AMD, INTC, ARM, 688041.SH · US · A-shares
CPU industry deep dive · 2026 H2
Agentic AI puts CPUs back at the center of data-center orchestration. Server CPU TAM revised higher toward ~$100–120B by 2030. Watch AMD/Arm globally and Hygon in China. See also AMD stock note and GPU utilization / AI infra deep dive.
Cite a section with a deep link, e.g. /en/r/cpu-industry-h2-2026#paradigm
Key snapshot
- Server CPU TAM 2030 (AMD)
- ~$120B
- Agentic AI CPU:GPU target ratio
- 1:1~1:2
- Q1'26 non-x86 server revenue share
- 47.9%
- 2026 supply/demand
- Shortage + hikes
Data as of 2026-07 · Sources: IDC, TrendForce, AMD / Arm management guidance, company filings, broker research, industry media
Thesis: Agentic AI pulls the CPU from “GPU sidekick” back to the data-center orchestration layer. Server CPU TAM is revised sharply higher toward ~$100–120B by 2030; shortages lift ASP, and both x86 and ARM benefit. Globally, watch AMD / NVIDIA / Arm; in China, Hygon plus Kunpeng localization.
1. Paradigm shift: why CPUs are tight and expensive again
Agentic AI orchestration bottleneck
Agent workloads lean on tool calls, API orchestration, and multi-task coordination—work that lands on the CPU. Studies show CPU can account for up to ~88% of end-to-end latency in tool-heavy agentic flows.
TrendForce: cluster CPU:GPU ratios migrate from ~1:4–1:8 toward 1:1–1:2; cores needed per GW of AI capacity may approach ~120 million.
Capacity and lead-time constraints
Advanced process and packaging capacity is crowded out by AI silicon; Intel / AMD server CPUs are near sell-out, with multi-month lead times and allocation shortfalls for some customers.
Multiple 2026 price hikes (consumer and server, often reported in the 10%–35% range) shift pricing power back to sellers.
Architecture diversification
AI racks (Grace/Vera and peers) plus CSP custom silicon (Graviton, Axion, Cobalt, etc.) push ARM toward parity with x86 on server revenue share.
IDC Q1'26: global server revenue ~$122.6B; non-x86 revenue share 47.9%. x86 still leads units, but the revenue gap is closing.
2. Market size and demand layers
Vendors and sell-side disagree on absolute dollars, but agree on direction: a high-growth decade for data-center CPUs.
| Source | Scope | Near-term | 2030 outlook | Implied CAGR | Takeaway |
|---|---|---|---|---|---|
| AMD (Lisa Su) | Server CPU TAM | — | >$120B | ~35% | Raised CAGR from 18% to 35% within seven months |
| Arm | Data-center CPU TAM | ~$24B | >$100B (FY31) | ~33% | Royalty model scales with industry expansion |
| BofA | Server CPU TAM | ~$43B (2026) | ~$125B | ~31% | Already embeds a high 2026 base |
| NVIDIA | Standalone CPU rack revenue | Guide ~$20B (FY) | — | — | Vera standalone CPU racks open a new category |
Demand structure: three layers
| Layer | Use cases | CPU role | Cycle |
|---|---|---|---|
| L1 AI clusters | Training / inference / agents | Orchestration + host + standalone racks | Very strong |
| L2 Cloud & enterprise | Virtualization, databases, ads/search | General-purpose workhorse | Strong |
| L3 PC / edge | Notebooks, industrial, endpoints | Efficiency & ecosystem | Mixed |
Q1'26 server revenue structure (illustrative)
Source: IDC · Q1 2026 · non-x86 YoY +107.6%; x86 −2.9% (supply constraints also contribute)
3. Competitive map: global four poles + multi-ISA China
Global: designers and platform vendors
| Company | ISA | Positioning | 2026 catalysts | Strengths | Risks | Stance |
|---|---|---|---|---|---|---|
| AMD | x86 | Server share aggressor | EPYC sold out; server CPU revenue surge | Perf/efficiency; share gains | Capacity; NVIDIA CPU competition | Overweight |
| Intel | x86 + foundry | Incumbent + process comeback | Price hikes, long lead times, 18A/Xeon refresh | Installed base; foundry narrative | Share loss; valuation ahead of earnings | Tactical / thematic |
| NVIDIA | ARM (Grace/Vera) | New AI-rack CPU entry | Standalone CPU racks; GPU bundling | System-level pricing power; channels | Ecosystem breadth; customer custom silicon | Satellite |
| Arm | IP licensing | Architecture standard-setter | CSP custom + AGI CPU racks | High operating leverage; asset-light | Customer in-house IP dilutes royalty | Overweight (leverage) |
| Apple / CSPs | Custom ARM | Vertical integration | M-series / Graviton iterations | HW/SW co-design; efficiency | Not sold broadly | Indirect beneficiaries |
China: two leaders, many followers; four ISA paths
| Vendor | Path | Flagship | Core markets | Ecosystem | Investment map |
|---|---|---|---|---|---|
| Hygon | x86-compatible | Hygon CPU / DCU | Finance, telecom, DC, AI | Low migration cost; CUDA-like DCU stack | 688041 core holding |
| Huawei Kunpeng | ARM | Kunpeng 9xx | Carriers, SOE, internet | openEuler + full stack; server share >20% | Indirect (OEMs / partners) |
| Phytium | ARM | Tengyun / desktop | Government desktop, Xinchuang | Desktop shipment lead | China Great Wall, etc. |
| Loongson | LoongArch (indigenous) | 3A/3C6000 | Gov, defense, industrial | Highest autonomy; ecosystem climbing | 688047 |
| Zhaoxin | x86 | Kaixian, etc. | Xinchuang PCs / some servers | Desktop compatibility | Not directly listed |
| Sunway | Indigenous (Alpha lineage) | Sunway series | HPC, specialty | Security niches | Not directly listed |
4. Technology & supply-chain pivots
Product and architecture trends
- Chiplet / advanced packaging is the battleground for density and yield
- CPU–accelerator heterogeneity: NVLink-C2C, Hygon HSL, LoongLink, etc. set system efficiency
- Standalone CPU racks are a new category: NVIDIA Vera, Arm AGI CPU Rack
- Efficiency: ARM / custom cloud silicon keep compressing classic dual-socket x86 general workloads
Upstream constraints
- Wafer: TSMC advanced nodes remain the lifeline for most fabless CPUs; Intel Foundry is the swing factor
- Memory / power / cooling tighten in lockstep with CPUs, lifting system ASP
- Export controls and entity lists keep reshaping China supply-chain paths
- EDA, IP, and advanced packaging tools remain long-cycle bottlenecks
5. Key company assessments
1. AMD — server CPU aggressor, highest elasticity
Multi-quarter highs in server CPU revenue; strength in both cloud and enterprise. EPYC is near sold-out in 2026—unit-driven with ASP upside under shortage. Instinct GPUs make AMD one of the few vendors that can monetize AI budgets on both CPU and accelerator.
Risks: NVIDIA standalone CPU racks siphoning demand; wafer capacity; PC cyclicality. Best as the global CPU-theme core. See AMD stock note.
2. Intel — shortage beneficiary under share + valuation pressure
Shortage and price hikes help near-term revenue and margin expectations; Foundry / 18A is the long-dated option. But x86 server share keeps leaking to AMD, and non-x86 approaches revenue parity—the long moat narrative is still under repair. Track the operating inflection; size a full position only after earnings and FCF catch the stock. See INTC research.
3. Arm Holdings — picks-and-shovels on architecture expansion
Almost every incremental ARM server (CSP custom, NVIDIA Grace/Vera, Ampere, Kunpeng/Phytium) reinforces the IP and royalty story. Data-center TAM guidance has been lifted above $100B; the asset-light model has high elasticity to industry expansion. Risks: large customers raising in-house share, lumpy license deals, and a non-cheap multiple.
4. Hygon (688041.SH) — dual-chip (x86 + DCU) domestic CPU market anchor
Investment case: The only scaled commercial domestic high-end x86 CPU plus a CUDA-like DCU stack; low migration cost for finance/telecom. Intel price hikes and AI inference clusters’ CPU attach open both Xinchuang and open-market runways.
Risks: Growth is visible, but the market cap is hypersensitive to growth and policy expectations. Watch gross margin (~58% in 2025, down YoY), SBC/R&D, advanced-node access, and export compliance.
Positioning: domestic core · manage valuation volatility
5. Loongson · Sugon · China Great Wall — satellite names under ISA divergence
Loongson (688047): fully indigenous LoongArch; 2025 revenue ~RMB 635M (+26%), gross margin ~47%, losses narrowing. Fits security-sensitive use cases; ecosystem and server volume remain the prove-it items.
Sugon: key OEM/ecosystem partner to Hygon; benefits from CPU/DCU ramp and liquid-cooled HPC orders—beta to Hygon and compute CapEx.
China Great Wall et al.: Phytium desktop/server OEM channels; more tied to Xinchuang order cadence, less elastic than design leaders.
6. Investment scorecard
Dimensions (1–5 each, max 30): demand fit, competitive position, growth visibility, earnings quality, valuation attractiveness, risk control
| Name | Demand | Position | Growth | Earnings | Valuation | Risk | Total | Stance |
|---|---|---|---|---|---|---|---|---|
| AMD | 5 | 5 | 5 | 5 | 3 | 4 | 27 | Global core |
| Arm | 5 | 5 | 5 | 4 | 3 | 4 | 26 | Levered overweight |
| NVIDIA (CPU) | 5 | 4 | 5 | 5 | 3 | 3 | 25 | Satellite add |
| Hygon | 5 | 5 | 5 | 4 | 2 | 3 | 24 | Domestic core |
| Intel | 4 | 3 | 3 | 3 | 2 | 3 | 18 | Theme watch |
| Loongson | 3 | 3 | 3 | 2 | 3 | 3 | 17 | Theme / satellite |
7. Portfolio suggestions
Global growth
AMD 45% + Arm 30% + NVIDIA 15% + cash/hedge 10%
Bet on server CPU shortage and ARM architecture expansion; avoid Intel valuation debates.
China Xinchuang / compute
Hygon 55% + Sugon 25% + Loongson/Great Wall 20%
Hygon owns the x86+AI spine; OEMs and indigenous ISA as satellites—watch valuation drawdowns.
Balanced (cross-market)
AMD 35% + Hygon 30% + Arm 20% + Intel theme 15%
Global shortage alpha + localization beta; Intel only as event-driven sleeve.
8. Summary
The 2026 CPU story is Agentic AI re-pricing “orchestration compute.” Shortage and price hikes are the near-term trade; architecture diversification and localization are the medium-term structure. Globally, lead with AMD / Arm; domestically, anchor on Hygon with OEM ecosystem as wings. Intel benefits from shortage but faces twin questions on share and valuation—better as a theme than a core holding.