Nvidia's AI CPU Push Threatens Intel and AMD
Meta's Muse personal agent rocketed to the No. 1 free iPhone app in the United States within 10 days of launch, providing an early consumer-scale example of how AI agents can increase demand for cloud CPUs as well as GPUs. The agent can research products, fill out forms, handle email, make reservations, and continue working after the user closes the app—fundamentally different from a chatbot that waits for a question, generates an answer, and stops.
What matters to semiconductor investors is where that work takes place. Meta says Muse operates in a secure virtual machine that isolates the agent's activity and the user's data. The environment supplies the browser, processing, memory, and storage needed to perform a series of tasks. While the large language model still depends heavily on GPUs for reasoning and inference, the agent also needs CPU resources to run software, direct the browser, move data, manage permissions, call tools, and keep the job operating.
Muse puts a consumer product around a change already taking place inside data centers. AI is moving from models that answer questions to agents that perform work. That transition expands the infrastructure bill beyond accelerators and creates a second source of growth in general-purpose computing. One agent does not require one physical CPU, but millions of agents can create substantial demand for virtual CPU hours, memory, storage, browser sessions, and secure cloud environments. A popular consumer agent can generate repeated, persistent computing activity from each user—an agent working for 20 minutes across several websites consumes a different mix of infrastructure than a chatbot producing a response in seconds.
This expansion does not automatically benefit Intel and AMD. Nvidia is preparing its Vera CPU for precisely these workloads, cloud providers are deploying processors of their own, and Arm Holdings supplies the architecture behind many of the challengers. The CPU market is growing, but its competitive structure is changing at the same time.
Nvidia introduced the Grace data-center CPU in 2021, but Vera represents a more ambitious move. Vera contains 88 custom Olympus cores built on the Arm architecture and is being positioned both within the Vera Rubin platform and as a stand-alone server processor. Nvidia designed it for work surrounding accelerated computing, including agent orchestration, data processing, analytics, tool use, and sandboxed code execution.
Management has cited visibility to approximately $20 billion in server CPU demand and said CPU revenue could more than double in fiscal 2028. Nvidia's advantage is not based on Vera having the industry's highest core count—AMD's sixth-generation EPYC 9006 family, code-named Venice, offers as many as 256 cores and 512 threads. Intel retains a large installed base, mature software compatibility, and deep enterprise relationships. But Nvidia approaches the purchase from another direction. A customer selecting a Vera Rubin rack has already chosen Nvidia GPUs, networking, interconnects, software, and system architecture. The CPU arrives as part of that design. Intel and AMD traditionally competed for an individual processor socket; Nvidia is attempting to make the socket one element of a larger platform decision.
That difference matters because AI systems are becoming more tightly integrated. Performance depends on how quickly data moves among the CPU, accelerator, memory, and network, not simply on the specification of one chip. Nvidia can optimize those connections across the rack and use CUDA and its networking products to keep the customer within its platform. The strategy does not require Vera to displace every Xeon or EPYC processor—Nvidia only needs to increase the amount of semiconductor content it captures from customers already buying its accelerators.
Nvidia's move into CPUs did not begin with Muse or Vera. In September 2020, Nvidia agreed to acquire Arm from SoftBank for $40 billion. Since many Nvidia competitors also rely on Arm licenses, regulators opposed the deal. The companies terminated the acquisition in February 2022. Nvidia forfeited the $1.25 billion prepaid at signing but retained a 20-year Arm architecture license. It later invested approximately $100 million in Arm's 2023 public offering, purchasing about 1.96 million shares, eventually selling the equity position. Grace and Vera show that Nvidia did not need to own Arm to build a CPU business around the architecture. The failed transaction changed the route, not the objective—Nvidia moved from an effort to own the architectural supplier to designing Arm-based processors and controlling how they operate inside Nvidia systems.
The Information Network estimates that AI data-center systems revenue will increase from approximately $560 billion in 2026 to $1.68 trillion in 2030, a compound annual growth rate of 31.6%. Accelerators will remain the largest component, rising from an estimated $350 billion to $970 billion. Nvidia and Arm may capture more of this growth than many investors expect, while Intel and AMD can increase CPU revenue and still lose share of the expanding market.