Skip to content 
AI data centers are different from traditional data centers in one important way: the amount of computing and data moving through the system is much higher.
GPUs, AI accelerators, HBM, high-speed switches, and networking devices all have to handle large amounts of data. On the PCB side, that means more high-speed signals, higher power density, tighter routing, and more demanding material and manufacturing requirements.
This is why AI server PCBs have become a major area of interest in the PCB industry.
In the first quarter of 2026, Taiwan’s PCB industry reached NT$245.6 billion in output, up 19.6% year over year. Industry data from TPCA and ITRI pointed to continued AI server demand as a driver for high-layer-count and HDI PCB products.
The change is not simply that AI data centers need more PCBs. The bigger change is that the PCBs themselves are becoming more complex.
Why AI Data Centers Need More PCBs
The PCB demand created by AI data centers is not limited to GPU server mainboards.
A typical AI computing system can include:
- GPU / AI accelerators
- CPUs
- HBM
- High-speed networking
- Ethernet / InfiniBand
- Storage
- Power systems
- Switches
- Backplanes
All of these systems need electrical connections, and many of them need to move data at very high speeds.
AI servers also tend to pack more accelerators into each system and rack. As computing density increases, the PCB has to handle more signals and more power connections within roughly the same physical space.
So the PCB demand created by AI infrastructure extends across the hardware system rather than being limited to one particular board.
How AI Server PCBs Differ From Conventional Server PCBs
From a PCB manufacturing perspective, the biggest differences are fairly straightforward: more layers, higher speeds, greater density, and higher power.
| Parameter | Conventional Server PCB | AI Server PCB |
|---|---|---|
| Main computing device | CPU | GPU / AI Accelerator |
| Layer count | Medium to high | High |
| Signal rate | Relatively lower | 56G / 112G / 224G and beyond |
| Materials | Standard FR-4 / mid-loss materials | Low-loss / ultra-low-loss materials |
| Routing | Conventional multilayer routing | High-density, high-speed routing |
| Vias | Mainly through vias | HDI / Microvias / Via-in-Pad |
| Impedance | Controlled impedance | Tighter impedance control |
| Power density | Relatively lower | Higher |
| Thermal requirements | Conventional | More demanding |
These differences show up directly in PCB manufacturing.
A conventional server board may mainly need to deal with layer count, impedance, and reliability. An AI server PCB may also need microvias, back drilling, low-loss materials, high-speed SerDes routing, and heavier copper for power delivery.
For a deeper look at the board-level requirements, AI Server PCB: Design Requirements, Materials & Manufacturing covers stackup, materials, signal integrity, power delivery, and manufacturing considerations in more detail.
Higher Layer Counts Are Becoming Common in AI PCBs
AI servers need to connect more devices while providing enough routing space for high-speed signals, power, and ground.
One of the most direct ways to create that space is to add more PCB layers.
Industry analysis from EMSG shows that PCB layer counts are increasing as AI server platforms become more complex. GB200 and GB300 platforms involve boards with more than 20 layers and, in some cases, more than 30 layers, while parts of next-generation platforms are moving toward 40 layers and beyond.
The difficult part is not simply adding more layers.
A 30- or 40-layer PCB has to deal with:
- Stackup
- Lamination
- Layer registration
- Copper balance
- Drilling
- Plating
- Impedance control
- Board thickness
- Warpage
As the number of layers increases, dimensional variation can accumulate through multiple manufacturing steps.
That makes high-layer-count PCB manufacturing a different challenge from producing a conventional multilayer board.
Why HDI and Microvias Are Increasingly Used in AI PCBs
AI processors commonly use high-density BGA packages.
If a board relies entirely on conventional through vias, routing space around a dense BGA can quickly become a problem.
This is one reason HDI technology is increasingly used in AI-related PCB designs.
Common structures include:
- Laser microvias
- Blind vias
- Buried vias
- Stacked microvias
- Via-in-pad
- Sequential lamination
Microvias allow connections between adjacent layers to be made in a much smaller area. Via-in-pad can also provide a direct connection from a BGA pad to an inner routing layer.
For more complex HDI structures, designs may use configurations such as 1+N+1, 2+N+2, or 3+N+3, depending on the package and routing requirements.
The HDI PCB Design Guide covers the relationship between HDI stackup, microvias, routing, signal integrity, and manufacturing.
There is also a practical manufacturing issue here. Smaller vias do not automatically mean better results. Laser drilling, copper filling, plating, layer registration, dielectric thickness, and thermal cycling all have to work together.
High-Speed Signal Requirements for AI PCBs
This is another area where AI server PCBs are moving away from conventional server boards.
GPUs, CPUs, HBM, and high-speed networking devices need to move large amounts of data. AI server designs commonly involve 56G, 112G, 224G PAM4, PCIe, NVLink, Ethernet, and other high-speed interfaces.
At these data rates, PCB geometry becomes part of the transmission system.
Controlled Impedance
Impedance cannot be treated as something that only exists in the PCB design software.
During fabrication, dielectric thickness, copper thickness, trace width, etching, and material properties all affect the final impedance.
That is why impedance control needs to start with the stackup and continue through PCB fabrication and testing.
SLLPCB’s High-Speed PCB Design Guidelines covers controlled impedance, stackup design, low-loss materials, differential routing, and other signal-integrity considerations.
Insertion Loss and Return Loss
As signal speeds increase, PCB loss becomes harder to ignore.
Dielectric loss, copper roughness, trace length, and transitions can all affect signal quality.
This is one of the reasons low-loss materials are increasingly used in high-speed AI PCB designs.
Via Stubs and Back Drilling
When a high-speed signal passes through a conventional through via, the unused section of the via can form a stub.
At high data rates, that stub can create an unwanted discontinuity.
Back drilling can remove the unused portion of the via and reduce its effect on the signal path.
It adds another manufacturing step, but accurate depth control and registration are particularly important on high-layer-count boards.
Why AI Server PCBs Use More Low-Loss Materials
Standard FR-4 is not automatically unsuitable for AI data centers.
The need for higher-performance materials usually depends on where the high-speed channels are and how much loss the system can tolerate.
At higher data rates, Dk, Df, copper roughness, dielectric thickness, and the overall stackup all affect signal performance.
For this reason, high-speed AI PCB designs may use low-loss or ultra-low-loss laminates.
Material grades such as M6, M7, M8, and M9 are increasingly discussed in connection with high-speed computing and networking applications.
EMSG’s industry analysis also highlights M7+ CCL and materials such as T-glass and NER-glass in the AI PCB supply chain.
The choice is not necessarily to use the highest-grade material on every layer.
A practical stackup may use higher-performance materials around critical high-speed channels while using more conventional materials elsewhere. The right combination depends on data rate, channel length, loss targets, stackup, and cost.
Higher AI Server Power Also Puts More Pressure on the PCB
AI accelerators consume a lot of power, and that puts additional pressure on the PCB power delivery system.
The board may need to handle:
- Power planes
- Ground planes
- Copper thickness
- Via current capacity
- Decoupling
- Thermal vias
- Copper balance
- Board warpage
Thicker copper and larger power planes can help carry higher current, but they also make etching, plating, lamination, and dimensional control more difficult.
Thermal management is another part of the same problem.
High-power processors generate significant heat, so the PCB may need thermal vias, large copper areas, heatsinks, cold plates, or other thermal structures as part of the overall system.
Large BGA packages also make board flatness more important. Uneven copper distribution and thermal cycling can contribute to warpage and reliability problems.
For AI power modules and dense compute boards, PCB-level thermal paths are becoming a more practical design consideration rather than something left entirely to the system level.
Why AI PCBs Are More Difficult to Manufacture
The difficult part is that many of these requirements can appear on the same board.
For example, an AI server PCB may combine:
- 30+ layers
- HDI
- Microvias
- Via-in-Pad
- Back drilling
- Low-loss laminate
- Fine-pitch BGA
- Controlled impedance
- Heavy copper
Each technology is already used in the PCB industry.
The problem comes from combining them.
Higher layer counts increase lamination and registration requirements.
HDI adds laser drilling, plating, and microvia reliability requirements.
Low-loss materials can require different lamination and fabrication controls.
High-speed routing requires tighter control of impedance and geometry.
Heavy copper creates additional challenges during etching and copper balancing.
So the manufacturing difficulty of an AI PCB usually comes from the combination of several demanding processes rather than from one single technology.
High Layer Count
More layers mean more lamination steps and more opportunities for dimensional variation to accumulate.
HDI Reliability
Microvia structures have to remain reliable through thermal cycling and assembly. Via geometry, copper plating, dielectric thickness, and lamination parameters all matter.
High-Speed Fabrication
Controlled impedance depends on maintaining the required dielectric thickness, copper thickness, trace geometry, and etching accuracy.
Back Drilling
Back drilling requires accurate depth and position control so the unwanted via stub is removed without affecting the required connection.
Board Flatness
Large AI server boards and large BGA packages make copper balance, lamination, material selection, and warpage control more important.
DFM and Yield
As geometries become tighter and layer counts increase, small process variations can have a larger effect on yield.
That is why DFM review is better done before fabrication rather than after the PCB layout has already been finalized.
How AI Data Centers Are Affecting PCB Material Supply
The change in PCB technology is also showing up in the supply chain.
AI data centers need more high-end copper-clad laminates, high-speed materials, and advanced packaging substrates.
Materials being discussed in connection with AI infrastructure include:
- M7+ CCL
- T-glass
- NER-glass
- ABF
- BT
These are not all standard materials for conventional server PCBs.
As AI server production grows, demand for specialized materials grows with it.
That can affect:
- Material availability
- Lead time
- Supplier allocation
- Material qualification
- PCB production capacity
For a high-speed AI PCB project, material availability can sometimes be just as important as PCB manufacturing capacity.
Why AI PCBs Cost More
AI PCB pricing is not simply based on board size.
Several factors can affect the final cost:
| Cost Factor | Effect on Cost |
|---|---|
| Layer count | More materials, lamination, drilling, and processing |
| HDI | Laser drilling and sequential lamination |
| Low-loss materials | Higher material cost |
| Back drilling | Additional drilling operation |
| Heavy copper | More difficult copper processing |
| Fine-pitch BGA | Tighter manufacturing requirements |
| Impedance control | Additional engineering and testing |
| Low volume | Higher cost per board |
| Material availability | Can affect cost and lead time |
A high-layer-count PCB with HDI and low-loss materials has a very different cost structure from a conventional multilayer board.
Material availability can also become a cost factor. If a project specifies a particular high-speed laminate and supply is limited, material lead time can affect the overall production schedule.
Where AI Data Center PCBs Are Used
AI-related PCBs are used across several parts of the data center hardware stack, including:
- AI training servers
- AI inference servers
- GPU servers
- AI accelerator platforms
- High-performance computing systems
- Data center servers
- Networking and switching equipment
- High-speed optical communication systems
The exact PCB requirements vary from one application to another.
A GPU board may place more emphasis on high layer count, HDI, power delivery, and high-speed routing, while a networking board may place more emphasis on signal integrity, low-loss materials, impedance control, and high-speed SerDes channels.
Where AI Data Center PCBs Are Heading
AI servers are still moving toward higher computing density.
On the PCB side, several changes are becoming increasingly visible:
- Higher layer counts
- More complex HDI
- Higher data rates
- Lower-loss materials
- Higher current density
- Tighter impedance control
- More back drilling
- Tighter manufacturing tolerances
The growth of AI server demand is already showing up in the PCB industry. Taiwan’s PCB output in the first quarter of 2026 increased 19.6% year over year, with AI server demand helping drive high-layer-count, HDI, and other advanced PCB products.
The important change is not simply that AI data centers are buying more PCBs.
The boards themselves are being asked to handle more layers, faster signals, higher power, denser routing, and more demanding materials at the same time.
That combination is what makes AI data center PCB manufacturing different from conventional server PCB production.
Frequently Asked Questions
How many layers does an AI server PCB need?
The required layer count depends on the GPU or accelerator architecture, HBM configuration, routing density, and stackup. Some current AI server platforms use more than 20 or 30 layers, while parts of next-generation designs are moving toward 40 layers and beyond.
Why do AI server PCBs require HDI?
AI processors commonly use high-density BGA packages. HDI, microvias, and via-in-pad structures provide additional routing space and allow connections to be made within a smaller area.
What PCB materials are used in AI data centers?
Conventional materials can still be used for lower-speed or less demanding circuits. High-speed channels may require low-loss or ultra-low-loss laminates. The actual material choice depends on data rate, channel length, loss requirements, stackup, and manufacturing conditions.
Why are AI server PCBs more expensive?
Higher layer counts, HDI, low-loss materials, back drilling, fine-pitch BGA, heavy copper, and tighter impedance requirements all add manufacturing complexity and cost.
What makes AI PCB manufacturing difficult?
The challenge is usually the combination of several demanding requirements. A single board may need high layer counts, HDI, low-loss materials, back drilling, fine-pitch BGA routing, heavy copper, and tight impedance control.
How does AI data center growth affect PCB supply?
AI data center growth increases demand for high-layer-count PCBs, HDI, high-speed laminates, and advanced substrates. This can put additional pressure on specialized material supply and advanced PCB manufacturing capacity.
