Free tools Windows power users keep installed
One-click scans. No signup required.
iTechGuides is reader-supported. When you buy through links on our site, we may earn an affiliate commission. As an Amazon Associate I earn from qualifying purchases. Learn more
NVIDIA AI servers depend on suppliers and manufacturing partners for far more than the GPUs. The supply chain includes semiconductor fabrication, memory, substrates, advanced packaging, testing, optical components, boards, networking, storage, power, cooling, chassis and system assembly. NVIDIA designs key compute and interconnect technologies, while external partners manufacture components and OEMs or ODMs integrate them into usable systems.
NVIDIA’s October 2026 announcement names six partners supporting capacity expansion in substrates, advanced packaging, semiconductor testing and optical engines. It also says Grace Blackwell platforms involve millions of parts, thousands of suppliers and dozens of OEMs and ODMs. The company does not publish a complete supplier-by-supplier list for every component in every server, so the names below identify announced partners, not an exhaustive bill of materials.
Which parts of an NVIDIA AI server come from suppliers?
A complete NVIDIA-based AI server is a stack of components, not a GPU in a box. NVIDIA supplies major accelerator and interconnect designs, but the finished system also needs processors, memory, networking, storage, power, cooling, management hardware and a chassis. Suppliers make many of those parts; an OEM or ODM selects and integrates them into a validated server.
The most useful way to understand the supply chain is to separate the work into three layers: manufacturing components, assembling a compute platform, and integrating that platform into a complete server or rack. The supplier for an individual part can differ by product generation, location and OEM, and NVIDIA’s public materials do not identify every source.
#1 Best Overall
- The H100 NVL graphics card is designed to scale the support of large language models, such as GPT3-175B, in mainstream PCIe-based server systems, providing up to 12X the throughput performance of HGX A100 systems when configured with 8 units.
- Equipped with advanced features, including 94GB of high-speed HBM3 memory, NVLink connectivity for enhanced inter-GPU communication, and an impressive memory bandwidth of 3938 GB/sec, the H100 NVL is built for high-performance AI inference tasks.
- The card showcases a robust performance spectrum across various compute types: 68 TFLOPS for FP64, 134 TFLOPS for both FP64 Tensor Core and FP32, escalating up to 7916 TFLOPS/TOPS for FP8 and INT8 Tensor Core operations, all benefiting from sparsity optimizations.
- It enables standard mainstream servers to deliver high-performance capabilities for generative AI inference, simplifying the deployment process for partners and solution providers with fast time to market and ease of scalability.
- The H100 NVL's power efficiency is optimized with a configurable maximum power consumption ranging between 2x 350-400W, supporting extensive computational tasks without excessive power usage.
| Layer | What it contributes | What the public information establishes |
|---|---|---|
| Component manufacturing | Semiconductor production, memory, substrates, packaging, testing, optical parts, boards, storage, power and cooling equipment. | NVIDIA identifies selected partners for substrate capacity, advanced packaging, testing and optical engines; it does not publish a complete supplier list for every component. |
| Platform integration | Combines GPUs with CPUs, system memory, interconnects, network adapters and management hardware. | DGX and HGX configurations show representative components, but the platforms are not interchangeable product labels. |
| Complete-system build | Adds chassis, cooling, power delivery, firmware, rack design, validation and service. | DGX is an integrated NVIDIA system specification. HGX is a platform that OEMs and channel partners use to produce complete servers. |
What does NVIDIA design, and what do outside companies manufacture?
NVIDIA’s role includes the accelerator and interconnect technologies that define the compute platform. Those designs still rely on a broader manufacturing chain: semiconductor fabrication and packaging turn designs into usable chips and modules, while other suppliers provide the supporting parts required to run them in a server.
In its October 2026 announcement, NVIDIA named Amkor, IBIDEN, King Yuan Electronics (KYEC), Samsung Electro-Mechanics, Siliconware Precision Industries (SPIL) and Unimicron in connection with capacity expansion for substrates, advanced packaging, semiconductor testing and optical engines. The announcement identifies these companies as partners in specific supply-chain activities; it does not establish which named company supplies every part in a particular server or how much of NVIDIA’s total supply each provides.
NVIDIA’s technical supply-chain article describes Grace Blackwell platforms as drawing on millions of parts and thousands of suppliers, with final systems built by dozens of OEMs and ODMs. That scale helps explain why “Who makes an NVIDIA server?” does not have a single-company answer: a platform design, a manufactured component and a finished server can each involve different organizations.
Which suppliers does NVIDIA name for substrates, packaging, testing and optical engines?
The October 2026 announcement names six partners and gives three capacity targets. These are announced plans, not evidence that the targets have already been reached or that each partner supplies all NVIDIA systems.
| Company or group | Activity identified by NVIDIA | Announced capacity figure | Qualification |
|---|---|---|---|
| Amkor | Interposer-wafer capacity | More than 300% growth | Targeted by the end of 2028, according to NVIDIA’s October 2026 announcement. |
| SPIL | Optical-engine capacity | 15x increase | Targeted by mid-2027, according to NVIDIA’s October 2026 announcement. |
| Substrate partners | Substrate capacity expansion | 2.5x increase | NVIDIA’s planned increase over three years; the announcement names IBIDEN, Samsung Electro-Mechanics and Unimicron among the partners. |
| King Yuan Electronics (KYEC) | Semiconductor testing | Not stated | The announcement names KYEC as a partner but provides no corresponding capacity multiple in the cited material. |
| Siliconware Precision Industries (SPIL) | Advanced packaging and optical engines | 15x increase for optical-engine capacity | The optical-engine target is specifically tied to mid-2027; no separate advanced-packaging capacity figure is stated. |
The announcement groups partners across capacity expansion rather than providing a detailed allocation of every process to every supplier. It therefore supports identifying the partner network and announced targets, but not assigning a particular company to a specific GPU’s complete manufacturing path.
What is inside a DGX H100 or H200 server?
NVIDIA’s DGX H100/H200 product guide describes an integrated system with eight H100 or H200 GPUs and two Intel Xeon CPUs. Its listed contents show how an AI server combines NVIDIA compute hardware with third-party processor, memory, networking, storage and power components.
| Component | DGX H100/H200 configuration in NVIDIA’s 2025 guide | What it does in the system |
|---|---|---|
| Accelerators | Eight H100 or H200 GPUs | Provide the main GPU compute resources. |
| Host processors | Two Intel Xeon CPUs | Handle host-side processing and coordinate system activity. |
| GPU interconnect | Four NVSwitch units | Provide the platform’s GPU interconnect layer. |
| Networking | ConnectX-7 networking cards | Connect the server to the relevant network fabrics. |
| System memory | 2 TB | Provides system memory alongside GPU memory. |
| Storage | NVMe storage | Provides local high-speed storage; the cited guide summary does not state its capacity. |
| Management | Baseboard management controller (BMC) | Supports system management functions. |
| Power | Six 3.3 kW power supplies | Supply power to the system. This listed rating is not a statement of typical operating consumption. |
These are specifications for the DGX H100/H200 system described in NVIDIA’s 2025 product documentation, not universal specifications for every server using H100 or H200 GPUs. Other manufacturers’ systems can differ in chassis, component selection, storage configuration, power design, cooling and service arrangements.
Rank #2
- Graphics Card Interface: Pci E
How does an HGX platform differ from a DGX server?
DGX is an integrated NVIDIA system specification; HGX is a platform that OEMs and channel partners turn into complete servers. That distinction matters when identifying who supplies the final chassis, cooling, firmware, service and other system-specific parts.
NVIDIA’s HGX AI Factory node configuration lists eight GPUs, one BlueField-3 DPU, eight BlueField-3 SuperNICs, two CPUs, at least 1.5 TB of system memory, NVMe storage and management hardware. NVIDIA’s reference architecture says hardware support is delivered through the fulfilling OEM and channel partners. The listed configuration describes a platform design, not a guarantee that every HGX server has identical parts or implementation.
For an HGX system, the OEM’s implementation can affect the exact chassis, cooling, firmware, service, component sourcing and rack design. A buyer comparing systems should therefore check the actual vendor specification and support terms rather than treating “HGX” alone as a complete server model.
What networking and interconnect parts does an NVIDIA AI server need?
The networking layer is more than the adapter installed in the server. NVIDIA identifies BlueField-3 SuperNICs, BlueField-3 DPUs, ConnectX-7 SmartNICs, switches, transceivers and cables as elements used across compute, storage, customer and management fabrics.
Recommended Free Tools
- Inside the server: GPUs communicate through the platform’s GPU interconnect; the DGX H100/H200 guide lists NVSwitch units. Network adapters or SuperNICs connect system resources to external fabrics.
- At the server boundary: A DPU or network adapter provides the server’s connection to a chosen network architecture. The HGX configuration lists BlueField-3 DPUs and SuperNICs, while DGX H100/H200 lists ConnectX-7 networking cards.
- Across the rack or cluster: Switches, transceivers and cables form the links among servers and to storage or other networks. The appropriate design depends on the intended compute, storage, customer and management fabrics.
“NVIDIA networking” does not identify every item in a deployment. A procurement specification still needs to state the selected fabric, switch and cabling design, as well as the adapter configuration and how the system is validated by its OEM.
Who supplies memory, storage, power, cooling and the chassis?
NVIDIA’s public component lists establish that these categories are part of the systems, but they do not provide a complete, product-by-product supplier roster for them. A server buyer should distinguish between a component category named in a configuration and a named company confirmed as its supplier.
- Memory: DGX H100/H200 documentation lists 2 TB of system memory; the HGX AI Factory node configuration specifies at least 1.5 TB. These figures refer to different configurations and should not be treated as a single standard for every NVIDIA server.
- Storage: Both configuration descriptions include NVMe storage. The cited summaries do not give a universal capacity or identify a supplier, so those details must be checked in the system’s own specification.
- Power: The DGX guide lists six 3.3 kW power supplies. That equipment rating is not the same as the server’s expected draw in a particular workload or facility.
- Cooling and chassis: These are part of making the platform a deployable system, but the exact design can vary with OEM, generation and rack configuration. The cited public material does not identify one supplier that applies to all systems.
- Management: DGX H100/H200 includes a BMC, and the HGX node configuration includes management hardware. The actual implementation is part of the system configuration rather than a universal supplier assignment.
Why do supplier capacity and system integration affect availability?
AI-server production depends on capacity across several linked stages, not only on the supply of GPUs. Substrates, advanced packaging, testing and optical components are among the activities for which NVIDIA announced expansion plans in October 2026. Capacity growth at one stage does not by itself establish that every other stage can meet demand or that finished systems will be available immediately.
Rank #3
- Discrete graphics card memory 40 GB
- Memory bandwidth (max) 1555 GB/s
- Graphics processor family NVIDIA
- Graphics processor A100
NVIDIA CEO Jensen Huang said, “Capacity is the constraint,” on October 21, 2026. In practical terms, a customer may need a complete validated system, not merely an accelerator: the system also requires compatible memory, networking, storage, power, cooling and integration. The OEM or ODM must assemble and support that combination, and rack-level deployment adds further design dependencies.
Quick wins for a faster PC:
Repair Windows errors before they cause bigger problemsFix Now →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →The announced targets are forward-looking milestones: NVIDIA planned to raise substrate capacity 2.5x over three years, Amkor targeted more than 300% growth in interposer-wafer capacity by the end of 2028, and SPIL targeted a 15x increase in optical-engine capacity by mid-2027. They indicate where capacity is being expanded, but do not quantify the total availability of H100, H200 or other systems.
What should buyers check when comparing NVIDIA GPU servers?
Use the platform name as a starting point, then compare the actual configuration and the organization responsible for delivering and supporting it. Two systems built around NVIDIA GPUs can differ in ways that affect workload fit, facility requirements and service.
- Confirm the GPU generation and memory. Check the exact GPU model and memory configuration in the vendor’s specification; do not infer it from a broad platform label.
- Inspect the GPU topology. Verify how the system implements NVLink and NVSwitch, including the actual number and arrangement of interconnect components.
- Identify the network fabric. Compare the server adapters or DPUs with the planned Ethernet or InfiniBand fabric, switches, transceivers and cables.
- Check host and system memory. Confirm the CPU model and system-memory amount for the specific configuration rather than relying on another product’s example.
- Review storage and management provisions. Ask for storage capacity and the management hardware or features included in the delivered system.
- Validate facility fit. Confirm power requirements, cooling approach, chassis dimensions and rack density with the system vendor and data-center operator.
- Establish responsibility for validation and support. Identify the fulfilling OEM or channel partner and confirm the applicable system validation, service and support terms.
This checklist is especially important when comparing a DGX system with an OEM-built HGX server: the GPU platform may share NVIDIA technologies, while the complete-system specification and support relationship differ.
Frequently Asked Questions
Does NVIDIA make its own AI server GPUs?
NVIDIA designs its accelerator technologies, but semiconductor production and later manufacturing steps depend on external partners. The public partner announcement discussed here names capacity support in selected stages, not the full fabrication chain for each GPU.
The Tool Desk
Outbyte PC Repair FREERepair Windows errors before they cause bigger problemsFix Now →Outbyte Driver Updater FREEScan for outdated or missing drivers - takes under a minuteDriver Scan →Does the DGX H100/H200 configuration use AMD CPUs?
No. NVIDIA’s 2025 DGX H100/H200 guide lists two Intel Xeon CPUs. That detail applies to the documented DGX configuration, not all NVIDIA GPU servers.
Can I identify a server’s component suppliers from the NVIDIA platform name?
Usually not. A platform label identifies a design or system family, but the OEM’s detailed specification and support documentation are needed to establish the delivered parts and responsible integrator.
Are NVIDIA’s announced supplier capacity increases guaranteed to reduce server wait times?
The announcements describe planned capacity targets at particular manufacturing stages, not delivery forecasts for finished servers. Availability also depends on other component capacity, system integration and the OEM’s ability to deliver a validated configuration.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
