UBS analysts contend that the rise of agentic AI is prompting a fundamental reworking of data center architecture, moving the market beyond a sole focus on GPU-heavy training toward complex orchestration of distributed workloads. The research argues this change will materially expand demand for server central processing units.
UBS estimates the server CPU total addressable market (TAM) could increase roughly fivefold over the next half decade - from a $30 billion baseline in 2025 to about $170 billion in 2030. The bank's view reflects expert inputs indicating that agentic AI deployments typically require three to five times more CPU cores per user than conventional workloads.
Market share and vendor implications
Within this enlarged market, the ARM instruction set appears positioned to capture a disproportionate slice of new demand. UBS projects ARM unit share will climb to between 40% and 45% by 2030, a sizable jump from an estimated 15% share in 2025. The bank has accordingly raised its price target for ARM from $175 to $245, citing an increased long-term EPS compound annual growth rate of 37%.
For other chipmakers, the opportunity varies by architectural strengths. Advanced Micro Devices Inc (AMD) is expected to benefit strongly from its established capabilities in high core counts and multithreading - attributes UBS views as critical for scaling the parallel subagents that agentic systems spawn.
Intel Corporation (Intel) is noted as attempting to close the performance gap with its forthcoming Coral Rapids offerings. UBS suggests Intel's more immediate upside may materialize from a spillover effect into the PC market as agentic workloads push computation toward end-user devices.
Workload trends and hardware attach rates
UBS highlights a growing tendency for agentic AI to execute workloads locally on end-user devices in order to harness otherwise unused compute capacity and to minimize cloud latency. That dynamic is expected to trigger a meaningful PC upgrade cycle, creating a secondary growth vector for x86 CPU vendors.
Technically, UBS reports a sharp step-change in CPU-to-accelerator attach rates for agentic inference. While traditional training workloads have historically required roughly 8 to 12 CPU cores per GPU, agentic systems are estimated to demand on the order of 80 to 120 cores per GPU. This stems from the need to provision an isolated execution environment - a "sandbox" - for each subagent an AI system spawns to complete tasks.
As a result, UBS anticipates a movement away from configurations that previously paired one CPU for every four GPUs toward tighter ratios such as one-to-two or even one-to-one in some deployments over the coming years.
Pricing and revenue projections
The research notes that these technical and architectural shifts carry material pricing implications. Central Processing Unit average selling prices (ASPs) for high-end AI-focused CPUs are likely to accelerate meaningfully. UBS cites examples of large core-count AI CPUs - including NVIDIA's 144-core Grace and AWS's 192-core Graviton 5 - which the bank expects could command unit prices in the $3,000 to $4,000 range.
Looking to 2030, UBS projects pro-forma CPU revenues of about $41 billion for AMD and about $39 billion for Intel. ARM's total CPU-related revenue is forecast at roughly $26 billion, of which UBS attributes $16 billion to device sales and $10 billion to royalty income.
These projections reflect UBS's current modeling of how agentic AI could reshape compute demand and vendor economics; they rest on the assumption that agentic deployments will substantially raise CPU core requirements and that hyperscalers and device makers will pursue denser, more power-efficient CPU architectures.