The core thesis is that AI agents are shifting the bottleneck from GPUs to CPUs, making data-center CPU stocks the next battleground after the GPU run in NVDA and MU. NVDA is already pushing Grace into Vera, ARM is trying to monetize its first own CPU, QCOM is pivoting from phones into a 2029 data-center target, and AMD is converting AI demand into rack-scale deals with hyperscalers. The last segments frame these as priced very differently: ARM near 50x sales, AMD rich but still scaling, QCOM under 4x forward sales, and NVDA far ahead but no longer the only CPU story. The watchlist turns on whether META, TSMC, and the announced deployment timelines actually hold, because every one of these CPU plays depends on those milestones.
$10,000 in NVDA at Chat GPT’s launch less than 4 years ago would be over $120,000 today; $10,000 in MU just last year would already be over $75,000 today.
Data centers were priced as a GPU business for the last 4 years: accelerators from NVDA, AMD, and AVGO were the core product, with power, cooling, racks, networking chips, memory, and CPUs built around them.
AI agents change the workload split: GPUs still run the math, but CPUs choose the next action, fetch files, move data, and manage tool calls and sub-agents, making CPU the next battleground.
NVDA’s Grace CPU has 72 cores, and the superchip links two Grace CPUs into 144 ARM Neoverse V2 cores; ARM owns the Neoverse data-center core design NVDA licenses, so NVDA pays ARM royalties on Grace.
NVDA has sold nearly 2.5 million Grace CPUs to date, and Vera is NVDA’s first fully custom CPU: 88 custom Olympus cores, ARM instruction set, and first deliveries last quarter to Anthropic, OpenAI, SpaceX, and ORCL.
NVDA’s latest earnings call reported over $75 billion in data center revenue for the quarter, with roughly $60 billion from compute and $15 billion from networking; Colette Kress said NVDA expects around $20 billion in CPU revenue this year and sees AI CPUs as a new $200 billion TAM.
ARM’s initial AGI CPU was delivered to multiple customers on July 29, ending a 35-year model in which ARM sold blueprints to AAPL, QCOM, NVDA, and AMZN and collected up to 5% royalties.
ARM’s AGI CPU has up to 136 cores, runs at 300W, uses TSMC 3 nm, and has META as lead partner; ARM claims more than double the performance per rack versus x86 and says demand exceeds $2 billion over 2 years, with capacity for only about half.
ARM revenue was $1.3 billion for the quarter, up 22% year over year; royalty revenue also rose 22%, data-center royalties more than doubled, and this was the first quarter ARM stopped reporting remaining performance obligations, last at $2.07 billion in March.
ARM’s gross margin is now the key watch item because the business is shifting from high-margin royalties to hardware sales with wafer costs, capacity risk, and hardware-style margins; real AGI-CPU revenue should show up in fiscal 2028.
ARM is priced near 50x sales and almost 39x forward sales, while the other stocks in the video sit between 4 and 14x; QCOM’s forward P/S is under 4, and the video frames ARM as already priced in.
QCOM’s modems were inside iPhones for years, and it still collects royalties on almost every smartphone; its Orion cores are built under an ARM architectural license, and QCOM’s Snapdragon chips are in most flagship Android devices.
QCOM reported July 29 earnings of $9.9 billion revenue and $2.21 EPS; automotive hit $1.6 billion, up 61% YoY for a 23rd straight quarter of double-digit growth, IoT was $1.8 billion up 9%, and handsets were $5.1 billion down 20%.
QCOM management said handset revenue has bottomed and announced double-digit chip price increases starting September 1; non-handset revenue is now carrying the company, with non-handset growth guided from 24% this fiscal year to over 60% next year.
QCOM’s June investor day target is more than $15 billion in data-center revenue by 2029 versus about $300 million this year, implying roughly 50x growth in 3 years; META signed a multi-generation CPU deal and Satya Nadella confirmed Azure will deploy QCOM accelerators starting next year.
AMD’s July 20 Helios announcement with MSFT pairs Instinct MI455X GPUs with sixth-generation Epic CPUs, code-named Venice, as a rack-scale system; MSFT is also launching NDMI455X V7, Azure HDV2, and Azure HX V2 on AMD hardware.
OpenAI signed for 6 gigawatts of AMD GPUs, META for 6 more gigawatts of Helios racks, Anthropic for another 2 gigawatts, and AMD also committed up to $5 billion in equity to Anthropic; the deals include warrants for up to 160 million AMD shares for OpenAI and 160 million for META at 1 cent each, described as up to 20% dilution to secure the deals.
AMD reported revenue of $11.5 billion, up 50% YoY; data-center revenue was $6.7 billion, up 107% YoY and 58% of total revenue, with next-quarter revenue guided to about $13 billion; AMD is described as 5x NVDA on P/E and 3x on forward P/E, with multiples rising as NVDA’s fall.
The market is framed as shifting from units sold to tokens and watts, with AI power deployed per watt becoming the real share metric; every company named relies on TSMC for chips, so delays or supply constraints at TSMC would hit NVDA, ARM, QCOM, and AMD alike.