I have a hard time believing it is a slam dunk for AMD. AMD already has considerable share in datacenter, while on the client side their volumes are not large. Not sure from where will they will be able to meet the excess demand if they haven’t reserved capacity at TSMC in advance. If what LBT says is true about CPU demand exploding being a recent phenomena, i don’t see how Lisa would have asked for meaningfully higher CPU volume from TSMC.
AMD will have a CPU Portfolio stretched over 4 TSMC primary Nodes in a few months so they will soak up whatever supply they can get on whatever node. They also waste a stupid amount of wafers on bad yielding MI products. They just have to divert like 10% of MI350 / MI400 wafers to CPU to probably like double their CPU DC output lol.
How much is the bottleneck wafers or packaging? Both use 2.5D I think but the EPYC needs larger interposer + a lot more chips to place (more time, lower throughout) ? TSMC super tight there too…
Its not advanced packaging, up to Zen5 all server parts just use the regular organic substrate. Vcache variants need the wafer bonding step but there is no interposer.
Intels architecture is just messed up. AMD’s compute dies work across Ryzen AND EPYC. Intel’s Panther Lake tiles? Client-only. $11.6B inventory they can’t redirect to datacenter. It sucks when you have inventory you can’t use. Been in that situation before
My brother, Fucking good The Information, who have no any supplying train check at all! Who has more competition, Lisa Su or LPT ? Surely, I will buy Lisa for more industry experience! This is why LPT can’t give the accurate forecast on the real demands of CPU , just meeting with the president here and there!
Yup this was always the problem. Intel DC products were so messed up in execution and we still haven’t seen Diamond Rapids since god knows how long ago it was supposed to come out. Client was always going to be okay as AMD still hasn’t managed 30% share there. But server… OMG who was in charge?!
BTW apart from everything else being pointed to by everyone else, this is the line that I found most interesting in the call -
“EUV wafer revenue grew from less than 1% of wafers out in 2023 to greater than 10% in 2025” — Zinsner
Houston we have a problem!
I originally thought that process ramp chart looked crazy when it was first published and WTF were the nodes taking that long to ramp. But go look at it again. Okay that chart is not revenue but still should advanced wafers not allow you to charge more? (Talking revenue not profits) Does it seem like EUV should be only around 10% now at the start of 2026?
How did AMD manage power issue on client product if they just used the same CPU chiplet on server into client product as well? And also normally client CPU will run higher freq. than server product as well, by just using the same CPU chiplet won't this limit client performance too?
Of course, but that's all AMD can do in the near future. Intel, in fact, started to allocate much more capacity to server CPU in 4Q25 when they found out normal server's demand will be much stronger than expected.
Intel products have a lot of problems, but I don't think sharing entire client chiplets makes sense in datacenter products. They already re-use the core IP blocks (P-core, E-core) across client/datacenter products, so that is a good interface for sharing across market segments. Although, it is an entirely separate discussion if they need two entirely different core (P-, E-core) architectures with different ISAs. But, it appears they recognized this and working towards a single unified core arch.
Also, LBT is pulling them in the opposite direction here as well - for e.g., on client cpu Intel removed hyper-threading because it no longer serves the purpose of performance vs. area tradeoff. But, LBT wants hyper-threading to be back on server cpus which runs against the goals of design minimization..
Thank you for a well written analysis laced with humor! Really appreciate this. From overall compute perspective for datacenters/AI- if I have to summarize, it looks like NVDA and the spoils to AMD. Will need to buy more $AMD stock now :-)
CPU demand surged in 4Q25, so AMD also didn't book enough capacity from supply chain to meet its customers demand either. Nevertheless, in terms of server CPU, AMD will still gain meaningful share from Intel.
Since increasing wafer in happened in 4Q25, wafer out will be in 1Q26. Considering you need some time for packaging and shipping, 2Q26 will indeed be the time Intel finally has sufficient supply to meet their customers' demand, which is exactly what Intel said in the earnings call.
I have a hard time believing it is a slam dunk for AMD. AMD already has considerable share in datacenter, while on the client side their volumes are not large. Not sure from where will they will be able to meet the excess demand if they haven’t reserved capacity at TSMC in advance. If what LBT says is true about CPU demand exploding being a recent phenomena, i don’t see how Lisa would have asked for meaningfully higher CPU volume from TSMC.
AMD will have a CPU Portfolio stretched over 4 TSMC primary Nodes in a few months so they will soak up whatever supply they can get on whatever node. They also waste a stupid amount of wafers on bad yielding MI products. They just have to divert like 10% of MI350 / MI400 wafers to CPU to probably like double their CPU DC output lol.
How much is the bottleneck wafers or packaging? Both use 2.5D I think but the EPYC needs larger interposer + a lot more chips to place (more time, lower throughout) ? TSMC super tight there too…
Its not advanced packaging, up to Zen5 all server parts just use the regular organic substrate. Vcache variants need the wafer bonding step but there is no interposer.
Hey Irrational, great post. Just wondering how you’re thinking about semicap after INTCs results?
Intels architecture is just messed up. AMD’s compute dies work across Ryzen AND EPYC. Intel’s Panther Lake tiles? Client-only. $11.6B inventory they can’t redirect to datacenter. It sucks when you have inventory you can’t use. Been in that situation before
My brother, Fucking good The Information, who have no any supplying train check at all! Who has more competition, Lisa Su or LPT ? Surely, I will buy Lisa for more industry experience! This is why LPT can’t give the accurate forecast on the real demands of CPU , just meeting with the president here and there!
Yup this was always the problem. Intel DC products were so messed up in execution and we still haven’t seen Diamond Rapids since god knows how long ago it was supposed to come out. Client was always going to be okay as AMD still hasn’t managed 30% share there. But server… OMG who was in charge?!
BTW apart from everything else being pointed to by everyone else, this is the line that I found most interesting in the call -
“EUV wafer revenue grew from less than 1% of wafers out in 2023 to greater than 10% in 2025” — Zinsner
Houston we have a problem!
I originally thought that process ramp chart looked crazy when it was first published and WTF were the nodes taking that long to ramp. But go look at it again. Okay that chart is not revenue but still should advanced wafers not allow you to charge more? (Talking revenue not profits) Does it seem like EUV should be only around 10% now at the start of 2026?
I love the post.
Struggling to see why Intel has enough upside to be #1 position.
Maybe LBT will fix everything. But it seems it will take quite a bit of time to unf*ck themselves.
There isn't much time to waste.
How did AMD manage power issue on client product if they just used the same CPU chiplet on server into client product as well? And also normally client CPU will run higher freq. than server product as well, by just using the same CPU chiplet won't this limit client performance too?
Server parts are the best, binned parts. Client gets the not so good parts, but the architecture is the same.
Of course, but that's all AMD can do in the near future. Intel, in fact, started to allocate much more capacity to server CPU in 4Q25 when they found out normal server's demand will be much stronger than expected.
Intel products have a lot of problems, but I don't think sharing entire client chiplets makes sense in datacenter products. They already re-use the core IP blocks (P-core, E-core) across client/datacenter products, so that is a good interface for sharing across market segments. Although, it is an entirely separate discussion if they need two entirely different core (P-, E-core) architectures with different ISAs. But, it appears they recognized this and working towards a single unified core arch.
Also, LBT is pulling them in the opposite direction here as well - for e.g., on client cpu Intel removed hyper-threading because it no longer serves the purpose of performance vs. area tradeoff. But, LBT wants hyper-threading to be back on server cpus which runs against the goals of design minimization..
Money isn’t real so Intel is basically losing nothing
Love the humour in this, especially around the Information’s crap reporting.
I love you!
Thank you for a well written analysis laced with humor! Really appreciate this. From overall compute perspective for datacenters/AI- if I have to summarize, it looks like NVDA and the spoils to AMD. Will need to buy more $AMD stock now :-)
It was a fun night as a resident IPG hater/ raptorlake is Intel’s bulldozer kinda guy.
I had not realized Intel was this thight on tooling and was better positioned to capitalize on AMD’s tightness but I guess not.
Good call on 18A once again
CPU demand surged in 4Q25, so AMD also didn't book enough capacity from supply chain to meet its customers demand either. Nevertheless, in terms of server CPU, AMD will still gain meaningful share from Intel.
Unlike Intel, AMD can move some of their client wafers to server and bank it. Intel is only moving low end client to high end.
Also, Intel just secured enough ABF supply very recently, the component shortage should ease in 2Q26.
Since increasing wafer in happened in 4Q25, wafer out will be in 1Q26. Considering you need some time for packaging and shipping, 2Q26 will indeed be the time Intel finally has sufficient supply to meet their customers' demand, which is exactly what Intel said in the earnings call.
We Lip-Bu-Stans only care about community-adjusted yield.; we’ll see you next quarter