture; after Arm enters chip sales, Meta is the number one customer.
AMD raised its 2030 market size forecast three times in under a year: from $60 billion to $120 billion, and then to $220 billion — versus 2025's $26 billion scale, that's quadrupling in four years, with the core driver entirely AI Agents. Arm's judgment: in the Agent era, CPU capacity per unit of power needs to reach 4x or more of today's level.
The market initially treated Muse as a negative for Amazon. That’s not the full picture. In the consumer agent boom, AWS is one of the winners: near-term capex pressure is matched by long-term, highly visible profit release, and the current valuation compression corresponds to $65B of upside.
Agent workloads are full-stack by nature: continuous reasoning loops rely on CPU, large context relies on memory, and high-frequency interaction relies on high-bandwidth, low-latency networking. Cisco data shows that an agent performing the same task generates 450% of the network traffic of a human; Deloitte estimates that inference compute will account for two-thirds in 2026.
AWS is precisely positioned here: its in-house Graviton CPU delivers 30%–40% better price/performance than comparable chips, the next-generation Graviton5 is ramping twice as fast, the top 1,000 EC2 customers have fully deployed it, and in-house hardware revenue nearly doubled quarter over quarter. At the same time, AWS pairs Trainium with NVIDIA GPUs, making it a natural fit for multi-architecture compute optimization.
For enterprises to deploy agents, they need more than compute — they need the full chain of development, orchestration, security, and operations. AWS spans from foundational platforms such as SageMaker and Bedrock to agent-level services such as AgentCore, Quick, Transform, and Continuum, covering the entire lifecycle so customers don’t have to assemble their own tech stack.
The results already validate this: in Q2, AWS’s order backlog was $496B, doubled year over year; its AI and in-house chip businesses each surpassed $25B in ARR. On the capacity side, compute will double by the end of 2027 compared with 2025, and most of the 2027 capacity and a large share of 2028 capacity have already been locked in by customers.
Muse’s emergence created a fork in the e-commerce industry in late September: Amazon completely blocked Muse’s shopping access on security grounds, defending its traffic gateway; Shopify, Walmart, Best Buy, Expedia, and others opened access across the board and closed the payment loop. The market initially fell broadly, then diverged decisively after 10-plus days: Shopify quickly recovered its losses, while EXPE, BKNG, and ABNB remained under pressure. The core reason is that the market finally figured it out: in the Agent era, gateway assets are depreciating, while infrastructure assets are appreciating.
Over the past two decades, the core internet windfall was gateway monopoly: consumer → platform → product. The platform controlled purchase intent and earned the richest profits through advertising and traffic distribution. Amazon’s 2025 advertising revenue of $68.6 billion is a reflection of gateway value. In the Agent era, the chain becomes: consumer → Muse → platform → product. Orders are still there, but the core steps of decision-making, price comparison, and screening are intercepted by the Agent, directly shrinking the value of search placements, recommendation rankings, and ad auctions. That is why platforms like EXPE, BKNG, YELP, and TRIP, which rely on search comparison and traffic distribution, remain under pressure: their core value is being bypassed, leaving only back-end fulfillment value.
Platforms like Shopify, which do not touch the consumer-facing gateway, derive their core value from merchant systems, inventory, orders, and payment checkout. The more transactions agents bring, the more service and system fees Shopify collects—so it benefits from ecosystem expansion. The same logic applies to payment-tool infrastructure: no matter how entry points change, clearing, anti-fraud, and refunds are essential, so it captures incremental volume all the way through.
Amazon’s e-commerce segment is the only platform that combines a top-tier gateway with full-stack infrastructure. Blocking protects its advertising base; opening captures incremental orders. This tug-of-war will continue. And its moat is still very high: Alexa for Shopping covers more than one hundred million users and uses first-party transaction data for precise price comparison, price alerts, and automatic ordering, lifting the U.S. site’s average order value by more than 40%; ecosystem stickiness remains strong.
The rollout of Muse is creating a complete Agent industry chain—from underlying communications at the bottom to compute at the top—with five layers benefiting in sequence:
Agents need to make phone calls, send text messages, and integrate with enterprise APIs to execute in the real world, so communications demand is exploding exponentially. TWLO covers omnichannel enterprise communications scenarios and is a mature middleware layer for connecting Agents to enterprises; BAND is deep in the underlying voice network—smaller but higher-beta—and is the core vehicle for Voice AI.
A massive number of Agents accessing the internet autonomously will require identity verification, bot management, API security, and edge computing. NET is the software entry point for Agents, tied to the developer ecosystem and managing access permissions for all Agents; AKAM uses its global nodes for edge inference and security protection, benefiting from real-time Agent use cases.
The biggest risk for Agents is permission overload: machines acting on behalf of people require a completely new security architecture. Security used to solve “how people access systems”; now it solves “how machines securely access systems on behalf of people.” OKTA handles identity authentication and permission management, PANW covers cloud security across the entire network, and CRWD focuses on endpoints and machine identity, giving rise to a brand-new hundred-billion-dollar market.
Once Agents finish making decisions, those decisions ultimately have to turn into transactions—inventory, orders, payments, and refunds are all essential. Shopify is the core transaction system for connecting Agents to merchants; card networks and digital wallets handle the settlement loop and benefit from the incremental transaction volume throughout.
Muse has validated that Agents are even more voracious consumers of compute than chatbot AI. Nebius has already secured a next-generation infrastructure contract with Meta worth up to $27B; the more users and the more complex the tasks, the higher the compute consumption handled by GPU cloud providers.
Google: Has the most complete technology stack, but its business model is self-contradictory. Core profit comes from active search advertising. In the Agent era, users no longer need to enter the search results page; Gemini completes the entire flow directly, and the value of ad slots shrinks accordingly. Its challenge isn’t building an Agent — it’s protecting its search advertising base in the Agent era.
Traditional customer service outsourcing (CNXC): Agents directly replace a large amount of repetitive human agent seats, and the legacy business remains under sustained pressure. NICE and FIVN are more mixed — traditional agent seats shrink, but pivoting to Agent orchestration and human-machine collaboration can capture new incremental demand.
The real significance of Muse is that it declares AI has moved from the “burning cash to train models” phase into a new phase of “landing and getting real work done.” GPUs will not be replaced, but CPUs will carry more and more weight; entry-point platforms face a re-rating of value, while infrastructure-type assets capture the full incremental upside; pricing for the new industrial chain’s journey from 0 to 1 has only just begun. The market spent three years figuring out how many GPUs AI needs—now, it’s time to figure out how many CPUs, how much infrastructure, and how many new business models AI needs.