Everything About AI
NextFin News - The artificial-intelligence trade, which had narrowed into a single bet on graphics chips, widened on Monday as Meta Platforms' new personal AI agent Muse convinced investors that the next phase of the AI buildout will run on central processing units as much as on accelerators. The S&P 500 climbed 1.49% to 7,764.70 and the Nasdaq Composite rose 2.26% to a record 27,122.09, its first record close since June, while Meta itself jumped 11.43% to $741.25 — its best single session since April 2025.
The Rally That Changed the AI Trade
The move was broad, but the leadership was specific. Advanced Micro Devices gained 10% and briefly crossed a $1 trillion market value for the first time, touching an intraday high of $613.92 before settling with a record close. Intel rose 12%, Arm Holdings climbed 17%, and the Philadelphia Semiconductor Index advanced 4.3% for a fifth straight session of gains. The biggest percentage winners were not the pure-play accelerator names; they were the CPU designers.
At the center of the rotation is Muse, the personal AI agent Meta introduced on September 8 under chief AI officer Alexandr Wang. Unlike the chatbots that defined the first wave of consumer AI, Muse is built to act on a user's behalf: sending emails, booking travel, filling out forms, negotiating bills, and completing purchases through a checkout integration with Stripe's Link. It connects to Meta's own apps, including Instagram and Facebook, as well as third-party services such as Spotify, Gmail, OpenTable, Ticketmaster and Shopify. Access runs through a dedicated app on iOS, Android and the web, plus WhatsApp.
Early usage data gave the market something it had been waiting for: proof that a consumer agent can reach scale quickly. A week after launch, Wang wrote that Muse had become the number-one app in Apple's App Store. In a note to clients on Monday, Wells Fargo analyst Ken Gawrelski cited SensorTower figures showing the standalone Muse app set a single-day U.S. download record of 264,000 on September 19, its third consecutive day above 200,000, and reached a peak of 448,000 daily active users on September 18 — ten days after launch. By comparison, OpenAI's ChatGPT took nearly a full year to cross 200,000 daily U.S. downloads and 49 days to reach 450,000 daily active users.
Believe Muse assistant and Instinct mark the transition to consumer AI assistants from chatbots.
Gawrelski wrote, referring to Muse and Instinct, the invite-only agent from Spear Street Technology, as twin signals that the category has shifted. He raised his price target on Meta to $796 from $640 and lifted his 2027 earnings-multiple assumption. The question the market is now asking is not whether agents can attract users. It is what agents do to the chip stack underneath them.
The Mechanism: Agents Multiply CPU Cores per Watt of AI
For the past two years, the AI investment thesis was simple and narrow: buy the companies that make the accelerators that train and run large language models. Graphics processing units did the heavy lifting because the work was batch-oriented — a user typed a prompt, the model generated a response, and the GPU went idle. Agentic workloads break that pattern. An agent does not answer one question and stop. It maintains a plan, loops over tools, holds state across minutes or days, and orchestrates calls to email, calendars, payment systems, maps and inventory databases. Those operations are serial, logic-heavy and memory-bound — the native territory of the CPU.
Meta has said as much in its own infrastructure commentary. In April the company said that as its agentic AI work advances, "compute requirements are evolving to demand more CPU," and that its data centers are "outgrowing the abilities of traditional CPUs." That is why Meta has been working with Arm and Amazon Web Services on new CPU infrastructure, and why it continues to develop its own MTIA accelerators rather than relying solely on merchant GPUs.
The transmission channel from "agent adoption" to "CPU demand" runs through core counts. In an interview after unveiling the AGI CPU at a March product event, Arm CEO Rene Haas said CPU demand is "off the charts," and argued that each gigawatt of data-center capacity in an agentic world will require roughly 120 million CPU cores, versus about 30 million in the prior generation of AI workloads — a fourfold increase. Arm introduced the AGI CPU in March, its first in-house-designed data-center processor aimed specifically at agentic inference, and has since said customer demand for the chip has exceeded $2 billion across fiscal 2027 and 2028, more than double what it disclosed at launch. The company also says CPU requirements are rising by at least four times as application workloads become agent-driven.
Meta is not a passive customer in that buildout. It is a lead partner and co-developer of the AGI CPU across multiple generations, with the chips intended to operate alongside Meta's MTIA accelerators. That matters because hyperscaler-designed silicon is where the volume will concentrate: Google has migrated 30% of its internal applications to its Axion chips, and Microsoft's Cobalt 200 is scaling into newer AI workloads. When a company the size of Meta co-designs a CPU and deploys it at scale, the merchant CPU market reprices around that anchor.
There is also an edge component. Meta has partnered with AMD to optimize Muse Glimmer, an open-source model built to run locally on consumer PCs rather than in the cloud. Local inference shifts spend from data-center GPU rentals to endpoint CPUs and neural processing units — a margin-accretive shift for Meta and a new volume pool for AMD. Reports have put the price of AMD server chips up roughly 40% amid the wave of agent demand, and AMD CEO Lisa Su has said the server CPU market could exceed $120 billion by the end of the decade, a 35% compound annual growth rate.
The second-order implication is a change in the CPU-to-GPU ratio inside the data center. If agentic orchestration adds a persistent CPU layer on top of model inference, then every dollar of AI capital expenditure buys more CPU content than it did in the chatbot era. That redistributes the infrastructure dollar across the compute stack, and it is why Monday's tape looked the way it did: Arm up 17% and Intel up 12% against a 4.3% rise in the broader semiconductor index.
Structural Demand, Not Just a Sentiment Rebound
The cyclical-versus-structural question decides the whole trade. The stock move itself is cyclical: chip names had lagged, positioning was light, and a single strong data point can produce a violent mean-reverting bounce. But the driver behind it is structural, and the evidence sits in the workload itself rather than in the price action.
Three things make it durable. First, the workload composition has changed permanently: agents add orchestration, tool use and state management on top of model inference, and those layers are CPU-native. Second, the economics reinforce it — running an agent locally on a device is far cheaper per task than routing every step through a cloud GPU, so cost pressure pushes workloads toward CPU-heavy edge and server designs. Third, the supply chain is committing: Arm is manufacturing the AGI CPU on TSMC's 3-nanometer process with broad shipments expected in the second half of the year, and it has already secured memory, wafer, packaging and testing capacity.
The strongest evidence that this is not a one-day sentiment spike is the pace of adoption. Muse reached 448,000 daily active users in ten days; ChatGPT needed 49 days to reach the same level. If even a fraction of those users keep the agent running in the background, the always-on orchestration layer becomes a permanent new line item in data-center bills. The market has not fully priced that shift: Meta entered the week up roughly 20% in September on product anticipation, but the breadth of Monday's move into CPU suppliers — rather than a narrow Meta-only bid — signals a repricing of the entire compute stack, not just one stock.
The Counter-Thesis: A Laggard Bounce in a Still-GPU World
The bear case deserves space because it attacks the core of the trade. The strongest version is not that agents are unimportant; it is that Monday's winners are simply the stocks that had the most room to run, and that the GPU-centric model of AI is not actually being displaced.
Nvidia's management made exactly this argument last week, downplaying core counts and arguing that single-core performance matters more for agentic AI than the fourfold core-count expansion Arm is promoting. Nvidia is also entering the CPU market itself: its Vera processor is designed for agentic workflows, and the company is collaborating with Intel on a custom Xeon with integrated NVLink so that x86 cores can act as coherent hosts for its Blackwell and Rubin clusters. If the agentic future is still orchestrated inside Nvidia racks, then the biggest beneficiary remains Nvidia, not Arm or Intel.
There is also a demand-quality question. Muse's download surge came a week after launch and was amplified by heavy promotion across Meta's apps. Launch humps are common; retention is what determines infrastructure spend. Threads, by comparison, was downloaded more than 4.3 million times in the U.S. on its launch day, and the Meta AI app recorded 108,000 U.S. downloads on its debut — both far above Muse's early numbers. The SensorTower data shows Muse is growing faster than ChatGPT did, but it does not yet show whether users keep the agent active or whether the tasks it performs are heavy enough to move CPU procurement.
Finally, there is a self-cannibalization risk for the merchant CPU market. The largest buyers of AI CPUs are the hyperscalers, and the largest hyperscalers are designing their own silicon. Meta co-develops with Arm; Google builds Axion; Microsoft builds Cobalt. That structure concentrates volume in captive designs and squeezes the merchant CPU suppliers — Intel and AMD — that retail investors are buying on Monday's momentum.
The counter-thesis is credible on a one-day horizon. It becomes a real threat to the structural call only if the adoption data rolls over or if hyperscaler CPU designs fail to show up in capital expenditure. Until then, the burden of proof sits with the bears to show that agents leave the CPU stack unchanged.
What to Watch: Connect, Earnings, and the Falsifying Signal
The market's read of Muse is a second-order trade dressed as a first-order one. The first-order move — Meta's stock up 11% on strong agent downloads — is straightforward. The second-order move — CPU designers outperforming because agents change the shape of compute demand — is where the durable money is, if the thesis holds.
The beneficiaries split by architecture. Arm wins if agentic workloads shift data-center designs toward Arm-based server CPUs, which is the bet embedded in its $2 billion demand figure. AMD wins on both sides: its server CPUs capture the x86 share of agent orchestration, and its partnership with Meta on Muse Glimmer gives it a foothold in local, on-device inference. Intel wins only if it can execute on the supply side — it is facing delivery delays and a supply crunch even as prices rise — and if its NVLink collaboration with Nvidia keeps x86 relevant inside AI racks.
The exposed are the investors who treated the AI trade as a single-factor bet on accelerators. A world in which CPU content per unit of AI work rises fourfold redistributes the infrastructure dollar, and portfolios built entirely around one layer of the stack will underperform a portfolio that spans the whole compute pyramid.
The forward look breaks into three horizons. In the short term, sentiment and positioning will dominate; Monday's gap higher can fade if the Meta Connect keynote on September 23 fails to deliver concrete agent metrics or new model announcements. Mark Zuckerberg takes the stage at 7 p.m. Eastern on September 23, and analysts expect demos of Muse and the underlying Muse Spark models, possibly a next-generation large language model code-named Watermelon, and updates on wearables including camera-free smart glasses and AI-powered hearables.
Over the medium term, the decisive data points are Meta's third-quarter earnings on October 28 and any guidance on CPU-related capital expenditure. Over the long term, the structural call stands or falls on whether agentic workloads remain CPU-intensive as models and tooling evolve.
The base case is that agents add a persistent CPU layer to AI infrastructure, lifting Arm, AMD and an executing Intel over the next four to six quarters. The upside case is that Muse's adoption curve keeps outrunning ChatGPT's, forcing rivals — which, according to reports, are scrambling to develop their own consumer agents — into a capital-spending race that tightens CPU supply further. The downside case is that agent retention rolls over after the launch hump, or that Nvidia's vertically integrated racks capture the orchestration layer and leave the merchant CPU market smaller than the fourfold core-count narrative implies.
The falsifying signal is specific: if Muse's U.S. daily active users fall back below 200,000 within 30 days of launch, or if Meta's October earnings show flat or declining CPU-related capex guidance despite the agent push, the structural call is wrong and Monday's rally was a cyclical bounce.
The AI trade did not end on Monday; it broadened. The market spent two years paying for the brain of the machine. It is now starting to pay for the hands.
Explore more exclusive insights at nextfin.ai.
