AI server ODMs benefit from rising CSP capital spending and increasing architectural complexity
AI summary card
AI server ODMs benefit from rising CSP capital spending and increasing architectural complexity
JPMorgan believes AI server ODMs remain key beneficiaries of CSP AI infrastructure expansion; Hon Hai and Quanta are better positioned in the NVL72 supply chain, while Celestica and Wiwynn lead in ASIC servers.
- 2026 capital spending guidance for the top four U.S. CSPs exceeds US$650 billion, up from about US$415 billion in 2025, with the incremental spending mainly going to higher GPU and memory content.
- NVL72 rack shipments are expected to rise to about 65,000 to 70,000 units, significantly above the prior year's roughly 25,000 to 27,000 units.
- Hon Hai and Quanta are the core NVL72 suppliers for major CSPs, with shares of about 40%-45% and 25%-30%, respectively.
- ASIC servers are more customized and have lower silicon ASPs, which is usually more favorable for ODM margins; TPU/Trillium and T3 are key highlights.
- Complex design capabilities, frequent product upgrades, and NPI experience continue to concentrate the competitive landscape among a small number of leading vendors.
Report interpretation
Overview
This report updates the market share matrix for AI server ODM systems and modules. JPMorgan believes ODMs, as system integrators, are key enablers for cloud service providers building AI infrastructure. Although server ODM margins may gradually decline over time as silicon ASPs rise, absolute profits are still expected to increase, driven by shipment volume and MVA growth.
Core views
The core view is that U.S. CSP capital spending faces upside risk, while demand for AI server modules and architectural innovation remains supported; Nvidia's GPU supply chain is generally stable, but Wiwynn has an opportunity for share gains as a second source for VR200 at Oracle and Meta; and ASIC servers have more durable ODM economics due to higher customization and lower customer switching risk. Hon Hai and Quanta are the top picks in the NVL72 supply chain, while Celestica and Wiwynn lead in ASIC servers.
Analysis framework
The report cross-validates the competitive landscape of AI server ODMs through CSP capital spending, NVL72 rack shipments, GPU module supplier matrices, Nvidia and AMD GPU system supply chains, ASIC server supply chains, and historical stock ratings.
Methodology notes
Break down primary and secondary ODM supplier relationships by customer, architecture, and system level
The report separates GPU modules, Nvidia GPU systems, AMD GPU systems, and ASIC servers, comparing primary suppliers, secondary suppliers, and potential new suppliers across different CSP-ODM relationships.
Use cloud service provider capital spending as a leading indicator for AI server demand
The report uses capital spending guidance and free cash flow ceilings for the top four U.S. CSPs to estimate potential upward revision room, and maps that to demand for GPUs, memory, and advanced server modules.
Asset mapping & comparison
Structured mapping from thesis to named assets (strengths, weaknesses, peers, risks).
- Hon Hai Precision (2317.TW)Core supplier of NVL72 and AI server systems
- Strengths
- It has the highest share in the NVL72 supply chains of major CSPs and possesses execution capabilities in GPU modules and system design.
- Weaknesses
- Server ODM margins may come under pressure as silicon ASPs rise.
- Comparison
- The report lists Hon Hai and Quanta as the top picks in the NVL72 supply chain.
- Risks
- CSP capital spending revisions may disappoint, supply chain share may be diluted by second-source suppliers, and architecture transitions may bring execution risk.
- Quanta Computer Inc. (2382.TW)Major supplier of NVL72 and ASIC servers for key CSPs
- Strengths
- It holds important primary-supplier positions in Nvidia systems for customers such as Microsoft, Meta, Amazon, and Google, and has broad coverage in the ASIC server matrix.
- Weaknesses
- It still faces competition from Hon Hai, Jabil, Wiwynn, and others in some customers and architectures.
- Comparison
- Together with Hon Hai, it forms the top-pick group in the NVL72 supply chain.
- Risks
- Second-source expansion, customer platform transitions, and capital spending volatility may affect share.
- Wiwynn Corp (6669.TW)Leading ODM in ASIC servers and a potential second source for VR200
- Strengths
- It has potential share upside as a second source for VR200 at Oracle and Meta, and holds a leading position in ASIC server programs such as T2/T3.
- Weaknesses
- It remains a potential or secondary supplier in parts of the GPU system supply chain.
- Comparison
- The report views Wiwynn and Celestica as leaders in the ASIC server segment.
- Risks
- The pace of second-source onboarding, customer certification timelines, and competitor capacity expansion may affect incremental upside.
- Celestica (CLS)Leading ODM in ASIC servers
- Strengths
- It has design and manufacturing advantages in ASIC server programs such as L6-L11 and benefits from more customized server demand.
- Weaknesses
- Its direct exposure in the primary Nvidia GPU rack supplier matrix is less prominent than that of Hon Hai and Quanta.
- Comparison
- The report lists Celestica and Wiwynn as leaders among ASIC plays.
- Risks
- ASIC project timing, customer concentration, and ODM manufacturing diffusion to second- or third-source suppliers may affect returns.
- WistronSupplier of GPU modules and Dell-related GPU systems
- Strengths
- It has an important position in Nvidia OAM/UBB and AMD MI300/MI400 modules and is deeply tied to Dell's L10 systems.
- Weaknesses
- It is highly dependent on Dell, and Dell is bringing in Compal as an incremental supplier.
- Comparison
- Compared with Hon Hai and Quanta, Wistron is more focused on modules and specific customer supply chains.
- Risks
- Dell supply chain diversification, module design competition, and certification risks for new entrants.
Key data
- 2026 capital spending guidance for the top four U.S. CSPsMore than US$650 billionThe report states this is above about US$415 billion in 2025, with the roughly US$245 billion increase mainly used for higher GPU and memory content.
- Potential capex implied by the FCF ceilingAbout US$750 billion to US$800 billionThe report believes capex budgets may move toward this implied upper limit over the next few quarters.
- NVL72 rack shipmentsAbout 65,000 to 70,000 unitsThe prior year was about 25,000 to 27,000 units.
- Hon Hai NVL72 shareAbout 40%-45%The report says Hon Hai is one of the main NVL72 rack suppliers for major CSPs.
- Quanta NVL72 shareAbout 25%-30%The report says Quanta is also a core NVL72 supplier for major CSPs.
- JPM 2026 estimated total NVL72 shipments68,500 unitsTable 1 shows 28,900 units for Hon Hai, 18,000 for Quanta, 14,000 for Wistron, and 7,600 for Others in 2026.
Impact & implications
For investors, the key driver for AI server ODMs is shifting from simple server assembly to complex system design, module integration, NPI experience, and supply chain execution. Upward revisions to capital spending, NVL72 volume ramp, ASIC platform iterations, and rising server architectural complexity all support leading ODMs in maintaining market share and growing absolute profits.
Risks
- Upward revisions to U.S. CSP capital spending may fall short of expectations, leading to lower-than-expected AI server orders and module demand.
- Server ODM margins may gradually decline as silicon ASPs rise.
- As manufacturing scales up, production may diffuse to second- or third-source suppliers, weakening the market share of leading suppliers.
- New AI architectures such as LPU, CPX, and Vera CPU rack raise the bar for innovation and execution, while also increasing project delivery complexity.
- Customer platform iterations, NPI timing, and uncertainty in supply chain certification may affect revenue recognition and profit realization.
What to watch
- Quarterly results of major ODMs and management guidance on AI server orders.
- Whether visibility on the 2027 CSP capital spending cycle improves.
- Whether NVL72 rack shipments reach the expected roughly 65,000 to 70,000 units.
- Whether Oracle and Meta formally introduce Wiwynn as a second source for VR200.
- The ramp-up pace of TPU/Trillium, T3, and other ASIC server platforms.
- Whether rising server architectural complexity continues to support the design premium and share concentration of leading ODMs.