Nvidia’s AI servers, such as the DGX and HGX systems, rely on a global supply chain for key components including memory (HBM and GDDR), networking silicon, power management ICs, and cooling solutions. While Nvidia designs the core GPU, CPU (Grace), and NVLink interconnects, it sources memory from SK Hynix, Samsung, and Micron; networking from its own Mellanox division plus optical module makers; and power/cooling from a range of specialized suppliers. This article details the main component categories, the suppliers behind them, and how the supply chain shapes AI server availability.
Background: Why Nvidia Doesn’t Make Everything
Nvidia is a fabless semiconductor company—it designs chips but outsources manufacturing to TSMC and Samsung. Similarly, it doesn’t produce memory, power electronics, or cooling hardware. Instead, it specifies the components and certifies suppliers through its partner ecosystem. This model lets Nvidia focus on architecture and software (CUDA, TensorRT) while leveraging the scale of specialized manufacturers. The AI server boom has made these suppliers critical: a shortage of any single component—like HBM memory—can delay entire server lines.
Core Silicon: GPUs, CPUs, and Interconnects
The heart of an Nvidia AI server is the GPU (e.g., H100, H200, A100) and, for Grace-Hopper systems, the Grace CPU. These are designed by Nvidia but fabricated by TSMC (and Samsung for some older nodes). The GPU die is packaged with HBM memory stacks on a silicon interposer, a process that requires advanced packaging from TSMC’s CoWoS lines. Nvidia also designs the NVLink and NVSwitch chips that connect GPUs, but these are also fabbed by TSMC.
Memory: HBM and GDDR from SK Hynix, Samsung, Micron
The most critical memory component is High Bandwidth Memory (HBM), used in all data-center GPUs. Nvidia sources HBM3 and HBM3e from three suppliers: SK Hynix (the lead supplier), Samsung, and Micron. Each has passed Nvidia’s rigorous qualification. The HBM stacks sit next to the GPU die and are essential for AI training performance. For lower-end or inference-oriented GPUs, Nvidia may use GDDR6 memory from the same trio, but HBM is the bottleneck in high-end servers.
Networking: Mellanox, Optical Modules, and Cables
Nvidia acquired Mellanox in 2020, giving it in-house InfiniBand and Ethernet networking chips. However, the actual optical transceivers, cables, and connectors come from external suppliers like Coherent, II-VI (now Coherent), Lumentum, and various Taiwanese and Chinese makers. These components are crucial for scaling AI clusters across thousands of GPUs. Nvidia’s NVLink also uses copper cables from suppliers like Amphenol and TE Connectivity.
Power and Cooling: From PSUs to Liquid Cooling
AI servers consume enormous power—an H100 GPU can draw up to 700W. Power supply units (PSUs) are sourced from Delta Electronics, Lite-On, and Chicony. Voltage regulators and power management ICs come from TI, Infineon, and MPS. For cooling, air-cooled servers use fans from Delta and AVC, while liquid-cooled systems (like the NVL72 rack) rely on cold plates, manifolds, and quick disconnects from suppliers such as Boyd, Cooler Master, and AVC. These components are often custom-designed to Nvidia’s specs.
How to Choose a Supplier: Nvidia’s Qualification Process
Nvidia doesn’t publicly list every supplier, but its certification process is known to be rigorous. Suppliers must:
- Meet performance specs (e.g., HBM bandwidth, thermal resistance).
- Pass reliability tests (thermal cycling, vibration, long-duration stress).
- Demonstrate manufacturing capacity to scale with Nvidia’s demand.
- Comply with export controls and supply-chain security requirements.
For buyers, this means that components from certified suppliers are generally interchangeable, but Nvidia often recommends specific SKUs for optimal performance.
Common Problems and Misconceptions
A common misconception is that Nvidia makes its own memory or networking gear. In reality, it relies on partners. Another issue is the assumption that all HBM is the same—it’s not. HBM3e from different suppliers may have slight timing differences, but Nvidia’s firmware handles them. Supply-chain bottlenecks, like the 2023 CoWoS capacity crunch, show how dependent Nvidia is on TSMC’s advanced packaging. Also, liquid cooling parts are not standardized, so mixing suppliers can cause leaks or performance issues.
Cost and Availability Trends
Component costs are not public, but HBM is known to be a significant cost driver—HBM3e can cost several times more than GDDR per GB. The overall BOM of an H100 server is estimated to be around $200,000, with memory and networking being major portions. Availability is often constrained by HBM supply and CoWoS capacity, leading to long lead times. Nvidia mitigates this by signing long-term agreements and pre-paying for capacity.
Alternatives and Future Directions
Nvidia could vertically integrate more, but it hasn’t. Instead, it’s diversifying suppliers (e.g., adding Samsung as a second HBM source) and designing custom interconnects. For buyers, alternatives include AMD’s MI300 series (which uses its own supply chain) or Intel’s Gaudi. However, Nvidia’s software ecosystem remains a key differentiator. Future servers may use more in-house networking chips, but memory and power will likely remain outsourced.
Frequently Asked Questions
Q: Does Nvidia manufacture its own memory?
No. Nvidia designs the GPU but sources HBM and GDDR memory from SK Hynix, Samsung, and Micron.
Q: Who makes the CPUs in Nvidia’s Grace servers?
The Grace CPU is designed by Nvidia but fabricated by TSMC. It’s based on ARM architecture.
Q: Are the networking chips in Nvidia servers made by Nvidia?
Yes, after the Mellanox acquisition, Nvidia designs its own InfiniBand and Ethernet chips, but optical modules and cables come from external suppliers.
Q: Why is HBM memory so critical and scarce?
HBM offers far higher bandwidth than GDDR, essential for AI training. Its complexity (stacked DRAM dies) limits production yield, causing shortages.
Q: Can I buy Nvidia AI server components separately?
Generally no. Nvidia sells complete servers or boards, not individual HBM or PSUs, to end users. Components are available only through OEM partners.
Related
Method: the standard formula for this calculation, rounded as each result states. See how the calculators work or report an error.
More cards in this drawer
- How to Choose an Amplifier for Magico A3 Speakers
- How to Ban a Player from Your FiveM Server
- Pokemon Go Friend Codes: How to Add & Use Them33.33 XP/day
- Google Algorithm Change: Original Reporting Boost
- Create Folder for Selected Item on Mac: Full Guide
- Data Center Location Questions for AI Workloads