Server CPU & AI Accelerator Market Shift: ARM Surges, x86 Retreats, Custom Silicon Challenges Nvidia

Server CPU Market: ARM's Comprehensive Rise and x86's Defensive Battle

Bank of America (BofA) global research reveals a decisive shift in the server CPU landscape. The total addressable market (TAM) is projected to expand from approximately $35 billion in 2025 to $210.6 billion in 2030, demonstrating robust growth despite AI accelerator hype.

Intel and AMD Share Compression

Intel : Value share forecast to plummet from 40.4% (2025) to 22.0% (2030).

AMD : Steady growth from 27.4% to 30.7%, securing the clear second position.

ARM Ecosystem Explosive Expansion

ARM Merchant (commercial ARM) : Share surges from 29.3% to 37.9%, becoming the single largest segment. Driven by cloud providers' pursuit of performance-per-watt (e.g., AWS Graviton, Ampere).

ARM Custom (in-house ARM) : Share grows from 2.7% to 9.4%, reflecting hyperscalers' trend toward custom silicon optimized for specific workloads.

This signals a data center transition from "general-purpose compute first" to "efficiency and customization first." ARM's performance-per-watt advantage is eating into traditional x86 territory.

AI Accelerator (XPU) Market: Nvidia's Dominance and the "Custom Chip" Encirclement

Measured in gigawatts (GW) of deployed data center capacity — a more physical proxy than chip shipments — annual deployment is projected to leap from 13.0 GW (2025) to 55.7 GW (2030), a 4× increase. Energy supply emerges as the core bottleneck for AI scaling.

Nvidia's Absolute Growth but Relative Dilution

Nvidia's absolute deployed capacity rises from ~10 GW to ~25 GW, yet its share of total capacity is visibly eroded by 2030.

Diversified Compute Landscape Formation

Hyperscaler custom ASICs : Google (TPU), AWS (Trainium), Meta (MTIA), Microsoft (Maia) show thickening capacity blocks in 2026–2030, reducing Nvidia dependency and optimizing TCO.

New entrants : OpenAI appears as a distinct block, hinting at chip design or large-scale cluster builds; Anthropic (TPU) reflects deep model-company–chip-vendor binding.

AMD pursuit : Grows steadily but remains smaller than Nvidia.

The market moves from "single dominant player" to "one superpower, multiple strong contenders." Hybrid deployment of custom ASICs plus merchant GPUs becomes standard for hyperscale data centers.

Summary and Outlook

Combining both charts yields three structural conclusions:

Heterogeneous compute architecture : Future data centers will not be x86 monocultures. CPU layer: ARM claims half the market. Accelerator layer: GPUs coexist with diverse custom ASICs (TPU, etc.).

Performance-per-watt is the core metric : ARM's CPU rise and custom XPU proliferation are both driven by energy efficiency. At GW-scale deployments, performance-per-watt directly dictates operating cost and expansion ceiling.

Vertical supply-chain integration : Large cloud and AI firms integrate upstream via custom silicon (ARM Custom + custom XPU) to build moats beyond commodity hardware.

By 2030, the compute market will be 6× larger, but its underlying logic has flipped: the era of general-purpose hardware dividends has ended; the age of full-scale customization and efficiency optimization has arrived.