AI Server PCB: Inside the 84 Billion Dollar Forecast

On 31 August 2026, a Goldman Sachs flash report substantially revised its outlook for the AI server printed circuit board market. The 2027 forecast was raised by 38 percent to 37.5 billion US dollars, and the 2028 figure was lifted to 84 billion dollars. Over the same period, the copper clad laminate market associated with AI infrastructure is projected to reach 22.1 billion dollars in 2027 and 48 billion dollars in 2028. The report estimates compound annual growth of 148 percent for AI PCB and 161 percent for AI-related CCL across 2026 to 2028.

Forecasts of that magnitude invite scepticism, and they should. But the more useful question is not whether the number is right. It is what has to be true inside a server for the number to be plausible at all.

Reading the Revision, Not the Number

The upward revision was not motivated by a change in expected server shipment volumes. It was motivated by what each server contains.

The report identifies four drivers: increasing PCB layer counts, growing penetration of HDI, upgrading to high-speed low-loss laminates, and higher PCB content per rack. Those four factors compound. A rack with more accelerator modules carries more boards. Boards with higher layer counts consume more laminate and more processing time per unit area. HDI adds microvia and via-filling steps. Low-loss material costs more and runs in a narrower process window.

Multiplying those effects produces growth that is not simply proportional to unit volume, which is precisely why the forecast moved so sharply without a corresponding change in server shipment estimates. The market is expanding because boards are becoming more valuable, not only because more boards are being made.

That distinction matters for anyone evaluating a supplier. A factory that can build twice as many conventional boards has not positioned itself for this market. The requirement is different process capability, not more of the same.

Growth From Value Per System

If AI server PCB growth came only from server count, it would be a conventional capacity cycle. The more significant part of the revision is that the technical content per unit is rising at the same time.

Data transfer rates between GPUs, CPUs, high bandwidth memory, and switch silicon continue to increase, while interconnect distances keep shrinking. That combination changes the PCB’s role. It is no longer simply a carrier that connects devices; it becomes part of the high-speed transmission infrastructure of the computing system, and its electrical characteristics determine whether the links close.

Once the board is part of the signal chain, every parameter that affects loss and impedance acquires economic value. Dielectric loss, copper foil roughness, via stub, layer-to-layer registration, and impedance tolerance all map directly onto whether a design works. Manufacturing capability becomes a design constraint rather than a purchasing decision, and the resulting price premium reflects genuine difficulty rather than scarcity alone.

The practical consequence of value-per-system growth is that PCB cost per rack rises faster than rack count. Buyers planning multi-year programmes should model board content as a rising line item, not a stable one, and should expect that each platform generation demands more from the manufacturing process than the last.

The Layer Count Ladder

The visible expression of this trend is layer count, and the ladder is steep.

Standard server mainboards are moving above sixteen layers. High-end AI platforms extend toward thirty and forty layers, and some high-speed backplanes and extreme interconnect scenarios are entering the region of seventy-eight layers. Each step up the ladder changes the manufacturing problem rather than simply repeating it.

More layers mean more lamination cycles. Each cycle introduces opportunities for material movement and registration error, and those errors accumulate through the stack, so layer-to-layer alignment on a thick board is fundamentally harder than on a thin one. Greater finished thickness raises the aspect ratio of every plated through hole, which makes uniform copper deposition along the hole barrel more difficult and increases the risk of thin or voided plating in the centre. High-speed channels frequently require back drilling to remove via stubs, adding a process step with its own depth tolerance.

Then there is the interaction with density. As chip input-output counts rise, conventional through holes and ordinary trace widths stop being sufficient to route signals out from under large packages. HDI and any-layer interconnect structures provide microvias that free routing channels, and mSAP processes producing lines at 0.075 mm and below create additional capacity in the densest regions. Those techniques originated in consumer electronics and are now moving into infrastructure boards.

Bringing all of that together, the increase in value is the product of three overlapping changes: more layers, finer lines, and higher-grade materials. Each one raises the requirement on process control, and the three interact. A stackup combining heavy copper for power, low-loss dielectric for high-speed channels, and microvias for escape routing demands a process window narrower than any single one of those requirements would imply. Confirming that a manufacturer holds those capabilities in production, rather than only in a sample, is what a PCB capability review is for, and it becomes decisive as the layer ladder climbs.

CCL Follows PCB: Forty-Eight Billion Dollars of Material

The forecast that AI-related copper clad laminate reaches 48 billion dollars by 2028 is the clearest evidence that value is migrating upstream as well as growing in place.

As signal frequencies rise, the limits of standard FR-4 in dielectric loss, stability, and copper foil roughness become visible. High-frequency and high-speed laminates, low-loss resin systems, low-profile copper foil, and higher-grade glass fabric all become more important. Each of those upgrades raises the material cost per square metre, and higher layer counts raise the number of square metres consumed per board.

This has a supply chain consequence that is easy to overlook. The resin systems and glass fabrics used in advanced laminates are also the ones in shortest supply, and demand for them comes from several directions at once. Optical modules need low-loss substrates for 800G and 1.6T transmission. Automotive central computing platforms need high-reliability multilayer boards. Robots and wearables increase consumption of flex and rigid-flex. Each application competes for the same upstream capacity, and that competition is felt at the laminate level long before it reaches a board quotation.

The result is that material capability, not just processing capability, increasingly determines who can deliver. A manufacturer with access to qualified material systems and documented alternatives is in a stronger position than one that buys opportunistically, and this is a difference that shows up in lead time before it shows up in price. Organising that supply is part of board manufacturing discipline, and it belongs in the same review as the process capability discussion.

Where the Technical Boundaries Spill Over

The forecast applies to AI servers, but the technology it describes does not stay inside the data centre.

The same demands appear in switches moving toward 800G and 1.6T, in optical modules requiring finer high-speed lines and tighter impedance control, and in network interface cards and backplanes. Across those products, differential impedance control has moved toward the plus or minus five percent range, and mSAP and high-density HDI have become central rather than optional. Board manufacturing, laminate supply, and optical communication are becoming a coupled industry chain rather than three separate ones.

That chain then extends further. Central computing architectures raise layer counts and high-speed interface counts in automotive main control boards. Unmanned aircraft and robots require lightweight, high-density connections that survive motion, which increases the use of rigid-flex and flexible circuits. Different end products converge on a similar technical direction: more layers, finer lines, higher power density, and stricter signal integrity requirements.

For hardware teams, the strategic reading is that supplier evaluation should focus on demonstrated process capability, verified material supply, and a complete verification chain including assembly-level testing. Boards at this level of complexity do not fail only because of fabrication; they also fail because a high-density package was soldered with marginal paste volume or a hidden joint defect passed inspection. That is why AI-capable PCBA production requires paste inspection, X-ray, and electrical test alongside board fabrication, and why the two stages should sit inside one quality system rather than in separate supplier relationships.

Whether the 2028 figure lands at 84 billion dollars is less important than the direction it describes. PCB content per rack is rising, boards are becoming harder to build, and the materials they require are the same ones under the most supply pressure. Programmes that plan around those three facts will be better placed than programmes that treat board manufacturing as a commodity purchase.

Frequently Asked Questions

Why was the AI server PCB forecast raised so sharply? Because the revision reflected higher technical content per server rather than higher shipment volumes: more layers, greater HDI penetration, upgraded low-loss materials, and more board content per rack.

What layer counts do AI server boards use? Standard server mainboards are moving above sixteen layers, high-end AI platforms extend toward thirty and forty layers, and some high-speed backplanes approach seventy-eight layers.

Why does CCL growth outpace PCB growth? Material upgrades raise cost per square metre while higher layer counts raise the area consumed per board, so laminate demand grows on both axes at once.

What limits a factory from serving this market? Not merely output capacity, but the ability to hold impedance tolerance, plating quality, and registration on high-layer-count boards with microvias and back drilling, in volume rather than in samples.

What should buyers watch in the next cycle? Material availability and process stability rather than headline price. Advanced laminates are shared across AI, optical, automotive and robotics demand, so supply pressure appears early and persists.