Demand for AI hardware has pushed supply pressure down to the lowest layers of the stack. This time the squeeze is on the printed circuit board (PCB): on some board types, delivery times now exceed two years, and a significant share of production capacity appears booked with orders through 2030. In other words, the problem is no longer just the chip, but the substrate the chip sits on.
What happened?
For the high-layer, high-density circuit boards that go into AI servers, demand has outstripped available production capacity. Manufacturers are quoting long delivery dates for new orders; on the most complex board types, the wait has climbed past two years. In short, there are buyers ready to pay, but no boards to deliver.
Why is this bottleneck different from memory?
We've watched the same pattern play out several times. First, accelerator GPUs grew scarce. Then the HBM alongside those chips — and DRAM memory more broadly — hit a bottleneck. Now it's the turn of a component no one puts in a headline: the circuit board. This isn't a coincidence; the demand shock flows top to bottom through the hardware stack, from the most visible layer to the most invisible.
What makes the PCB critical is that it's assumed to be a cheap, ordinary commodity. But the boards used in AI servers are nothing of the sort. They require many layers, specialized materials that can withstand high frequencies, and tight error tolerances. The number of facilities capable of producing these boards is limited. And building a new plant takes a long time, just as it does with chips: construction, specialized equipment and skilled labor add up to years.
What partly eased the memory situation was that DRAM is a standardized product with many manufacturers. On the PCB side, there's less flexibility for substitution. Shifting a high-layer board design to another supplier overnight isn't easy; certification and testing processes slow things down. For that reason, this bottleneck may prove stickier than the earlier ones.
Who feels it, and how?
The most direct impact is on the data centers building AI servers. Even if the chip and memory problems are solved, the system can't be completed if the board doesn't arrive. The bottleneck shifts to the slowest link in the stack, and that link sets the schedule for projects.
| Layer | Visibility of the bottleneck | Substitutability | Pace of capacity growth |
|---|---|---|---|
| GPU / accelerator | High — headline-worthy | Low | Slow |
| HBM / DRAM memory | Medium | Medium — multiple manufacturers | Medium |
| Printed circuit board (PCB) | Low — rarely tracked | Low | Slow |
The second impact is on pricing. As delivery times stretch out, priority goes to whoever pays the most. Large cloud providers can lock in capacity in advance through long-term agreements. Smaller players and late orderers fall to the back of the line. This tilts the balance of power in AI infrastructure a bit further toward the big buyers.
There's an indirect risk in adjacent sectors as well. Once the specialized boards headed for AI servers swallow up capacity, ordinary electronics products sharing the same facilities can get stuck in the queue too. The demand shock tends to spill beyond where it started.
What's next?
There are a few things to watch. First, whether PCB makers announce capacity investments; because until a new line comes online, delivery times won't ease. Second, whether the bottleneck shifts down yet another layer — raw materials such as specialized resins, copper foil and fiberglass could be the next squeeze point. Third, whether the big buyers move toward vertical integration.
The picture shows this: AI's physical cost doesn't pile up in a single part, but across the entire chain. When you widen one link, the bottleneck doesn't disappear, it just moves. The PCB crisis probably isn't the last stop either; it's merely a sign that the chain's least-discussed link has now reached its limit too.
Yorumlar (0)
Henüz yorum yok. İlk yorumu sen yaz.
Yorum yapmak için Sinyal'i indir
Yorumlar Sinyal hesabıyla yapılır. Mobil uygulamada giriş yap, yorum bırak.