TSMC investments and expansion 2021

[moinmoin]

Since TSMC is currently a crucial bottleneck for the industry I think this deserves an own thread.

TSMC Board of Directors Meeting Resolutions|Taiwan Semiconductor Manufacturing Company Limited

Hsinchu, Taiwan, R.O.C., Feb. 9, 2021 – TSMC (TWSE: 2330, NYSE: TSM) today held a meeting of the Board of Directors, which passed the following resolutions:

pr.tsmc.com

Notable passed resolutions:

4. Approved capital appropriations of approximately US$11,794.8 million (approximately NT$324,293.27 million) for purposes including: 1) Fab construction, and installation of fab facility systems; 2) Installation and upgrade of advanced technology capacity; 3) Installation of mature and specialty technology capacity; 4) Installation and upgrade of advanced packaging capacity; 5) Second quarter 2021 R&D capital investments and sustaining capital expenditures.

5. Approved the establishment of a wholly-owned subsidiary in Japan to expand our 3DIC material research, with a paid-in capital of not more than ¥18.6 billion (approximately US$186 million).

6. Approved the issuance of unsecured corporate bonds in the domestic market for an amount not to exceed NT$120 billion (approximately US$4.4 billion), and the provision of a guarantee to TSMC Global, a wholly-owned foreign subsidiary of TSMC, for its issuance of US dollar-denominated senior unsecured corporate bonds for an amount not to exceed US$4.5 billion, to finance TSMC’s capacity expansion and/or pollution prevention related expenditures.

 

[darkswordsman17]

Nothing that wasn't basically known or all but guaranteed but I guess good to know it from about the most official way (board passing it).

By the time that US fab gets up and running the...I think it was 6nm is not gonna be cutting edge, although I have a hunch we're gonna be on these 7-5nm processes for an extended period of time so it won't be exactly out of date either.

TSMC obviously was going to be spending a lot of money to upgrade fabs. It'd be interesting to know if they're getting aggressive in moving 7nm to newer or if its older fabs. I think the latter would be a better move due to the EUV machine situation.

 

[moinmoin]

It'd be interesting to know if they're getting aggressive in moving 7nm to newer or if its older fabs.

My understanding is that unlike Intel TSMC is not converting its fabs but pretty much always expanding existing ones and building new ones. All nodes TSMC ever offered are still in production.

 

[Hans de Vries]

Capex 2021 of TSMC expected to increase to an incredible 28 Billion $

TSMC hikes capex to record $28bn as chip race heats up

Taiwanese titan logs highest revenue in 2020 and says 5G will fuel further growth

asia.nikkei.com

 

[CakeMonster]

Can you really go wrong with building tons of capacity on current nodes? I mean I remember back when 28nm went into everything and stayed relevant for a long time. Seems demand is even bigger now.

 

[TheELF]

Capex 2021 of TSMC expected to increase to an incredible 28 Billion $

Are these Taiwan dollars? Because TSMC "only" made 18.5bil this year in US dollars, so they would be spending 10 bil more than what they made in the whole year.

 

[Hitman928]

Are these Taiwan dollars? Because TSMC "only" made 18.5bil this year in US dollars, so they would be spending 10 bil more than what they made in the whole year.

Eh? TSMC had just a hair under $48B of revenue last year in USD.

Edit: Ok, I see you are looking at their net profit. I think you are confused about revenue/expenditures/profit. TSMC spent about about $17B in expenditures in 2020 while taking in $48B in revenue with $18.5B net profit. So increasing expenditures to up to $28B, even if revenue doesn't increase, means they would still have a net profit for the year. Obviously they expect revenue to increase as well.

 

[TheELF]

TSMC spent about about $17B in expenditures in 2020

If they did it wasn't in US$ (or in NTD)
These are the US $ numbers TSMC has published for 2020
45 bil rev - 21 bil direct cost=24bil - 5 bil expenditures =19bil income (not net)

Net Revenue $ 45,505
Cost of Revenue 21,342
Gross Profit 24,163
Total Operating Expenses 4,929
Income from Operations 19,258

https://investor.tsmc.com/english/encrypt/files/encrypt_file/reports/2021-02/FS.pdf

 

[Hitman928]

If they did it wasn't in US$ (or in NTD)
These are the US $ numbers TSMC has published for 2020
45 bil rev - 21 bil direct cost=24bil - 5 bil expenditures =19bil income (not net)

Net Revenue $ 45,505
Cost of Revenue 21,342
Gross Profit 24,163
Total Operating Expenses 4,929
Income from Operations 19,258

https://investor.tsmc.com/english/encrypt/files/encrypt_file/reports/2021-02/FS.pdf

You do see that you listed $21.3 as cost of revenue, right? The capital expenditure will be some portion of that, though the details get fuzzy in terms of what is counted when for different companies, but the main point is that you didn't pick out the reported capex which was $17.2B in 2020. So again, increasing that capex to even $28B isn't some unobtainable number as you suggested in your earlier post.

The company, fresh off a record financial performance in 2020, is budgeting between $25 billion to $28 billion for capital spending this year, up from $17.2 billion in 2020

TSMC hikes capex to record $28bn as chip race heats up - Nikkei Asia

 

[TheELF]

Never mind...
Turns out TSMC made more money from amortization and bonds than they did by making wafers, so they had enough money to invest 17bil and put another 7bil in their coffers.

 

[TheELF]

You do see that you listed $21.3 as cost of revenue, right? The capital expenditure will be some portion of that,

Nope, that's direct cost only, paying the electric bill and stuff.

 

[Doug S]

Can you really go wrong with building tons of capacity on current nodes? I mean I remember back when 28nm went into everything and stayed relevant for a long time. Seems demand is even bigger now.

How many customers do you think there will be for 5nm five years from now? Apple, Qualcomm and AMD will all have moved on, except maybe for chipsets for AMD and low end stuff for Qualcomm. There will be some "following" customers who take older nodes, but there's a limit to how much they will consume.

Every node leaves behind customers who NEVER progress smaller than that node. Some because of physical constraints (shrinks help RF components less and less the smaller you get) Others because the design/mask costs are higher and higher with every smaller node, and they simply don't have enough volume to economically justify shifting to a smaller node.

TSMC has to balance who is willing to buy how much 5nm now and in the next couple years with who is willing to buy it and how much in 2025 and 2030. With so few companies having their own fabs there is more demand for leading edge capacity - but that doesn't translate into equivalent demand for lagging capacity because the companies that had their own fabs were mostly like Intel and they retired processes after a few generations and don't e.g. still run 45nm wafers.

 

[scannall]

How many customers do you think there will be for 5nm five years from now? Apple, Qualcomm and AMD will all have moved on, except maybe for chipsets for AMD and low end stuff for Qualcomm. There will be some "following" customers who take older nodes, but there's a limit to how much they will consume.

Every node leaves behind customers who NEVER progress smaller than that node. Some because of physical constraints (shrinks help RF components less and less the smaller you get) Others because the design/mask costs are higher and higher with every smaller node, and they simply don't have enough volume to economically justify shifting to a smaller node.

TSMC has to balance who is willing to buy how much 5nm now and in the next couple years with who is willing to buy it and how much in 2025 and 2030. With so few companies having their own fabs there is more demand for leading edge capacity - but that doesn't translate into equivalent demand for lagging capacity because the companies that had their own fabs were mostly like Intel and they retired processes after a few generations and don't e.g. still run 45nm wafers.

At this point, 5nm looks like it will be a very long lasting node. The number of companies will to chase even smaller node will continue to decrease as costs for them balloon. Masks, wafer costs etc.

 

[blckgrffn]

How many customers do you think there will be for 5nm five years from now? Apple, Qualcomm and AMD will all have moved on, except maybe for chipsets for AMD and low end stuff for Qualcomm. There will be some "following" customers who take older nodes, but there's a limit to how much they will consume.

Every node leaves behind customers who NEVER progress smaller than that node. Some because of physical constraints (shrinks help RF components less and less the smaller you get) Others because the design/mask costs are higher and higher with every smaller node, and they simply don't have enough volume to economically justify shifting to a smaller node.

TSMC has to balance who is willing to buy how much 5nm now and in the next couple years with who is willing to buy it and how much in 2025 and 2030. With so few companies having their own fabs there is more demand for leading edge capacity - but that doesn't translate into equivalent demand for lagging capacity because the companies that had their own fabs were mostly like Intel and they retired processes after a few generations and don't e.g. still run 45nm wafers.

I hope you're right but with the speed of things now we could still be seeing tons of things at 5nm (or freshened up derivative) - refreshed game consoles, heck even desktop CPUs. There isn't enough 7nm now and I would expect that 3nm+ processes are going to consumed around that time with the best and highest paying customers.

I don't have the TSMC roadmap memorized, but if the progress on leading edge nodes slows at all then I would think they could be selling more 5nm (++ or whatever) then compared to what they are selling now.

Haha, I wonder how far out they are allowing capacity orders? Do you have any idea?

 

[Hitman928]

Nope, that's direct cost only, paying the electric bill and stuff.

No, what you are referring to is opex. Are you really suggesting that TSMC spent $17B on capex and $21B on opex last year on $45.5B revenue? Please show me the math how you get to $18.5B net profit as reported by TSMC using those figures.

 

[trexfromouterspace]

No, what you are referring to is opex. Are you really suggesting that TSMC spent $17B on capex and $21B on opex last year on $45.5B revenue? Please show me the math how you get to $18.5B net profit as reported by TSMC using those figures.

17 + 21 = 38. 45.5 - 38 = 18.5. Duh.

 

[rainy]

17 + 21 = 38. 45.5 - 38 = 18.5. Duh.

Sorry, it doesn't work, it's just 7.5.

 

[guachi]

Never mind...
Turns out TSMC made more money from amortization and bonds than they did by making wafers, so they had enough money to invest 17bil and put another 7bil in their coffers.

This isn't a profit/loss statement. It's a statement of cash flows. Most of us think of profit/loss when we discuss whether a company "made money" not how the company manages its cash flows.

What the cash flow statement tells me is that the company is in great shape and generates large cash flows from continuing operations. That it's expanding its physical holdings as acquisition of new property exceeded depreciation on existing property. That even with strong dividends the company is still rapidly expanding its cash holdings billions of dollars a year.

 

[Doug S]

This isn't a profit/loss statement. It's a statement of cash flows. Most of us think of profit/loss when we discuss whether a company "made money" not how the company manages its cash flows.

What the cash flow statement tells me is that the company is in great shape and generates large cash flows from continuing operations. That it's expanding its physical holdings as acquisition of new property exceeded depreciation on existing property. That even with strong dividends the company is still rapidly expanding its cash holdings billions of dollars a year.

They expanded their cash holdings by almost exactly (only off by $5 million) the net amount of new bonds issued. The sale of bonds to increase their cash holdings was probably in preparation for their planned large increased in capital spending for 2021.

They are likely run like US companies used to be in 50 years ago, and US utilities are still run today - their goal is to pay out their entire net income in the form of dividends. Either the tax laws are a bit different in Taiwan (safe assumption) or the "standard business practices" there are more like what "standard business practices" used to be in the US.

 

[Doug S]

FWIW, in their Feb. 9 board meeting TSMC said it plans to have two new bond issues raising a total of $8.9 billion to help with capacity expansion (i.e. that huge jump to $28 billion of capital expenditure, and most likely related to the rumored Intel deal)

 

[Doug S]

Another data point I saw in an article discussing TSMC's results - Apple accounted for over 20% of TSMC's 2020 revenue, and 5nm and 7nm in total accounted for 41% of TSMC's 2020 revenue.

While Apple was still making some rather modest use of TSMC's 10nm in 2020, we can deduce from the above that around HALF of TSMC's 2020 revenue on their leading edge nodes of 5nm and 7nm was Apple.

 

[Mopetar]

I had assumed that more of TSMC's revenue would come from 7nm or newer nodes. I know that not everything needs to be manufactured on bleeding edge nodes, but it's still strange to think that even at the much lower prices per wafer, that so much of their revenue is accounted for by these older nodes.

 

[Doug S]

I had assumed that more of TSMC's revenue would come from 7nm or newer nodes. I know that not everything needs to be manufactured on bleeding edge nodes, but it's still strange to think that even at the much lower prices per wafer, that so much of their revenue is accounted for by these older nodes.

TSMC pretty much doesn't retire any fabs, they just drop the price as the older processes get more depreciated and get customers who can't afford newer processes, customers with products that don't need newer processes, and customers with very long lived products that need to be made with the exact same parts for decades (true of many products in heavily regulated industries where it is costly and may take years to re-qualify a product using newer parts)

For at least the last decade (maybe longer) they've identified some processes as "long lived" and others as not. For example 20nm and 10nm were both designated as having a shorter life and won't be around nearly as long as 28nm, 16nm and 7nm. That way companies that need to be able to get the same part in 2035 they get today know to choose 16nm instead of 10nm etc.

 

[moinmoin]

Looks like Apple shifting to newer nodes has released some of its capacity for Q2 and onward.

https://twitter.com/x/status/1370335123078455297

IC vendors grab extra TSMC 5/7nm capacity for 2Q21

Part of TSMC's 5nm and 7nm process capacity for Apple has been released as the US vendor is transitioning from old to new offerings, with the foundry capacity soon shared out by other major chipmaking clients, according to sources from Taiwan's IC design sector.

www.digitimes.com

 

[CakeMonster]

For at least the last decade (maybe longer) they've identified some processes as "long lived" and others as not. For example 20nm and 10nm were both designated as having a shorter life and won't be around nearly as long as 28nm, 16nm and 7nm. That way companies that need to be able to get the same part in 2035 they get today know to choose 16nm instead of 10nm etc.

This is really interesting, do we know what these decisions are based on? Like if 28/16 turned out to be more durable or have lower error rates than 20 or anything like that?