TSMC Reports Third Quarter EPS of NT$12.54

[lefty]

TSMC seem to be opening quite a lot of new fabs. Three in Arizona, 2 in Japan and 1 in Germany. All of them are more expensive to run than in Taiwan. I just wonder if they can find enough customers to keep the fabs going at 100%.

 

[Daniel Nenni]

TSMC seem to be opening quite a lot of new fabs. Three in Arizona, 2 in Japan and 1 in Germany. All of them are more expensive to run than in Taiwan. I just wonder if they can find enough customers to keep the fabs going at 100%.

I think some of TSMC's top customers will want to make wafers in AZ just to say they are US made. TSMC also got State and Federal money for AZ which is good for the bottom line. I'm not sure if they will build more fabs in AZ, that will depend on customer push. Last I heard four were planned.

Japan and Dresden are a different animal, they are joint development deals with customers built in. It will be interesting to see how that goes but I do applaud TSMC for the diversification. My bet is that they do real well and TSMC will do more of those type of deals, spreading the risk amongst partners and customers.

I do not however see TSMC doing joint venture deals with the financial community like Intel is doing with Apollo. That does not make sense to me but I am not a financial analyst.

 

[MKWVentures]

Did Pat G say publicly that Intel design groups can choose which foundry? Or was it Bob Swan? With chiplets Intel can use 18A for the CPU/GPU and TSMC N3 or Intel 3 for the supporting chiplets if necessary but from what I have heard Intel 18A is solid for internal use moving forward. Pat G said Intel was all-in on 18A which I believe is literal.

Pat Said that Intel Product groups can choose who they want. But he could also tell intel foundry the price to the product group is 50% of TSMC price. Not really much independent until they don't all report to Pat.

previous products were dual designed for TSMC and Intel internal so a choice could be made. Is that still true for 18A Products? or are the operating without a net? 20A was still POR until Q1 2024.

Where exactly do they plan to make all these 18A wafers?

 

[MKWVentures]

TSMC seem to be opening quite a lot of new fabs. Three in Arizona, 2 in Japan and 1 in Germany. All of them are more expensive to run than in Taiwan. I just wonder if they can find enough customers to keep the fabs going at 100%.

need to separate mythical fabs from real fabs. Intel said they will open 6-7 new fabs by 2029

 

[tedc]

Pat Said that Intel Product groups can choose who they want. But he could also tell intel foundry the price to the product group is 50% of TSMC price. Not really much independent until they don't all report to Pat.

According to Intel's own press release: "Instead of recognizing manufacturing costs that were previously allocated to the product operating segments, [Intel Product groups] will be charged a market-based price by Intel Foundry."

If the "market" is TSMC, assuming that just means TSMC pricing. If 18A is all internal, does that mean Intel Foundry sets the same price internally as TSMC, regardless of what margin is?

 

[Mooredaddy]

I think some of TSMC's top customers will want to make wafers in AZ just to say they are US made. TSMC also got State and Federal money for AZ which is good for the bottom line. I'm not sure if they will build more fabs in AZ, that will depend on customer push. Last I heard four were planned.

Japan and Dresden are a different animal, they are joint development deals with customers built in. It will be interesting to see how that goes but I do applaud TSMC for the diversification. My bet is that they do real well and TSMC will do more of those type of deals, spreading the risk amongst partners and customers.

I do not however see TSMC doing joint venture deals with the financial community like Intel is doing with Apollo. That does not make sense to me but I am not a financial analyst.

It doesn’t. It’s desperation. They mortgaged future profits. TSMC would never do the same thing.

 

[MKWVentures]

According to Intel's own press release: "Instead of recognizing manufacturing costs that were previously allocated to the product operating segments, [Intel Product groups] will be charged a market-based price by Intel Foundry."

If the "market" is TSMC, assuming that just means TSMC pricing. If 18A is all internal, does that mean Intel Foundry sets the same price internally as TSMC, regardless of what margin is?

The foundry charges market price as an accounting tool. Which is essentially TSMC price. If they didnt do that, no product group would ever use Intel

When Charging market price, which is TSMC Price, Intel foundry loses 7-10B per year.... . TSMC has Positive 50% Margin. Intel has Negative 30% margin. This has been true for many years. This has been published by Intel repeatedly

When you do the math, the answer is that the full burdened cost is 2x TSMCs cost. There a lots of reasons why and Intel is trying to fix that. Intel says they will fix that on 18A and that their cost will match TSMC Cost. It is not clear how this is possible but that is their commit.

Intel Product Group is doing quite well now when charged TSMC prices and using TSMC Chips

 

[Daniel Nenni]

Pat Said that Intel Product groups can choose who they want. But he could also tell intel foundry the price to the product group is 50% of TSMC price. Not really much independent until they don't all report to Pat.

If the Intel design groups choose TSMC below 3nm I do not see Pat G keeping his job. Hopefully 18A, BSPD, and EMID are compelling for internal use. EMID and BPSD were designed for Intel products, right?

 

[tedc]

Sounds like it's not a true "choice" for the Intel Product team then, is it? It's like putting your thumb on the scale. As long as Intel Foundry can match TSMC offerings (regardless of costs), it'd be difficult to justify selecting TSMC (politically anyway).

 

[Barnsley]

Really hard to believe IFS can ever stand on its own, nevermind take share from TSMC if margins are so deeply negative while TSMC's are 57% and climbing. How will they ever be able to offer discounts to offset switching cost without going bankrupt if this is the situation they are already in. I understand scale is what they want/need and the thesis is. However, TSMC will always have a superior cost structure. And if Intel design is supposed to be independent and allowed to pick whatever process they want where does that leave IFS? Furthermore, why would anyone take such a risky bet going to IFS when we are clearly in such a critical moment for computing. No one can afford to fall behind by trying to save a nickel and dime here or there.

If TSMC margin is constantly rising , it is at whos' expense?

Their employees?
Their suppliers?
Their customers?

Or just the continuous improvement as they reduce waste themselves?

 

[Yunus]

The unexpected jump in margins is likely related to cuLitho, TSMC has recently started using cuLitho, a computation library that should speed up some mask related processing at a fraction of the power budget needed by CPUs. CC Wei is giving a hint to this in his answer to the first question in the earning call:

And think about it, let me use, 1% productivity gain that were almost equal to about TWD 1 billion to TSMC. And this is a tangible ROI benefit. And I believe we cannot be the only one company that have benefited from this AI application

 

[MKWVentures]

If the Intel design groups choose TSMC below 3nm I do not see Pat G keeping his job. Hopefully 18A, BSPD, and EMID are compelling for internal use. EMID and BPSD were designed for Intel products, right

Assuming Intel delivers on process, the challenge is Fab economics. Rumor is that 20A works but Intel could not afford to build and ramp the fab without losing another ~$3B due to low 18A/20A volume so it was moved to TSMC. Arrow lake 20A was GAA BSP per announcement in 2021. It seems to work fine on N3.

My experience has been that Fab economics makes 80% of decisions unless you are rolling in cash. Hence foundry/fabless strategies are so successful. Some really smart person once said "Fabless: the transformation of the Semiconductor Industry" ..

 

[Daniel Nenni]

Assuming Intel delivers on process, the challenge is Fab economics. Rumor is that 20A works but Intel could not afford to build and ramp the fab without losing another ~$3B due to low 18A/20A volume so it was moved to TSMC. Arrow lake 20A was GAA BSP per announcement in 2021. It seems to work fine on N3.

My experience has been that Fab economics makes 80% of decisions unless you are rolling in cash. Hence foundry/fabless strategies are so successful. Some really smart person once said "Fabless: the transformation of the Semiconductor Industry" ..

I agree completely. IDM foundries really depend on their internal business to fund the foundry. I don't think Samsung has ever made a profit on their external foundry business after Apple left but they do internal SoC chips and are dominant memory players so the IDM Foundry business model makes sense. If Intel outsources designs then the IDM Foundry model makes less sense.

Intel is converting 20A to 18A, correct? Making more room for 18A business?

Also, if an Intel is using TSMC then why would a foundry customer chose Intel Foundry? If Intel is using TSMC then how does Intel get economies of scale (better margins) on their manufacturing business to better compete with TSMC?

 

[MKWVentures]

I agree completely. IDM foundries really depend on their internal business to fund the foundry. I don't think Samsung has ever made a profit on their external foundry business after Apple left but they do internal SoC chips and are dominant memory players so the IDM Foundry business model makes sense. If Intel outsources designs then the IDM Foundry model makes less sense.

Intel is converting 20A to 18A, correct? Making more room for 18A business?

Also, if an Intel is using TSMC then why would a foundry customer chose Intel Foundry? If Intel is using TSMC then how does Intel get economies of scale (better margins) on their manufacturing business to better compete with TSMC?

20A is killed as of now. The Fabs will wait for 18A. the timing of this is very challenging

The original model was that scale will double with foundry and allow financial success and tech leadership. The Scale is not on track (foundry commits on 18A is still low) and the cost reduction is not on track (hence current adjustment). Also the UMC/Tower projects need to deliver to use old fabs/nodes

 

[Daniel Nenni]

Also the UMC/Tower projects need to deliver to use old fabs/nodes

Have you heard anything on the UMC and Tower deals? Everytime I ask Intel I get shoulder shrugs.

 

[Jure]

20A is killed as of now. The Fabs will wait for 18A. the timing of this is very challenging

The original model was that scale will double with foundry and allow financial success and tech leadership. The Scale is not on track (foundry commits on 18A is still low) and the cost reduction is not on track (hence current adjustment). Also the UMC/Tower projects need to deliver to use old fabs/nodes

Low external foundry commits or internal too? How many wpm are Panther Lake and Clearwater Forest projected to use?

Edit: Sorry for asking for numbers. It's just insane to think that Intel used to be able to fill multiple fabs (and still could not make enough of Skylake at the peak of the demand) with their own products and now they are at a point where they are unable to fill a single fab...

 

[BruceA]

If TSMC margin is constantly rising , it is at whos' expense?

Their employees?
Their suppliers?
Their customers?

Or just the continuous improvement as they reduce waste themselves?

Simple business of having scale, efficiency broad value portfolio and continued squeezing more value out of investments.

Leading edge is valuable and can get more value and TSMC is most certainly going to charge more given the margins Nvidia, AMD and others are getting. It is like printing money, what Intel used to do for decades with prices and product leadership and that has never been more important for competitive products!

Also remember all of TSMs legacy fabs are mostly to all depreciated and just have operational cost. Even their N5 and N7 nodes are starting to come off depreciation. Coming leading edge nodes have huge RD, as well as depreciation, but scale too giving them huge advantages compared to Intel and Samsung.

Intel has the same burden of RD, capex to wafer capacity ( probably more dollars per WS) but NONE of the legacy fab cash generation.

Past year and especially this and next year have to also pay TSMC for wafers on N3, N5 and N7 wafer yet Foundry gets nothing but negative RD and CapEx and minimal money from old nodes w/o scale.

As Pat has said it will be money losing and long time till break even. It is a long slog and bet the company. If executable and exucted they should be on the beginning of a long climb back to relevance and leadership, 2025 is the first real test of it with PantherLake.

Not sure why so many haters out there. The business needs a strong TSMC and a comparable competitive Intel, three would be better, competition is always good!

 

[BruceA]

Low external foundry commits or internal too? How many wpm are Panther Lake and Clearwater Forest projected to use?

Edit: Sorry for asking for numbers. It's just insane to think that Intel used to be able to fill multiple fabs (and still could not make enough of Skylake at the peak of the demand) with their own products and now they are at a point where they are unable to fill a single fab...

The totally PC market is 200M plus but slowing wiith ARM share rising the battle will be Lunar, Panther against the other companies. No reason it can’t be 100M plus, that isn’t enough and thus the need for Foundry.

Also remember the total area on leading est is decling with Chiplets moving GPU and IO and other control off the leading edge.

 

[MKWVentures]

The totally PC market is 200M plus but slowing wiith ARM share rising the battle will be Lunar, Panther against the other companies. No reason it can’t be 100M plus, that isn’t enough and thus the need for Foundry.

Also remember the total area on leading est is decling with Chiplets moving GPU and IO and other control off the leading edge.

@Jure

18A will eventually be able to fill fabs but new products don't ramp as fast as the used to. As I stated before I amfab saying a full fab is 50K/month. Intel old fabs didnt always get to that level. Also going to GAA/BSP is a huge jump. it is not like gong from 10-7 or even 14-10. Intel reported out on these issues with Fab 34..... they could have kept I3/4 in oregon but they wanted to get the new fab up. So they took a hit on cost and efficiency until Fab 34 ramps.

Assuming 18A works and Intel is truthful, they whole Intel issue will an accounting Fab financing challenge until Foundry revenue materializes. And I haven't worked the impact of the JV/SCIP in.... lots of clauses in those contracts will impact decisions.

 

[dkr1986]

The totally PC market is 200M plus but slowing wiith ARM share rising the battle will be Lunar, Panther against the other companies. No reason it can’t be 100M plus, that isn’t enough and thus the need for Foundry.

Also remember the total area on leading est is decling with Chiplets moving GPU and IO and other control off the leading edge.

Based on what I see online (1), notebook alone accounts for 200 million units per year in 2025 and beyond. Desktop is another 70 million units per year. Per (2), Intel holds 78% market share in 2024 for CPUs in personal computing. ARM is not a threat in PCs as of now (example - please look at Snapdragon X elite - no one can find it outside US and no is probably buying it in US also - QCOM recently cancelled the dev kit). But assuming 75% market share, that is 200+ million units for Intel alone. It makes sense considering Intel shipped 50 million units in Q4 of 2023 per (3). If I am not mistaken, 2023 was a down year for PC sales. So 200 million units/yr for Intel is likely conservative for 2025 and beyond. Please note that Windows 10 end of life is October 2025 and AI PC wave (if there is one).

Panther lake is supposed to be laptop only so we can roughly say 150 million units in 2025. For Xeon I could not get a clear number to mention here but some information here about prior products in (9). Saying its about 1 million per year may be okay with GPUs getting all the attention as opposed to 50 million in 4 years for first gen Xeon (1 million per month).

Considering, the Apollo SCIP came with a minimum wafer commitments (That is probably why Meteorlake was moved to Ireland in my opinion to avoid paying fines), I am guessing Intel will release the Arrow Lake refresh desktop chips on Intel 3 if the leak are true.

I am pretty sure I made many mistakes in coming up with these numbers. But it looks like Intel has enough to fill the fabs on the leading edge nodes like Intel 18A and Intel 3 with recent slowdown in fab expansions. Only question is will the nodes ramp to HVM without any issue and be on time (execution). I am sure (hoping) Intel has more external customers in pipeline.

Recent leaks & news indicate the following,
Intel 18A = used for Panther lake (1x) CPU tile per chip, Clearwater Forrest (12x) CPU tiles per chip. Also Microsoft server chip (4) , Amazon network fabric chip (5), Ericson 5G product (6) and Faraday server chip (7). + Secure Enclave products for DoD (8).
Intel 3 - Granite Rapids, Sierra Forrest, Clearwater Forrest base tile. Some panther lake SKUs will have Intel 3 GPU tiles per leaks. Arrow lake refresh (leaks indicate it is either cancelled/ still alive - no concrete news yet)
Intel 7 - Raptor lake (some leaks about P core only Barlett lake for embedded market next year compatible with desktop LGA1700 socket), Sierra Forrest & Granite rapids base tile. Clearwater Forrest IO tile.

1-https://www.statista.com/statistics/272595/global-shipments-forecast-for-tablets-laptops-and-desktop-pcs/
2-https://www.extremetech.com/computing/intel-holds-78-global-market-share-for-cpus-analyst
3-https://wccftech.com/intel-global-desktop-notebook-pc-cpu-shipments-50-million-units-3-times-more-amd-apple-combined/
4-https://www.extremetech.com/computing/intel-microsoft-shock-with-custom-chip-deal-on-intel-18a-process
5-https://press.aboutamazon.com/aws/2024/9/intel-and-aws-expand-strategic-collaboration-helping-advance-u-s-based-chip-manufacturing
6-https://www.ericsson.com/en/news/2023/7/ericsson-and-intel-expand-strategic-collaboration-to-advance-next-generation-optimized-5g-infrastructure#:~:text=Ericsson%20(NASDAQ%3A%20ERIC)%20and,next%2Dgeneration%20optimized%205G%20infrastructure.
7-https://www.faraday-tech.com/html/News/pressRelease/ENG_01_0403.jsp
8-https://www.tomshardware.com/tech-industry/intel-confirms-dollar3-billion-award-for-secure-enclave-18a-chips-coming-to-us-military
9-https://www.servethehome.com/on-ice-lake-intel-xeon-volumes-and-market-penetration-q3-2021/