One AI server, from the GPU at its core to the rack on the outside, depends on a whole row of Taiwanese companies. The most advanced AI chips are designed in the US, and memory is led by the US and South Korea. But putting all the parts together into a working AI server is Taiwan’s home turf.
This piece extends the Taiwan semiconductor supply chain to the server-system layer and splits a whole AI server into seven links. If you only want the narrower assembly group, read ODM stocks and AI server assembly.
What an AI Server Is Made Of
Today’s AI server is mostly a “full rack-scale system.” Take NVIDIA’s GB300 NVL72: a single rack integrates 72 GPUs and 36 CPUs, plus switches, power, and cooling, all tied together with liquid cooling. Pull it apart and you get roughly six layers:
- Compute: GPU, CPU, HBM
- Board-level: motherboard, printed circuit board (PCB), IC substrate, connectors
- Interconnect: NVLink, InfiniBand, Ethernet switches
- Power: power supplies, busbars, backup batteries
- Cooling: cold plates, manifolds, quick disconnects, cooling distribution unit (CDU)
- Mechanicals: rack, chassis, trays, rails, cable harnesses
Taiwan has a representative supplier in almost every layer. That is why one AI server can move so many concept stocks.
Core-Data Snapshot
| Topic | Value | Timing / Nature |
|---|---|---|
| 2026 global AI server shipment growth | Nearly 31% year on year | TrendForce forecast, raised in August 2026 |
| Taiwan’s share of global AI server production | About 90% (production/assembly basis) | MIC / Ministry of Economic Affairs estimate |
| ASIC-based AI server share | About 27% in 2026; GPU systems still the majority | TrendForce forecast, April 2026 |
| GB300 NVL72 per-rack power | About 130 to 142 kilowatts | System-maker specs, not published by NVIDIA |
| Vera Rubin NVL72 per-rack power | About 240 kilowatts (Gigabyte solution) | System-maker spec, varies by configuration |
| AI chip liquid-cooling penetration | About 33% in 2025 → about 53% in 2026 → nearly 60% in 2027 | TrendForce forecast, August 2026 |
Taiwan’s Division-of-Labor Map: Seven Links
These are industry-role descriptions. Being named does not mean a firm has won orders or will benefit.
| Link | What they do | Representative Taiwanese firms |
|---|---|---|
| System / rack assembly (ODM) | Integrate GPU, CPU, power, and cooling into systems and ship full racks | Hon Hai/Foxconn, Quanta, Wistron, Wiwynn, Inventec, Gigabyte |
| Cooling / liquid cooling | Cold plates, manifolds, quick disconnects, CDU, vapor chambers | Asia Vital Components, Auras, Jentech |
| Power | Power supplies, busbars, power shelves | Delta Electronics, Lite-On |
| Networking switches | Data center switches, white-box networking | Accton |
| Chassis / mechanicals | Racks, chassis, trays | Chenbro, AIC |
| Rails / connectors | Server rails, high-speed connectors, cabling | King Slide, BizLink, Lotes, ACES |
| PCB / substrate | Printed circuit boards, copper-clad laminates, ABF substrates | Gold Circuit, Zhen Ding, Elite Material, Unimicron, Nan Ya PCB, Kinsus |
When NVIDIA announced Vera Rubin was in full production in May 2026, its list of system-maker partners included Taiwanese firms such as Hon Hai, Quanta’s QCT, Wistron, Wiwynn, Inventec, and Gigabyte.
From GB300 to Vera Rubin: What’s Changing in the Supply Chain
NVIDIA began volume shipments of Vera Rubin in August 2026, and Hon Hai said it would start volume production of Vera Rubin racks in Q3 2026. Even so, most shipments still come from the Blackwell generation, which includes GB300. A generational shift moves the supply chain’s center of gravity, and this round brings three changes.
Liquid cooling becomes standard: a GB300 rack already draws over 100 kilowatts, too much for air cooling. Vera Rubin NVL72 goes further with a fanless, fully liquid-cooled design. Cold plates, CDUs, quick disconnects, and leak detection all matter more.
Power moves to higher voltage: power density is pushing the power architecture to higher voltages (such as 800VDC), and power shelves, busbars, and backup batteries are being upgraded with it. NVIDIA’s 800VDC partner list includes Taiwanese firms such as Delta, Lite-On, and BizLink, and NVIDIA plans full adoption in its next-generation racks in 2027.
The ODM role is changing: NVIDIA keeps moving toward full-rack solutions. Part of the ODM’s work is shifting from motherboard design to rack-level integration, yield, and cooling/power integration. That is an opportunity for Taiwanese firms, and also margin pressure. BOM (bill of materials) positions change across generations, so “having a related product” is not the same as “shipping at volume.”
How to Read “AI Server Stocks”
The market puts companies from all seven links above into the AI server theme. What matters is where each company sits and what it does. Then, when you read the news, you can quickly tell whether it affects cooling, power, or system assembly.
Keep two things in mind. “About 90% from Taiwan” is a production/assembly figure and does not mean the profit stays with Taiwanese firms. Real customers and orders are mostly confidential, and the pairings the market draws are often analyst guesswork. This piece only describes who does what. It does not rank stocks and is not investment advice.
For the supply chain built around NVIDIA, read the NVIDIA supply chain. For why cooling is so critical, read the liquid-cooling gate. For all eight gates of the chain, go back to the supply-chain overview.
Penchan’s summary
AI servers are the part of Taiwan’s AI supply chain with the broadest participation. From system and rack assembly to cooling, power, networking, mechanicals, connectors, and substrates, every link has representative firms.
From GB300 to Vera Rubin, liquid cooling has become standard, while high-voltage power is heading toward the next rack generation. Use this map to read the news, not as a stock-picking list.