You are currently viewing SemiWiki as a guest which gives you limited access to the site. To view blog comments and experience other SemiWiki features you must be a registered member. Registration is fast, simple, and absolutely free so please, join our community today!

TSMC's capex is a cloud over a huge industry

Robert Maire

Moderator
TSMC's capex is a cloud over a huge industry party. Record business is straining the industry. China & Trump are wildcards- Waiting for the hangover...

Party like its 1999 -
We recently attended as a speaker and panel moderator talking about M&A and China, SEMI ISS 2017 (Industry Strategy Symposium). SEMI IS is a gathering of semiconductor equipment industry executives at a retreat for 3 days. In many ways its better than Semicon West as it is a concentration of industry types without those pesky customers and analysts to distract them. The mood was great and there was record attendance. Every companies business is rocking. During the same time period, a number of companies pre-announced better than expected Q4.

We spoke to a large number of suppliers to the industry and everyone is struggling to keep up because business is so strong. The industry is not only hitting on all cylinders but has the turbocharger on.

Q1 looks to be in the bag as orders have already filled production slots. We are in one of those "perfect storm " points where almost everyone is buying, 3D NAND, foundry, logic etc;.

The only downside I heard was that it was too good to last and everyone was wondering "how much longer?". There was also quite a bit of nervousness in conversations that included the words "Trump" and "China" in the same sentence

Front End Flying -
Front end equipment is going great with Dep and etch leading the way. We heard that some products have become "production limited" as suppliers and equipment makers can't ramp fast enough and are simply out of capacity as they ship everything that isn't nailed down. Everyone is struggling to keep up.

No rest for the weary- Normally most companies in the space are shut down the week between Christmas and New Year and in weak times, employees are "given" 2-3 weeks off. No factories were closed this year as everyone worked through the holidays. This pace and stress is continuing into Q1 without let up.

KLAC & LRCX will likely exceed guide-
Our checks seem to indicate strong levels of business for both companies in front of their Q1 reports. We think that Lam's Dep business is super strong as they could be both taking share and having success with new products. Etch not so much as we think Applied and others are keeping up the pressure especially at non-critical etch. It hardly matters as business will be at record levels regardless. KLA continues to reap the benefit of new products.

TSMC = flat Capex, H2 slowdown-
Its sounds like 2017 will be a mirror image of 2016 for capital spending by TSMC. TSMC has been spending beaucoup dollars as it ramps 10NM spend in advance of ramping production for Apple's next gen processor at the end of Q2. The spending pattern appears to be a typical "bell shaped" curve centered around the end of 2016. TSMC told us their quarterly spend ramped in 2016; Q1- $1.15B, Q2- $2.26B, Q3- $3.25B and Q4- $3.53B. If we use TSMC's guidance of flat capex in 2017 with a H1 bias we can extrapolate and simply reverse those 2016 numbers as TSMC slows down spend as it ramps up production at 10NM; Q1/2017= $3.5B, Q2=$3.25B, Q3= $2.25B and Q4= $1.15B- probably not a bad guess. This suggests a sharp roll off that equipment companies need to offset with someone else spending (hopefully).

This TSMC capex announcement will obviously add to fears of a second half slowdown.

95% "Reuse" -
One number that took us by surprise that we heard mentioned several times was 95% reuse of equipment between 10NM and 7NM. While Intel has been great at reuse, others have not been as focused but it sounds like the extreme costs of fabs has made them get focus. Its clear that 10NM will be a "lite" node as everyone agrees it's a quick "pitstop" on the way to 7NM. So quick that GloFo is skipping it entirely (A wise move...). Much like 22NM being a "lite" node on the way to 14/16NM. Other detailed presentations we saw seemed to provide technical support for 7NM being a "tweaked" 10NM node. Sounds like litho won't change much until we get to EUV at 5NM. There will be more dep and etch steps but not as big a bump as we are experiencing in the transition to 10NM.

If the level of reuse actually turns out to be this high it could be scary for equipment makers as TSMC & Intel and others can reduce spend for a longer period. We continue to hear about delays in Intel's "1276" (7NM)process.

The issue of reuse is likely more of a problem for yield management companies as a lot of tools were bought for 10NM but will likely be "reused" for 7NM, while etch and Dep tools will also be reused, additional steps will still require additional new purchases.

China looms large-
Many of the discussions and presentations at the conference were about China and the record number of fabs planned. Even if a third of those fabs turn out to be typical Chinese hyperbole its still a big number. If you do the math, China could be a bigger capex spender, outpacing both Korea and Taiwan by sometime in 2018. The only issue as far as equipment companies goes is that most of those projects are for older technology using refurbed old tools. Many "new/old" tools have been outsourced or sold off to sub-contracters to manufacture as they are no longer attractive to standard new tool makers.There is already a shortage of some older tools to refurb and tool makers could see a bump up in their service revenues to serve this market.

The bigger question is that of emerging semiconductor equipment tool makers in China. There are more than a dozen spanning every segment needed to make chips. The oldest, largest and most successful of which is AMEC started by Gerald Yin (who spoke at the conference), ex of AMAT and LAM. Gerald is a smart etch scientist and business man and has already seen $500M invested in AMEC. Rumor has it that his etch tools have already invaded both the east and west coasts of the US in well known fabs. He has also done a good job of entering the MOCVD market with what appears to be a very good tool that has already seen very strong orders. AMEC will obviously have the home court advantage in China that Aixtron was just shut out of.

M&A Malease -
Our panel and all the attendees seemed to agree that M&A opportunities for the big 5 (ASML, AMAT, LRCX, KLAC & TEL) is significantly reduced. Many at the show expressed surprise that the KLAM deal was blocked but ASML/Hermes sailed through. There still is a lot of opportunity for small and medium size and "roll up" deals to occur but right now, because business is so good, no one has time and everybody thinks their valuation isn't high enough. We will likely have to wait for the eventual "roll over" before we see some renewed M&A activity.

The Stocks-
We think we could see a "pop" in LRCX, AMAT & KLAC in the near terms with great results and a better Q2 guide. Pre-announcements by suppliers support this view. But beyond the near term, things get fuzzy. If anything, concerns about a slowdown will increase with TSMC's H1 capex bias.

The stocks multiples are still low on a relative basis as investors are waiting for the "roll over". No one wants to bid the stocks up to high levels just as business goes off a cliff. While the cyclicality is not like the cyclicality of the bad old days that lasted for years with everyone losing money, the spending news from TSMC is a reminder of the concentrated spending problem in the industry. If memory were to slow down at the same time TSMC slows in the second half we are in for a wild ride, however we see no significant signs of that (yet).

Right now memory , especially 3D NAND remains very strong. Demand and pricing are good as it appears to be a price elastic market that is switching to SSDs. Yield is still not great and there is a little bit of concern that when yields are fixed that supply could.

The current up cycle has taken everyone by surprise, if they tell you otherwise they are lying. No one planned for it, and everyone is struggling. If there is this much upside surprise left in the industry the corollary is that there is same the same amount of downside surprise left. So much for the industry being less volatile....

The stocks that are somewhat left out of the good times are the materials and consumable stocks (the defensive stocks) who are by their nature much less volatile but much more dependable and predictable.

The bigger question on investors minds now is how long to hold on until taking profits; sell now, or after Q1, or after Q2???

Its hard to say that when things are doing so well right now. With TSMC's spending pattern clear it seems to be up to the direction of memory spending to determine overall capex and if a memory spend increase could offset a H2 TSMC weakness. We would look for cracks in memory pricing or weakened demand. We are in the post-holiday period where demand seasonally softens and prices usually drop for memory so we also have to be careful not to confuse seasonal patterns with longer term trends. The speed of China's ramp could also offset a H2 slowing of TSMC but the timing of those fabs and the amount of leading edge equipment is hard to determine.

Given these issues, multiple expansions are going to be hard to come by but the stocks can easily go up on EPS expansion which will likely be the case for the near term.
 

hist78

Member
95% "Reuse" -
One number that took us by surprise that we heard mentioned several times was 95% reuse of equipment between 10NM and 7NM. While Intel has been great at reuse, others have not been as focused but it sounds like the extreme costs of fabs has made them get focus. Its clear that 10NM will be a "lite" node as everyone agrees it's a quick "pitstop" on the way to 7NM. So quick that GloFo is skipping it entirely (A wise move...). Much like 22NM being a "lite" node on the way to 14/16NM. Other detailed presentations we saw seemed to provide technical support for 7NM being a "tweaked" 10NM node. Sounds like litho won't change much until we get to EUV at 5NM. There will be more dep and etch steps but not as big a bump as we are experiencing in the transition to 10NM.

If the level of reuse actually turns out to be this high it could be scary for equipment makers as TSMC & Intel and others can reduce spend for a longer period. We continue to hear about delays in Intel's "1276" (7NM)process.

The issue of reuse is likely more of a problem for yield management companies as a lot of tools were bought for 10NM but will likely be "reused" for 7NM, while etch and Dep tools will also be reused, additional steps will still require additional new purchases.
One factor needs to be considered is if 7nm is a "long life" node (just like 28nm does to TSMC), then the 10nm to 7nm high equipment conversion rate might not be that significant to the potential 5nm fab equipment demand generated from TSMC. Assume 5nm is on track to come to the market.

If 7nm is a long life node, then TSMC must keep most its 10/7nm manufacturing capacity to serve not only those early 7nm adopters but also those other late comers for many years to come. TSMC can't just reassign all the equipment for 10/7nm to 5nm production.
 

Daniel Nenni

Admin
Staff member
One factor needs to be considered is if 7nm is a "long life" node (just like 28nm does to TSMC), then the 10nm to 7nm high equipment conversion rate might not be that significant to the potential 5nm fab equipment demand generated from TSMC. Assume 5nm is on track to come to the market.

If 7nm is a long life node, then TSMC must keep most its 10/7nm manufacturing capacity to serve not only those early 7nm adopters but also those other late comers for many years to come. TSMC can't just reassign all the equipment for 10/7nm to 5nm production.
7nm is most certainly a long node, like 16nm (16nm, 16FF+, 16FFC, 12nm). It is my understanding that TSMC will build more fabs for 5nm and re purpose them for 3nm. In addition to the S0C 7nm (APPL, QCOM, MDTK), TSMC is working with NVIDIA on an HPC version. There will also be EUV versions of 7nm. From what I know there are only a handful of 10nm customers so that will be a quick node like 20nm.
 

name99

Member
"TSMC = flat Capex, H2 slowdown"

I assume this is a very limited "state of 2017 H2" prediction, which I guess matters to some financial people, but not in the grand scheme of things.
I imagine the TSMC schedule looks something like
H1 2017: 10nm including A10X
H2 2017: 10nm in bulk including A11
H1 2018: 7nm including A11X
H2 2018: 7nm in bulk including A12
H1 2019: 7nm+ including A12X

BUT (and this is the point) I assume 7nm+ is the fancy version of 7nm meaning EUV.
Which means big capex during 2018?
That's not a slam dunk, of course, but it seems to me a likely and sensible point at which to introduce EUV.

There's also the question of some much innovation at the first 7nm node? Do we get GAA at that point? Or SiGe? Both of those will presumably require new equipment. And both of those could (in principle...) be retrofitted to 10nm. ie the value proposition is you use the cheaper lithography of 10nm but you use GAA and/or SiGe that were first introduced to the 7nm node to get more value out of your 10nm design?

There's always been an assumption of more or less monotonic movement to higher nodes --- you took all of better lithography AND better materials AND new transistor design as a single package. This certainly made sense when lithography was "easy", it's not clear to me that it makes as much sense now.
I can see the foundries, concerned with maximum equipment reuse and with trying to satisfy all customer needs, switching to a rather less monotonic model, a model where innovations that are essential at smaller nodes, once perfected, are moved back to older nodes so that they can be used by customers who don't have the budget for triple patterning or EUV masks or whatever, but who could still get value out of GAA transistors or better materials, running at the 16 or 10nm nodes.

Obviously this is a departure from historical practice, but everything being done now is a departure from historical practice. No reason the business model will stay unchanged...


As for 10nm being a lite node, I'm wary of how people interpret that. We heard the same thing about 20nm, but what does that mean in practice?
- Apple certainly got value from the node.
- Various other vendors also used the node and (I assume) derived value.
- As far as I know TSMC did not lose money on the node. AND
- They still offer it and plan to keep offering it for a while, along with tweaks and improvements along the way.

So I'm not sure what "lite" node is supposed to mean, EXCEPT that it only stays at the leading edge for one year rather than two.
But people seem to read a lot more into "lite node" than that simple chronological fact, anything from "no real improvement" to "massive waste of R&D money".

Part of me suspects that the people throwing out the term "lite node" are the competitors who don't have such a node of their own and are trying to spread a sour grapes thing. They're welcome to do that, but I think it's the job of the journalists to fight back against that framing rather than accepting it. (For example, it would be valuable, for purposes of understanding this point, to ask the TSMC guys for numbers regarding the volume and revenue they see on their 20nm node, and how they see that evolving over the next two or three years.)
 

hist78

Member
Part of me suspects that the people throwing out the term "lite node" are the competitors who don't have such a node of their own and are trying to spread a sour grapes thing.
I think some fabless companies who really don't have the financial and technical resources to rapidly develop and deploy each new node technology. So for practical reasons, they have to skip certain technology generations. AMD is probably the obvious example.

On the foundry side they are facing similar challenges. I think only the top three, Intel, TSMC, and Samsung, can afford to keep doing this billion dollar game each year repeatedly. For the rest of foundries they can only adopt the "island hopping" approach and wish each two to three years they can claim they are one of the top leading edge foundries "again".

The problem is that during the World War II in the Pacific theater, when the Allies adopted the "island hopping" strategy, their ultimate goal is very clear: to invade Japan and to get Japan's unconditional surrender.

In the semiconductor/foundry industry, skipping a node (island hopping) is probably the right decision many non-billion club members can do. But how long can they keep doing this when the end goal is far from clear or might not even exist at all?
 
Last edited:
Top