The AI hardware crunch has reached the humble circuit board

zohaibahd

Posts: 1,001   +21
Staff
Ripple effect: PC makers have spent most of this year chasing memory and graphics chips as AI data centers bought up whatever they could. Now the circuit boards those chips sit on are reportedly running short, too. The same factories are busy with far pricier boards for AI servers, so PC orders are getting pushed to the back of the line.

While planning their 2027 lineups, executives at HP and Dell have started worrying about whether they'll get enough boards, Nikkei Asia reports. Even with PC shipments possibly sliding again next year, one manager at an American PC maker said the company still can't get enough of them.

The trouble is that a high-end AI server board can stack dozens of layers of low-loss laminate and specialty glass fabric. So each one ties up far more material and factory time than a desktop motherboard ever would. Cloud giants are happy to pay up for them – Amazon even bought a roughly NT$1.59 billion stake in Taiwanese board maker Gold Circuit Electronics in September.

Every one of those boards starts as a sheet of copper-clad laminate. One Taiwanese PC maker told Nikkei that laminate now costs at least 12 times what it did in July. Meanwhile, one of the world's biggest laminate suppliers says its order book already stretches into 2030. That's despite plans to triple its capacity by 2028.

So board makers are quoting longer and longer waits. One told Nikkei that lead times on a good chunk of its products have doubled. Another said a fresh order placed today wouldn't arrive until late 2028.

New factories won't fix that overnight, either, because they're short on engineers. The scramble to hire has reportedly pushed pay across the PCB industry up by 20% to 30%.

Outside the PC world, Canon is already feeling it. The camera and medical equipment maker said this week that AI-driven shortages of boards and the parts mounted on them have held up production since August. Those shortages are one reason it now expects to lose about ¥90 billion in medical equipment sales this year.

PC buyers are still paying for the last few shortages. Memory and storage prices have climbed all year, so much that Gartner thinks sub-$500 PCs could disappear by 2028. Even an RTX 5080 now sells for close to $2,000 at US retailers. Acer's CEO still thinks PC prices could fall by 2027. But a board ordered today might not even ship by then.

Permalink to story:

 
It's all connected.

I just wonder - who's supposed to use these AI centers if they can't get devices that actually connects to them?
 
It's all connected.

I just wonder - who's supposed to use these AI centers if they can't get devices that actually connects to them?
It’s just a few years… supply will eventually meet demand and prices will stabilize. Most consumers don’t “need” a new PC anyways… wait a few years and it will be fine.
 
It’s just a few years… supply will eventually meet demand and prices will stabilize. Most consumers don’t “need” a new PC anyways… wait a few years and it will be fine.
I would argue that demand doesn't exist because people are buying capacity for things they don't need or can't use. if we are making twice as much *insert whatever* as can realisticly even be used, do we have a demand or a supply problem? The demand is almost entirely artificle. People are buying in demand tech simply because it is in demand, not because they can or want to use it. It's a real underpants gnome problem
 
I would argue that demand doesn't exist because people are buying capacity for things they don't need or can't use. if we are making twice as much *insert whatever* as can realisticly even be used, do we have a demand or a supply problem? The demand is almost entirely artificle. People are buying in demand tech simply because it is in demand, not because they can or want to use it. It's a real underpants gnome problem
Semantics :)

Whether they’re needed or not, they ARE being purchased and used… that will change as demand falls off and manufacturing ramps up…
 
Semantics :)

Whether they’re needed or not, they ARE being purchased and used… that will change as demand falls off and manufacturing ramps up…
I think we are starting to see a return to reason. I recently bit the bullet and bought a 9070XT for $850 but if the bubble pops right now, it would still be months before prices start to come down. I will say that once the bubble pops and all the added infustructure that's been built for chip, memory and PC manufacturing(and the infustructure that supports THAT), we will have probably several years of VERY cheap components again. all those that expanded capacity was built with debt, they have to pay that off and they aren't going to pay it off by not using.
 
I think we are starting to see a return to reason. I recently bit the bullet and bought a 9070XT for $850 but if the bubble pops right now, it would still be months before prices start to come down. I will say that once the bubble pops and all the added infustructure that's been built for chip, memory and PC manufacturing(and the infustructure that supports THAT), we will have probably several years of VERY cheap components again. all those that expanded capacity was built with debt, they have to pay that off and they aren't going to pay it off by not using.
Agreed - I’m looking forward to building an EPYC server in a few years… I figure I’ll be able to get a second hand one pretty cheap in 2030-31…
 
Agreed - I’m looking forward to building an EPYC server in a few years… I figure I’ll be able to get a second hand one pretty cheap in 2030-31…
Until I can find something that I can actually do that uses lots of compute, my 2017 xeon cluster is fine. I have 5 5900X's(they started out as 1800X's) but the b350 boards I'm using are locked in a way that prevents me from doing all kinds of stuff like SSH, hardware raid, ring clustering. I've really only found 2 things to do with a compute cluster and that is computational fluid dynamic(CFD) sims and architectural loads. if you can think of anything interesting I could do on them let me know.
 
Last edited:
Until I can find something that I can actually do that uses lots of compute, my 2017 xeon cluster is fine. I have 5 5900X's(they started out as 1800X's) but the b350 boards I'm using are locked in a way that prevents me from doing all kinds of stuff like SSH, hardware raid, ring clustering. I've really only found 2 things to do with a compute cluster and that is computational fluid dynamic(CFD) sims and architectural loads. if you can think of anything interesting I could do on them let me know.
Well, I suppose you could host a bunch of Minecraft servers :)

I’d be doing 3d rendering…
 
Back