AI Server PCB Supply: Why Effective Capacity Matters More Than Labels

On September 9, 2026, analysts Ben Bajarin and Jay Goldberg debated the AI compute supply chain on the semiconductor podcast The Circuit. Bajarin argued that 2027 may be the year when supply constraints are most visible, with new capacity progressively released through 2028 easing the balance. Goldberg expressed concern that large scale capital expenditure across wafer foundry and memory segments could convert into capacity around 2028, potentially restoring loose supply and renewed price pressure. Although their conclusions differ, both point to the same underlying issue: the next phase of the AI industry cannot be assessed by demand growth alone. Every critical link in the supply chain has to expand in step.

Server Shipments Depend on the Slowest Link

An AI server cannot be delivered just because GPUs are available. HBM, advanced packaging, IC substrates, high speed networking, power delivery and thermal systems all have to arrive within a comparable window. Bajarin has previously noted that AI infrastructure only becomes deployable compute when chips, memory, networking, power and cooling complete together, and that a delay in any one of them can hold back an entire system.

This weak link effect changes how the PCB industry should read AI demand. Server and switch PCB requirements do originate from compute expansion, but the final shipping cadence is not determined solely by PCB capacity. If HBM, packaging or power infrastructure is constrained, complete server delivery is affected regardless of how many boards are ready.

Strong AI demand therefore does not automatically imply that every category of PCB capacity will remain tight over the long term. The distinction that matters is which PCB products have genuinely long material qualification, process validation and yield ramp cycles, and which can be expanded relatively quickly once equipment is installed.

High Barriers Do Not Eliminate Price Competition Forever

PCBs used in AI servers and high speed switches are concentrating around higher layer counts, lower loss materials and stricter signal integrity requirements.

As high speed SerDes rates increase, material loss, impedance consistency, back drill stub control and layer-to-layer registration all become more critical. These products are indeed harder to scale rapidly than conventional multilayer boards. Adding equipment is only the first step. Low loss materials require re-establishing lamination, drilling and line width compensation parameters. High layer count structures require control of dimensional change and interlayer offset. Customer qualification has to be completed before capacity becomes truly effective.

But a technical barrier does not mean high end PCB is detached from the semiconductor cycle entirely. If server, substrate, memory and wafer manufacturing capacity all increase around 2028 while end demand growth falls below earlier expectations, price competition can still emerge.

The difference lies in where the competition concentrates. New high end PCB capacity generally forms more slowly than conventional capacity, so competition tends to focus on yield, delivery reliability and customer qualification rather than on square meters of output. A supplier that has qualified a material and established a stable process window competes differently from one that has merely installed equipment.

IC Substrates and System PCBs Are Not the Same Category

Supply chain analysis also needs to separate IC substrates from server system boards.

ABF and similar IC substrates sit within the chip packaging stage, where requirements for line precision, material systems and manufacturing environment are closer to semiconductor packaging than to board fabrication. High layer count server PCBs handle system interconnect between GPUs, CPUs, switch chips and large numbers of high speed interfaces.

Both are driven by AI demand, but their expansion logic and technical barriers differ substantially, and ordinary HDI cannot simply substitute for an ABF substrate. Treating them as one category called high end PCB obscures the fact that one is a packaging component and the other is a system interconnect board.

For PCB manufacturers, the more realistic opportunity concentrates in high layer counts, high speed materials, back drilling, impedance control and selected high density interconnect regions. Capability spanning 1 to 40 layer fabrication, 1 to 5 stage HDI, back drilling support and differential impedance control within plus or minus five percent in high speed scenarios, connectable to assembly, is where PCB fabrication capability contributes to AI hardware programs. For projects still in research validation and small batch introduction, the value of that capability lies in engineering validation of complex boards and multi-process coordination, not in replacing IC substrate manufacturing.

What Actually Survives a Cycle

If Bajarin’s assessment is correct, the central problem in 2027 remains insufficient supply, and the supply chain needs to maximize effective output. If Goldberg’s concern materializes, some capacity becomes loose from 2028 onward and PCB companies shift their focus to yield, cost and customer structure.

The two scenarios appear opposite, but the requirements they place on manufacturers are similar.

A high layer count is not the same thing as effective capacity. Advanced material is not the same thing as high margin. Only production lines where the material has been qualified, the process window is stable, yield can be replicated consistently, and delivery can follow the customer’s cadence actually generate revenue.

Those conditions are measurable. Material qualification means the specific laminate has been validated through lamination, drilling and reliability testing, not merely selected from a datasheet. Process window stability means parameter variation over months stays inside a band rather than drifting. Replicable yield means the distribution is known and controlled, not that a best-case batch was achieved once. Delivery cadence means the line can commit to a schedule and hold it.

Under a quality management system that records these attributes at the lot level, a manufacturer can demonstrate effective capacity rather than assert it. For customers, asking for process capability data and yield distribution is a more reliable way to assess a supplier than reviewing equipment lists.

Where Engineering Validation Fits

The intermediate stage between prototype and volume production is where a great deal of AI hardware work currently sits, and it has distinctive requirements.

Programs at this stage need multiple design iterations, moderate quantities and fast turnaround, combined with measurement discipline so that each iteration produces comparable data. That is a different capability profile from volume manufacturing. It depends on PCB manufacturing responsiveness and on the ability to run AI hardware PCBA assembly in small batches without losing process control.

Suppliers who serve this stage well accumulate material and process data that becomes the foundation for volume qualification later. That is why participating in a customer’s development cycle early has strategic value beyond the immediate order, and why the competitive dynamic in high end PCB is shifting from who has the largest capacity to who is engaged earliest in the design cycle.

Reading the 2027 and 2028 Debate Correctly

The more useful conclusion from the debate is not a prediction about which year marks a cycle peak. It is a distinction between ordinary capacity expansion and effective high end capacity.

As the AI industry enters a more mature capital expenditure cycle, earning a profit purely from demand growth becomes more difficult. Value is more likely to remain in the high speed, high layer count PCB segments where expansion is slow, qualification cycles are long and manufacturing stability requirements are high.

That has practical implications on both sides of the market. Buyers should evaluate suppliers on documented process capability rather than on declared technology level, because the ability to build a demonstration board is not the same as the ability to deliver production volumes at consistent yield. Suppliers should invest in the measurement and documentation systems that convert equipment into qualified capacity, because that is what customers will be purchasing once the current supply tightness resolves.

The same logic applies to programs that combine high speed and high power requirements, where high volume PCB assembly capability matters only after the process is stable enough to scale. Building that stability during the validation phase, rather than attempting to establish it during ramp, is what determines whether a supplier participates in the next cycle or watches it from outside.

Why Back Drilling Became a Qualifying Criterion

Back drilling illustrates how a single process step can separate qualified capacity from declared capability, and it is worth examining because it appears in most high speed AI server specifications.

A plated through hole that passes entirely through a thick board leaves a stub below the layer where the signal actually transitions. That stub behaves as a resonant structure, degrading the high frequency response of the channel. Back drilling removes the unused portion of the barrel, restoring the channel’s insertion loss and return loss characteristics.

The process difficulty is depth control. Remove too little and the stub remains, so the electrical benefit is partial. Remove too much and the remaining barrel may be too short to make a reliable connection to the target layer, or the drill may intrude on an adjacent signal layer. On a board with many high speed channels at differing depths, the operation has to be programmed per hole with depth accuracy that holds across the panel and across production lots.

Because the consequence of error is internal and largely invisible at final inspection, back drilling capability can only be assessed through data: depth measurement records, cross section results and electrical verification of insertion loss on test coupons. A supplier that performs back drilling and can show those records is demonstrating something a supplier that lists it as an available process is not. For AI server programs, that distinction determines whether a specification that looked achievable at quotation remains achievable at volume, which is precisely the difference between effective capacity and a capability claim.

The Practical Takeaway

AI demand will continue to pull the electronics supply chain forward for the foreseeable future. What changes as the industry matures is that the constraint stops being demand and becomes synchronization.

Chips, memory, packaging, substrates, boards, power and cooling all have to complete together for deployable compute to exist. In that environment, a PCB supplier’s value is defined by the portion of the chain it can hold stable, and by how early it can participate in a program before the design is fixed.

For engineering teams planning AI hardware, the practical step is to confirm that the manufacturing partner’s claimed capability is backed by process data from production, not prototype builds. For manufacturers, the work is to make that data available as a matter of routine. The companies that do this will find that the 2027 supply question and the 2028 oversupply question require the same answer, which is genuine, documented, replicable capability.