Mark Liu once was a manager in a US fab, and never got a chance to enter the vp level. He quit and became the ceo of tsmc later. Did you learn anything here? something wrong with the US system.
https://www.crunchbase.com/person/mark-liu-2
if they had kept Liu and promoted him to be the ceo. there is no tsmc competition now. tsmc could just be a 2nd tier company poses no threat, like GF.
Without Mark Liu I would assume TSMC will be exactly where it is now. TSMC has and had many other brilliant managers and engineers. You also assume that Jr Manager Mark Liu needed to be the head of integration ahead of Holt and later Bohr. Considering that team delivered strain 6 years ahead of Liu's team, HKMG 4 years ahead, and finFET 4 years ahead it seems fantastical to say intel would be in a much better state if they took one random GL or at best an EM and rocketed him to the top of a company who at the time was being led by Grove/Moore and had CB waiting in the wings after incorporating lean and SPC principles to intel's manufacturing and Chau changing intel's development model while simultaneously leading the charge on HKMG.
1) The foundry business started with IDMs using extra capacity for fabless customers. Japan was big on this which is why I spent the first part of my career in Japan. The problem is that IDMs competed with customers and put them second in line when capacity tightened. This created an opening for the pure play foundry business model which is in place today.
Considering there has never been reports of this being an issue at Samsung even when their capacity is tight I don't think this is a concern the same way it was back then. Definitely a back of the mind concern for many fabless firms, but for Samsung Foundry and Intel foundry it is in their best interest to not follow the example of the Japanese IDMs. Fortunately the semi-independence is a boon here.
2) TSMC builds fabs based on customer pre-pays and wafer agreements. Since TSMC is the only game in town they can do that.
Technically intel has that one 18A pre-pay and if memory serves some advanced package ones. But I suspect that this is the exception rather than the rule.
Intel cannot and does not really know how much capacity they will need until it is too late. Clearly Intel is optimistically building for the foundry business but at their own peril as we now know
Reading between the lines on earnings calls it does sound like intel hoped to be further along with more external wins by now. However I don't think that is the main issue causing IS, and OH push outs. Even if intel foundry ends up as some gangbuster thing that was more successful than anyone has ever seen from a new foundry we are still talking a minority of intel's total wafer volumes. Further evidence for this is that intel claims their plan for 2027 break even assumes pretty insignificant external revenue. On the schedule intel set in 2021/22, by 2027 OH would have been up and Germany would have been spinning up. On the newly shared schedule intel claims that Ohio should be spinning up in that 2027 timeframe. So to me that reads like intel was not building Ohio or Germany on the promise of filling those exclusively with external customers. Rather it seems that capacity was mostly intended to fill internal wafer demand.
With the above information my read is that while foundry might be softer than intel would have liked I suspect that the real reason for shuffling things around is internal product projections being worse.
[table]
[TR]
[TD]Fab[/TD]
[TD]Original Advertised Process[/TD]
[TD]Suspected Process[/TD]
[/TR]
[TR]
[TD]OH 1/2[/TD]
[TD]18A[/TD]
[TD]14A[/TD]
[/TR]
[TR]
[TD]Germany 1/2[/TD]
[TD]post 18A (presumably 14A)[/TD]
[TD]10A[/TD]
[/TR]
[TR]
[TD]Fab 38[/TD]
[TD]never publicly announced[/TD]
[TD]intel 3 because at IFS direct connect intel expected to produce more intel 3 than 18A in the long term and intel claims 42-62 will be 18A fabs. Currently only 34/44 are publicly slated for intel 3 so a third fab will tie things up.[/TD]
[/TR]
[TR]
[TD]Fab 48[/TD]
[TD]never publicly announced rumored to be paused while a process tech is selected[/TD]
[TD]intel 3, 18A, or other depending on capacity needs[/TD]
[/TR]
[/table]
Intel clearly was under the impression that the growth from the COVID DC and laptop booms would continue forever. Intel's revenue back then was like 77-79B per year with talks of 100B being right on the horizon as the PC TAM grew to 300M units per year. What we got to see was 54B in 2023 and intel being on pace for a flat 2024. To further support this argument there is also intel claiming how they got caught off guard with too much inventory in 2022, 2023, and now again in 2024. With a revenue swing like that it is not hard to imagine how 5 intel 18A fabs would go from about right to way too much. It also just seems more likely to me that the intel demand side of the equation was wrong rather than over estimating foundry by multiple orders of magnitude. But I could be wrong I don't make intel's LRPs so I am only speculating based on the publicly available data.
4) Advanced packaging is a very low margin business in comparison to selling chips. As an entry into the foundry business it makes sense, but Intel doing independent packaging does not.
If IF was just a wafer fab sure. But they are also one of the world's largest OSATs. To that side of the business ADV package and test is a huge boon to the GMs from like 20% that really profitable OSATs do. Although it might not be as bad for intel as a regular OSAT since intel just does FCBGA as their lowest value thing rather than doing tons of wirebonding business for charity like other OSATs.
It would be a very serious business pivot.
Intel foundry was always pitched as opening up intel's wafer fab services AND OSAT services. Also at this point getting some upside 30-40% GM adv packaging deals sounds like a great way to hit breakeven sooner.
Also, it is not clear to me that Intel has a technical packaging advantage over TSMC. Delivering bleeding edge technology to foundry customers must be done carefully without risk of delays or failure.
I think the story is more nuanced than TSMC or intel let on. I think the concept of EMIB is objectively superior to TSMC's large COWS passive interposers due to the significantly lower cost and not being limited by litho field limits. TSMC objectively leads intel on hybrid bonding tech with intel only catching up next year just as TSMC seems like they will move to that 3-4 micron bump pitch range. COWS has 3D MIM caps but if memory serves the 3D MIM caps found in MTL's foveros base die are better, and as far as I know out in the wild EMIB products don't have 3D MIM. Intel has had high density organic for longer than TSMC. intel claims their singulated die test is better than what any other OSAT can achieve, and I would assume TSMC also falls in that camp with the OSATs. But I don't know enough about die test or TSMC's capabilities to really have an opinion if this claim from intel has any merit.
TLDR TSMC does some stuff better than intel, and intel does some stuff better than TSMC. Both are trying to rapidly catch up in the areas the other is strong in.
1. Can Intel stay as IDM most of the time and provide foundry services to cover some costs?
2. Can Intel moderate its pace of fab building to meet its internal/external demands to be prudent financially?
3. AI demands should be large enough. Can Intel adjust its product portfolio to fill the fabs better?
My take is that this just takes longer than many people care to admit. As an example GF took a decade to even hit breakeven and that was after dropping all TD, the China fab construction, and the then planned New York expansions. They were also jumped started by merging with the number 3 foundry of the day only for them now to be the 4th largest (albiet with the caveat that Samsung's numbers are supported by internal). Samsung for it's part started in 2005 and didn't get a non internal customer (which includes Apple until they started doing their own custom APs) until either Apple first custom SOC in 2012 or QCOM in 2016 depending on how you want to count it. If intel hits break even 6 years after starting their foundry dreams and launches external customer products on advanced CMOS in under 6 years years that is by far and away more successful than any other new entrant to the foundry market has ever been.
www.crn.com
