Intel splits Foundry from Design

[Artificer60]

lets say a 18A fab is 50k wafers per month. you can take chiplet sizes and estimate volume. The part people miss is that volumes do not ramp quickly unless you are doing a massive fab conversion. example: In Q4 2023 meteor lake was being outsold by raptor lake 8:1. in fact Alder lake was outselling meteor lake. and the benefit of chiplets is that you do not scale all parts of the chip... so some stay on 3 or 5 or 6.

Why would you assume an Intel fab runs 50K a month? Everything I've read indicates that TSMC's Arizona fab is designed to run 20K per month. I suspect that Intel's fabs are sized similarly, especially the older fabs that have been around a while. What do your numbers say if you downsize the wafer volume to match TSMC? Unless you are considering the Arizona fab complex to be a single fab? It appears you are also neglecting the entire legacy Intel chip volume as shown in the table below. While clearly not a cash cow, it is still significantly better than leaving the fab idle. I think the right question is whether or not Intel will have enough volume on 18A and 14A to offset the $7B per year the fabs are currently losing. TSMC makes a lot of money off their trailing edge capacity. Why wouldn't you expect Intel to do the same? Particularly going forward when 18A becomes trailing edge capacity.

 

[Artificer60]

Oh and the reason I think the 50K estimate is too large is because TSMC is spending ~$12B on their AZ fab while Intel is spending ~$10B per shell on their AZ fabs. Assuming equal cost = equal size of the fab I would expect Intel to be builing fabs that have a capacity of around 20K wafer starts per month as well. I'm also assuming the extra $2B TSMC is spending coverss the added cost of starting up a green field site.

 

[jorgequinonez]

Why would you assume an Intel fab runs 50K a month? Everything I've read indicates that TSMC's Arizona fab is designed to run 20K per month. I suspect that Intel's fabs are sized similarly, especially the older fabs that have been around a while. What do your numbers say if you downsize the wafer volume to match TSMC? Unless you are considering the Arizona fab complex to be a single fab? It appears you are also neglecting the entire legacy Intel chip volume as shown in the table below. While clearly not a cash cow, it is still significantly better than leaving the fab idle. I think the right question is whether or not Intel will have enough volume on 18A and 14A to offset the $7B per year the fabs are currently losing. TSMC makes a lot of money off their trailing edge capacity. Why wouldn't you expect Intel to do the same? Particularly going forward when 18A becomes trailing edge capacity.

View attachment 2295

My understanding is that Intel won't be able to monetize Fabs 12, 22, & 32 until they finish rolling out the 12nm process that they're codeveloping with UMC. They quoted a 2027 date. The current processes on those fabs are only set up for internal use. They're not available nor set up for IFS clients.

 

[hist78]

lets say a 18A fab is 50k wafers per month. you can take chiplet sizes and estimate volume. The part people miss is that volumes do not ramp quickly unless you are doing a massive fab conversion. example: In Q4 2023 meteor lake was being outsold by raptor lake 8:1. in fact Alder lake was outselling meteor lake. and the benefit of chiplets is that you do not scale all parts of the chip... so some stay on 3 or 5 or 6.

Current Intel model is to get half a fab of foundry and half a fab of Intel internal on 18A in 2026. Both are behind and the impact on depreciation and cash flow requires a full fab.

Simple example: I could move the CPU chiplet for ALL Intel Client products (assuming ALL are chiplet based) into a little over 1/2 a fab (unless the yields are terrible). I could make ALL datacenter processors in One 18A fab. If these optimistic items were to happen, even the fastest conversion would mean this does happen until 2028 timeframe.

In order to need two fabs, you need ALL of the Samsungs (non-samsung) leading edge foundry business PLUS ALL Intel CPUs to be made on 18A PLUS Intel to start moving GPUs, SOCs to 18A. Or Intel needs to steal double digit leading edge share from TSMC.... which only happens IF TSMC massively messes up.

The Game changer on this is the WHALE. Amazon Fabric is not a whale, Microsoft is not a whale. government work is not a whale until 2030.
Qualcomm, Nividia, AMD, Apple, going all in is a whale... @Daniel Nenni knows far more than me on this part.

Feel free to do the math and tell me where I am wrong. we can review the spreadsheets

"Government work is not a whale until 2030."

I assume you're referring to military and national security-related applications. For example:

1. 1. The new B-21 Raider bomber has a target total production of 100 units by the end of 2030.
2. 
   
3. 2. The F-35 Lightning II will reach full-rate production at 156 planes per year.
4. 
   
5. 3. Fewer than 300 military or intelligence satellites are currently in service.
6. 
   
7. 4. The Javelin anti-tank missile is set to have a production rate of 3,960 units per year by late 2026.
8. 
   
9. 5. The U.S. Air Force plans to purchase up to 475 T-7 Red Hawk trainer jets from Boeing-SAAB.
10. 
    
11. 6. The U.S. Navy plans to buy up to 516 Naval Strike Missiles (NSMs) to replace Harpoon missiles over a five-year period from 2024 to 2028.

This is a niche market with small production volumes and highly diversified designs, requirements, and functionalities. It's not a large-scale market for a modern semiconductor fab.

 

[semiman]

In theory, they can raise more capitals by selling foundry stocks(which will be cheaper than design + fabrication company stocks) while keeping foundry ownership(keep 51% of stocks). But in order to pursuade investors, Intel design still needs to commit large volumes of wafers.

 

[kevin01]

My understanding is that Intel won't be able to monetize Fabs 12, 22, & 32 until they finish rolling out the 12nm process that they're codeveloping with UMC. They quoted a 2027 date. The current processes on those fabs are only set up for internal use. They're not available nor set up for IFS clients.

Mark my word, for the mature processes, China is going to eat their lunches (including TSMC's). Just look at how much money China is spending on DUV machines, the competition will be so intense that their already-depreciated DUV machines won't save them. The mature processes prices will collapse.

As a consequence, TSMC may lose its golden gooses -- mature processes, and therefore start to lose competitive advantages in a few years.

 

[Artificer60]

My understanding is that Intel won't be able to monetize Fabs 12, 22, & 32 until they finish rolling out the 12nm process that they're codeveloping with UMC. They quoted a 2027 date. The current processes on those fabs are only set up for internal use. They're not available nor set up for IFS clients.

I believe you are correct that the fabs are only set up to run process for Intel products, but my understanding is that only the CPU (and maybe graphics) chiplets are going to run on the leading edge technode. Intel would have to make the other chiplets on their trailing edge nodes (or outsource them). So I think there is still a chance to monetize at least part of that capacity as I doubt Intel needs the output from 3 fabs just to do the development work with UMC.

 

[jorgequinonez]

Mark my word, for the mature processes, China is going to eat their lunches (including TSMC's). Just look at how much money China is spending on DUV machines, the competition will be so intense that their already-depreciated DUV machines won't save them. The mature processes prices will collapse.

As a consequence, TSMC may lose its golden gooses -- mature processes, and therefore start to lose competitive advantages in a few years.

If the US government can slap import tariffs on PRC EVs and solar, they could easily do the same on chips fabbed in the PRC to keep them from dumping.

 

[Artificer60]

Mark my word, for the mature processes, China is going to eat their lunches (including TSMC's). Just look at how much money China is spending on DUV machines, the competition will be so intense that their already-depreciated DUV machines won't save them. The mature processes prices will collapse.

As a consequence, TSMC may lose its golden gooses -- mature processes, and therefore start to lose competitive advantages in a few years.

You lost me here. I think Intel pretty well proved that it isn't possible to economically do leading edge technodes with DUV with their spectacular 10nm faceplant. Or are you saying that China will undercut the monetization of trailing edge processes that TSMC currently enjoys?

 

[samwilde]

Old processes without EUV are still around 50% of TSMC revenues.

At the low end they will face competition from China. While at the higher end they will be facing competition from Intel.
At least that is the theory. We will see in 2 years time if the Chinese and Intel fab expansions go smoothly or not.

 

[kevin01]

You lost me here. I think Intel pretty well proved that it isn't possible to economically do leading edge technodes with DUV with their spectacular 10nm faceplant. Or are you saying that China will undercut the monetization of trailing edge processes that TSMC currently enjoys?

The latter. TSMC's mature processes partially funded its leading edge process R&D, while Intel's products partially funded its leading edge process R&D.

If the US government can slap import tariffs on PRC EVs and solar, they could easily do the same on chips fabbed in the PRC to keep them from dumping.

This is true. However, most other countries won't do this. For example, right now, Chinese electric vehicles are doing quite well outside of U.S.

In Semiconductor, Chinese manufacturers cannot obtain EUV machines, so leading edge processes are safe (for quite a while, at least).

Old processes without EUV are still around 50% of TSMC revenues.

View attachment 2297

At the low end they will face competition from China. While at the higher end they will be facing competition from Intel.
At least that is the theory. We will see in 2 years time if the Chinese and Intel fab expansions go smoothly or not.

Right. The profits from the lower end nodes are probably much higher than 50%, because these equipments are fully paid for by now.

 

[ChrisGar]

The latter. TSMC's mature processes partially funded its leading edge process R&D, while Intel's products partially funded its leading edge process R&D.
This is true. However, most other countries won't do this. For example, right now, Chinese electric vehicles are doing quite well outside of U.S.
In Semiconductor, Chinese manufacturers cannot obtain EUV machines, so leading edge processes are safe (for quite a while, at least).

Right. The profits from the lower end nodes are probably much higher than 50%, because these equipments are fully paid for by now.

Sounds like you think TSMC profit margin at 16nm and above are higher than 5nm/3nm. Really ?

 

[Xebec]

"Government work is not a whale until 2030."

I assume you're referring to military and national security-related applications. For example:

1. 1. The new B-21 Raider bomber has a target total production of 100 units by the end of 2030.
2. 
   
3. 2. The F-35 Lightning II will reach full-rate production at 156 planes per year.
4. 
   
5. 3. Fewer than 300 military or intelligence satellites are currently in service.
6. 
   
7. 4. The Javelin anti-tank missile is set to have a production rate of 3,960 units per year by late 2026.
8. 
   
9. 5. The U.S. Air Force plans to purchase up to 475 T-7 Red Hawk trainer jets from Boeing-SAAB.
10. 
    
11. 6. The U.S. Navy plans to buy up to 516 Naval Strike Missiles (NSMs) to replace Harpoon missiles over a five-year period from 2024 to 2028.

This is a niche market with small production volumes and highly diversified designs, requirements, and functionalities. It's not a large-scale market for a modern semiconductor fab.

The military does have a number of research data centers, and IIRC used to (>10 years ago) operate their own fabs to make these chips. I think the US govt large order is referring to the Secure Enclave order that’ll take time to design/ramp.

DoD data center demand seems to be increasing with uses of various collaboration tools (M365, etc.) being made more secure and placed in non-commercial data centers.

 

[tooLongInEDA]

The latter. TSMC's mature processes partially funded its leading edge process R&D, while Intel's products partially funded its leading edge process R&D.


This is true. However, most other countries won't do this. For example, right now, Chinese electric vehicles are doing quite well outside of U.S.

In Semiconductor, Chinese manufacturers cannot obtain EUV machines, so leading edge processes are safe (for quite a while, at least).


Right. The profits from the lower end nodes are probably much higher than 50%, because these equipments are fully paid for by now.

Hate to burst your bubble, but you clearly missed the memo (4th July 2024) where the EU massively increased tariffs on imported Chinese EVs.

"The European Union has raised tariffs on Chinese electric vehicles, as Brussels takes action to protect the bloc's motor industry.

The new tariffs on individual manufactures range from 17.4% to 37.6%, which is on top of a 10% duty that was already in place for all electric cars imported from China.
"
I'm not saying this is a) a good idea, or b) that it will actually work. But it's certainly happening. In a pretty large market for cars.

 

[tooLongInEDA]

Mark my word, for the mature processes, China is going to eat their lunches (including TSMC's). Just look at how much money China is spending on DUV machines, the competition will be so intense that their already-depreciated DUV machines won't save them. The mature processes prices will collapse.

As a consequence, TSMC may lose its golden gooses -- mature processes, and therefore start to lose competitive advantages in a few years.

Really ?

Just how long do you think the Chinese can carry ever increasing losses here ? There's already a collosal write-off for all the excess equipment they've bought. They then want to compond this by selling well below cost ? TSMC's base cost is presumably lower as they've legitimately depreciated their assets (as you've recognised).

And you think Western countries will just stand back and allow dumping into their markets ? Did they sit back and do nothing at all with steel when it was Japan ? Are they just sitting back and ignoring Chinese EVs now ? No.

One day the forensic accountants will go over the Chinese accounting books from this period. Not for some time. But the report will be a pretty interesting read.

 

[nghanayem]

Sounds like you think TSMC profit margin at 16nm and above are higher than 5nm/3nm. Really ?

Because TSMC passes only a trivial amount of the wafer cost savings as a node ramps to full production and as deprecation expenses go down. In their earnings TSMC for the past few quarters has said N3 drags down their corporate margins (which is normal as this happens with EVERY new process). The slower N3 ramp on account of the N3 vs N3E deilema and longer cycle times probably extend the period of below corporate avg margin. But then again price raises will help this occur faster. And then there is the wafer volume aspect. Obviously ASPs are far lower on old processes but if they have the same revenue that shows just how absurd TSMC’s trailing edge volume is with their never ramp down process node philosophy.

 

[MKWVentures]

Why would you assume an Intel fab runs 50K a month? Everything I've read indicates that TSMC's Arizona fab is designed to run 20K per month. I suspect that Intel's fabs are sized similarly, especially the older fabs that have been around a while. What do your numbers say if you downsize the wafer volume to match TSMC? Unless you are considering the Arizona fab complex to be a single fab? It appears you are also neglecting the entire legacy Intel chip volume as shown in the table below. While clearly not a cash cow, it is still significantly better than leaving the fab idle. I think the right question is whether or not Intel will have enough volume on 18A and 14A to offset the $7B per year the fabs are currently losing. TSMC makes a lot of money off their trailing edge capacity. Why wouldn't you expect Intel to do the same? Particularly going forward when 18A becomes trailing edge capacity.

View attachment 2295

50K per month is plan for a efficient fab. I was talking about 18A. the older fabs have a different plan (maybe 42 will be converted someday). The legacy fabs should be used for legacy. Intel fab goals and TSMC fab goals are not for 20K/month fabs

 

[M. Y. Zuo]

I don’t see what the split companies’ competitive advantages would be… does anyone know?

i.e. even though the supposed ‘synergy’ has been mostly theoretical in the past few years for Intel, that’s still one more advantage than two companies going their own separate ways and trying to compete independently against well established and entrenched competitors…

 

[hist78]

"Government work is not a whale until 2030."

I assume you're referring to military and national security-related applications. For example:

1. 1. The new B-21 Raider bomber has a target total production of 100 units by the end of 2030.
2. 
   
3. 2. The F-35 Lightning II will reach full-rate production at 156 planes per year.
4. 
   
5. 3. Fewer than 300 military or intelligence satellites are currently in service.
6. 
   
7. 4. The Javelin anti-tank missile is set to have a production rate of 3,960 units per year by late 2026.
8. 
   
9. 5. The U.S. Air Force plans to purchase up to 475 T-7 Red Hawk trainer jets from Boeing-SAAB.
10. 
    
11. 6. The U.S. Navy plans to buy up to 516 Naval Strike Missiles (NSMs) to replace Harpoon missiles over a five-year period from 2024 to 2028.

This is a niche market with small production volumes and highly diversified designs, requirements, and functionalities. It's not a large-scale market for a modern semiconductor fab.

Basically, the U.S. government might be a generous 'whale' when it comes to giving free grant money to the semiconductor industry, but it is not a whale when procuring small quantities of specific chips for military or national security-related applications. If not careful, a semiconductor company could lose big money in what seems like a lucrative government contract.

 

[kevin01]

Really ?

Just how long do you think the Chinese can carry ever increasing losses here ? There's already a collosal write-off for all the excess equipment they've bought. They then want to compond this by selling well below cost ? TSMC's base cost is presumably lower as they've legitimately depreciated their assets (as you've recognised).

And you think Western countries will just stand back and allow dumping into their markets ? Did they sit back and do nothing at all with steel when it was Japan ? Are they just sitting back and ignoring Chinese EVs now ? No.

One day the forensic accountants will go over the Chinese accounting books from this period. Not for some time. But the report will be a pretty interesting read.

Once the equipments were bought, it is sunk cost. Depreciated or not, it is just an accounting concept. They have a strong incentive to fully utilize these DUV machines rather than letting them sit idle.

To see how bad the competition can be, In H1 2024, China has spent more money on chipmaking equipments than TW, US, KR, JP combined. https://www.benzinga.com/markets/as...-korea-combined-what-it-means-for-nvidia-asml

Talking about anti-dumping meausures, EU cares about cars much more than where semis get made, as EU are large car makers themselves, but not so much in terms of semi.