Chapter 19
Servers, Storage & Compute Platforms
Chapter 19: Servers, Storage & Compute Platforms
19.1 Overview
This is the final assembly layer. The chips designed in Chapter 6, fabricated in Chapter 7, packaged in Chapter 9, connected through the networking silicon of Chapter 10 and the optics of Chapter 11, powered by the systems of Chapters 13-14, cooled by the technology of Chapter 15, and housed in the facilities of Chapters 16-17 must be integrated into complete server systems before they can run AI workloads. This chapter covers the OEMs and ODMs that assemble those servers, plus the PCB/motherboard supply chain and power supply units that are integral to server design.
The global server market reached $112.4 billion in Q3 2025 alone, up 61% year-over-year. GPU-equipped servers grew 49.4% YoY and now account for more than half of total server market revenue. The most dramatic shift in the market is the dominance of ODM Direct sales: in Q3 2025, direct shipments from Original Design Manufacturers to hyperscalers grew 112.2% to $66.8 billion, representing 59.4% of the total server market. Traditional branded OEMs are losing share to this direct channel 1.
Among branded OEMs, Dell Technologies leads with $9.3 billion in Q3 2025 server revenue (+37.2% YoY) but holds only 8.3% market share. Dell reported record Q1 FY2026 (May 2025) servers and networking revenue of $6.3 billion (+16% YoY), and $12.1 billion in AI orders in a single quarter, surpassing all FY2025 AI shipments. Its AI server backlog stands at $14.4 billion 2.
Supermicro has had a turbulent trajectory. After phenomenal growth of 55% YoY in Q4 2024 and a statistical tie with Dell for #1 OEM position, Supermicro lost 13.2% in Q3 2025 and fell to 4.0% market share. Production delays, accounting controversies, and intensifying competition contributed to the decline. However, Supermicro remains a key player in GPU-dense, liquid-cooled server designs and continues to innovate in high-density rack architectures 1.
HPE (Hewlett Packard Enterprise) is struggling more than its peers, recording a 2.3% decline in Q3 2025 and falling to 3.0% market share. HPE’s January 2025 acquisition of Juniper Networks for $14 billion strengthens its networking portfolio but does not address its core server market share erosion. HPE’s Cray line, used for HPC and AI supercomputing, is a differentiated asset, particularly with its CoolIT liquid cooling alliance 1.
The ODM tier is where the volume action is. Quanta Computer, Foxconn (Hon Hai), Wiwynn, Inventec, and Compal/Wistron build the vast majority of servers deployed by hyperscalers. These Taiwanese ODMs design and manufacture custom server platforms to hyperscaler specifications, often incorporating proprietary board layouts, cooling solutions, and power delivery systems. Quanta and Foxconn are the two largest ODMs, collectively representing a significant share of the $66.8 billion ODM Direct segment 13.
The server layer is also where AMD’s ZT Systems acquisition ($4.9 billion, closed 2025 6) becomes relevant. ZT Systems designs and manufactures AI server racks for hyperscalers, giving AMD an integrated platform to pair its EPYC CPUs and Instinct GPUs with custom server hardware. This vertical integration play mirrors NVIDIA’s strategy of selling complete systems (DGX, HGX, GB200 NVL72) rather than just chips (see Chapter 6).
Storage is a secondary but growing requirement. AI training datasets and model checkpoints require high-speed, high-capacity storage. Pure Storage, VAST Data (private), and traditional vendors (Dell EMC, NetApp, HPE) compete in this space. As inference workloads scale, the storage requirements shift from raw capacity to IOPS (input/output operations per second), favoring all-flash NVMe architectures.
19.2 Market Sizing & Growth
Global server market: Q3 2025 revenue $112.4 billion (+61% YoY). GPU servers >50% of total revenue. ODM Direct: $66.8B (59.4% share, +112.2% YoY). Full-year 2024: more than half of revenue from GPU-embedded servers 1.
Dell servers and networking: Q1 FY2026 revenue $6.3 billion (+16% YoY, record). AI orders $12.1 billion in one quarter. AI server backlog $14.4 billion. ISG segment revenue $10.3 billion (+12% YoY) 2.
Supermicro: Q3 2025 market share 4.0% (down from higher). Despite decline, remains important for GPU-dense liquid-cooled designs. Revenue volatile due to production and accounting issues 1.
HPE: Q3 2025 market share 3.0% (-2.3% YoY). Juniper acquisition ($14B) strengthens networking. Cray HPC/AI line with CoolIT liquid cooling is differentiated 1.
ODM Direct: $66.8 billion in Q3 2025 (+112.2% YoY). 59.4% market share. Key players: Quanta Computer, Foxconn, Wiwynn, Inventec. These companies supply directly to AWS, Google, Meta, Microsoft 13.
US data center server market (by power density): AI racks now operate at 50-120 kW per rack. Next-generation racks (Rubin) expected at 200+ kW. Each rack requires custom server chassis, power delivery, and liquid cooling integration 3.
19.3 Supply Chain Flowchart
SERVERS, STORAGE & COMPUTE PLATFORMS
|
|---> NVIDIA REFERENCE DESIGNS (sold as complete systems)
| DGX B200: flagship AI training server
| HGX B200: GPU baseboard for OEM integration
| GB200 NVL72: rack-scale system (72 GPUs, 36 Grace CPUs)
| MGX: modular GPU server platform
| Built by: Foxconn, Quanta, Wistron, Dell, HPE, Supermicro
|
|---> BRANDED OEMs (design + sell under own brand)
| Dell Technologies: PowerEdge XE series (AI-optimized)
| Q1 FY2026 server revenue $6.3B; $14.4B AI backlog
| HPE: ProLiant, Cray (HPC/AI); Juniper acquisition ($14B)
| Supermicro: GPU-dense, liquid-cooled rack-scale systems
| Volatility: production delays, accounting issues
| Lenovo: ThinkSystem AI servers; growing in enterprise
| Cisco: UCS (Unified Computing System); NVIDIA integration
|
|---> ODMs (design + manufacture for hyperscalers directly)
| Quanta Computer (2382.TW): largest server ODM globally
| Foxconn/Hon Hai (2317.TW): servers + GPU racks for NVIDIA, Apple
| Wiwynn (6669.TW): specialized cloud/AI server ODM
| Inventec (2356.TW): GPU server ODM; liquid cooling patents
| Wistron/Wistron NeWeb (3231.TW): server + networking ODM
| Compal Electronics (2324.TW): server ODM
| ZT Systems → AMD (acquired 2025, $4.9B): AI rack-scale systems
|
|---> GPU SERVER ARCHITECTURE (inside the server)
| GPU board: 4-8 GPUs per baseboard (HGX) or 72 per rack (NVL72)
| CPU: AMD EPYC or Intel Xeon or NVIDIA Grace (Arm)
| Memory: HBM on GPU (Chapter 8), DDR5 on CPU
| NVLink: scale-up interconnect between GPUs (Chapter 10)
| PCIe Gen 5/6: CPU-to-GPU, CPU-to-NIC, CPU-to-storage
| Retimers: Astera Labs Aries (Chapter 10)
| Power delivery: 48V to GPU VRMs (Monolithic Power) (Chapter 14)
| Cooling: direct-to-chip liquid cooling (Chapter 15)
| NICs/DPUs: NVIDIA ConnectX/BlueField, Marvell OCTEON (Chapter 10)
|
+---> STORAGE (for AI training datasets and checkpoints)
Pure Storage (PSTG): all-flash NVMe; AI-optimized storage
VAST Data (Private): universal storage for AI pipelines
Dell EMC PowerStore/PowerScale: enterprise AI storage
NetApp: cloud-connected storage for AI workloads
HPE Alletra: hybrid cloud storage
Samsung/SK Hynix/Micron: enterprise SSDs (Chapter 8)19.4 Key Companies
| Company | Ticker | Exchange | Approx. Mkt Cap | Role in Buildout | Key Metric |
|---|---|---|---|---|---|
| Dell Technologies | DELL | NYSE | ~$169B | #1 branded server OEM; PowerEdge AI servers | Q1 FY2026 server revenue $6.3B; AI backlog $14.4B |
| Supermicro | SMCI | NASDAQ | ~$21.2B | GPU-dense, liquid-cooled server specialist | Q3 2025 share 4.0% (volatile); key for NVIDIA designs |
| HPE | HPE | NYSE | ~$41.6B | ProLiant servers, Cray HPC/AI; Juniper acquisition ($14B) | Q3 2025 share 3.0%; Cray liquid-cooled AI differentiated |
| Quanta Computer | 2382 | TWSE | ~$45.0B | Largest server ODM; builds for AWS, Google, Meta | Major share of $66.8B ODM Direct segment |
| Foxconn (Hon Hai) | 2317 | TWSE | ~$70.0B | Second-largest ODM; builds NVIDIA GPU racks, Apple servers | Massive manufacturing scale; AI server revenue growing fast |
| Wiwynn | 6669 | TWSE | ~$20.0B | Specialized cloud/AI server ODM | Key supplier to Microsoft, Meta for GPU servers |
| Inventec | 2356 | TWSE | ~$8.0B | GPU server ODM; strong liquid cooling patent portfolio | Ranked 4th globally in liquid cooling patents |
| Lenovo | 0992 | HKEX | ~$15.0B | ThinkSystem AI servers; growing enterprise AI market | Q3 2025: 3.6% share (+26.1% YoY) |
| Pure Storage | PSTG | NYSE | ~$20.0B | All-flash NVMe storage for AI workloads | Leading AI-optimized storage vendor |
| NetApp | NTAP | NASDAQ | ~$23.0B | Cloud-connected storage for AI/ML data pipelines | All-flash arrays certified in NVIDIA DGX reference architectures; StorageGRID object storage |
| VAST Data | Private | Private | ~$9.0B (last round) | Universal storage platform for AI pipelines | Targeting hyperscaler and enterprise AI data management |
| ZT Systems (AMD) | Private | AMD subsidiary | Acquired $4.9B | AI rack-scale server design + manufacturing | Gives AMD integrated GPU + server platform capability |
| Celestica | CLS | NYSE/TSX | ~$43.2B | Hyperscaler AI server JDM/ODM; optical transceiver manufacturing | Revenue ~$10B; among fastest-growing EMS companies on AI server demand |
| Flex Ltd | FLEX | NASDAQ | ~$53.0B | Major contract manufacturer for AI servers, networking, and power hardware | Revenue ~$26B; serves hyperscaler and enterprise AI hardware programs |
| Sanmina Corp | SANM | NASDAQ | ~$5.0B | High-complexity EMS; AI server PCB fabrication, optical assemblies | Revenue ~$7.5B; US-based manufacturing for defense and enterprise AI systems |
| TTM Technologies | TTMI | NASDAQ | ~$3.0B | Advanced PCB manufacturer for AI servers, 5G infrastructure, aerospace | High-layer-count HDI PCBs required for GPU server motherboards and backplanes |
| Arrow Electronics | ARW | NYSE | ~$11.0B | Largest semiconductor distributor; critical supply chain link between chip makers and OEMs | Revenue ~$28B; aggregates inventory across hundreds of component suppliers for AI hardware builds |
| Avnet | AVT | NASDAQ | ~$7.1B | Second-largest electronic components distributor; supply chain solutions for AI hardware | Revenue ~$23B; provides just-in-time component availability and design support for server OEMs |
The manufacturing and distribution layer between chip designers and deployed AI infrastructure is often invisible but operationally critical. Celestica and Flex manufacture complete AI server systems under JDM contracts with hyperscalers. Sanmina fabricates the high-complexity PCBs and optical subassemblies. TTM Technologies supplies the high-layer-count HDI printed circuit boards that GPU server motherboards require; each successive GPU generation demands more power delivery layers and higher signal integrity, pushing PCB complexity upward. Arrow and Avnet aggregate components from hundreds of suppliers and provide just-in-time inventory positioning that determines whether server OEMs can scale production at the pace hyperscalers require. A component shortage at the distributor level can delay server assembly even when GPUs are available.
19.5 Bottleneck Analysis
GPU allocation determines server shipments (SEVERE): Server OEMs and ODMs can only ship AI servers as fast as they receive GPUs from NVIDIA (and to a lesser extent AMD). NVIDIA allocates GPU supply based on long-term relationships and prepayment agreements. Dell’s AI backlog grew from $14.4 billion (Q1 FY2026) to $43 billion by the end of FY2026 after the company booked $64.1 billion in cumulative AI orders. Dell Vice Chairman Jeff Clarke described “the supply environment is as tight as we have ever seen, and input costs are moving higher,” with DRAM spot prices up 5.5x in six months 7. The server layer is derivative of the chip layer; its bottleneck is upstream 2.
ODM concentration in Taiwan (MODERATE-HIGH): Quanta, Foxconn, Wiwynn, Inventec, and Wistron are all headquartered in Taiwan. While they operate manufacturing facilities globally (including China, Vietnam, Mexico, and the US), their engineering and management are concentrated on the island. A Taiwan Strait crisis would disrupt the entire server supply chain. This risk is partially mitigated by geographic diversification of manufacturing, but core design teams remain in Taiwan 3.
Liquid cooling integration complexity (MODERATE): Every new GPU generation requires redesigned server chassis with updated cold plates, manifolds, and coolant connections. The transition from air-cooled to liquid-cooled servers adds 30-50% to server cost and introduces reliability concerns (leaks, corrosion, maintenance). Supermicro’s competitive advantage in rapid liquid-cooled server design has been an important differentiator, though its execution challenges show the difficulty of scaling this capability 1.
Server power delivery at 100+ kW per rack (MODERATE): A GB200 NVL72 rack draws approximately 120 kW. Delivering this much power reliably to a single rack requires heavy-gauge power cables, high-current busbars, and custom power distribution. Standard data center power infrastructure was designed for 5-15 kW racks. The 10x increase in per-rack power consumption requires corresponding upgrades across the power chain (see Chapters 14-15).
19.6 Risks
ODM disintermediation of branded OEMs: As hyperscalers become more sophisticated in server design, they increasingly bypass Dell, HPE, and Supermicro to work directly with ODMs. Dell’s 8.3% and HPE’s 3.0% market shares are small and shrinking relative to the 59.4% ODM Direct segment 1. If this trend continues, branded OEMs become marginalized in the AI server market, surviving mainly in enterprise. Dell’s counter is to offer integrated solutions (servers + storage + networking + services) that ODMs cannot match for enterprise customers.
Supermicro execution risk: Supermicro’s accounting controversies, auditor changes, and production delays have eroded market confidence despite strong products. If these issues persist, hyperscalers may reduce reliance on Supermicro in favor of Dell, HPE, or direct ODM relationships. The company’s specialized expertise in GPU-dense liquid-cooled designs is valuable, but execution discipline is non-negotiable for mission-critical infrastructure.
Inference shifts server requirements: Training clusters require maximum GPU density and NVLink bandwidth. Inference workloads may require different server architectures: more CPUs, less GPU memory, lower power per unit. If inference becomes the dominant workload (as many predict), the optimal server design changes, potentially favoring cost-optimized configurations over the current GPU-maximized designs. This could benefit lower-cost ODMs and custom ASIC-based platforms over premium GPU servers.
First principles check: Why does the server layer matter if it is just assembly? Because integration is harder than it looks. A GB200 NVL72 rack contains 72 GPUs, 36 CPUs, 5,184 copper cables, hundreds of optical transceivers, liquid cooling manifolds, and custom power delivery. The tolerances for thermal management, signal integrity, and power delivery are extreme. A poorly assembled server can cause GPU throttling, network errors, or even hardware failure. The ODMs and OEMs in this chapter have decades of experience in high-reliability systems integration. This is not consumer electronics assembly; it is mission-critical infrastructure engineering.