Can TSMC Make Enough Chips?
The Question Everyone Wants to Know
Author’s Note: You can watch the video for this newsletter below
I wrote this video before the Q4 2020 results came out. After which, it seems like management had been grappling with the same question. Their answer was “No”. As the world’s industrial machine grinds down due to a lack of semiconductor capacity, TSMC is bulking up its fab capacity as fast as it possibly can - which apparently includes moving a thousand engineers down to Tainan, where N5 chips are being made.
If you like the content that the Asianometry channel covers and focuses on like technology, history and the economy, consider checking out the Patreon. Patreon subscribers get early access to videos before they get released to the public. There are also audio-only versions as well. Worth checking out, and thanks.
It's an exciting holiday season for an electronics buyer. You got some really interesting new products on the market. If you can get them, that is.
There is the new 5G iPhone in three sizes and with a sexy new design. Apple is really ramping up marketing alongside the telecoms to push the 5G angle and sell a huge amount of these new phones.
Apple has also begun transitioning its macs from Intel chips to its own Apple Silicon. The first generation M1 chip has impressed (and in some cases shocked) reviewers with its performance and battery life. The MacBook Air is Apple's most popular Mac and they sell millions of units a year. How many? Who knows, Apple has been working hard to obfuscate their unit numbers over the years.
You also have a new generation of gaming consoles: The Xbox Series X and S as well as the Playstation 5. The gaming industry has never been better, what with the global pandemic and all keeping everyone cooped up inside.
Consoles are long-lived hardware items. They sell a lot of units over many years. The PS4 has sold over 100 million units since its release in 2013. The prior X-box One is estimated to have sold over 40 million since 2013.
And then you got the two large overarching trends relating to the 5G transition (Qualcomm and the like churning out chips for 5G-enabled mobile phones) and cloud computing (with so many people around the world being locked in their homes).
All of these trends will require an immense amount of chips to drive them. The chips for the first three products, those from Apple and the gaming consoles, are exclusively made by TSMC. Being one of the last two fabs using the latest advanced processes, TSMC is a major fab for companies operating in the latter two markets.
Shortages are the usual case with any new product launch. There is always more demand than supply. But the pandemic has interrupted supply chains everywhere AND sent demand levels to unprecedented new heights. So it is hard to say whether or not the shortages that we are seeing in the PS5, Xbox, MacBook Air, etc are related to TSMC's chips or because of some other unknown component that is hard to get in good supply. But what the foundry is making is indeed critical to the final product being ready.
While waiting for the M1 MacBook Air to become available in Taiwan, I began to think about the sheer amount of chips that TSMC needs to make in order to satisfy all of its customers while maintaining market share. It has thus led to this video. It is less a video essay than a collection of thoughts and ideas.
Reviewing Capacity
TSMC makes about 12 million 12-inch equivalent wafers a year. Each wafer, depending on various factors, can result in tens of thousands of chips. But not all of these wafers are made with the most advanced processes.
Almost all of the company's wafers are fabbed in Taiwan, in facilities in Tainan, Hsinchu, and Taichung. There are two fabs on Mainland China (Shanghai and Nanjing) and one up in Washington state, but those two are not using the latest fabbing process - right now, the 5nm process.
(Again another reminder. The 5nm name is an industry-wide marketing label and does not correspond to any actual measurement on the chip die itself. The company actually calls the process N5 in its press correspondences and conferences.)
There is another N5 fab being built in Arizona, but it is rather small and is likely to be for government or location-sensitive applications. In addition, by the time the foundry is ready in 2 or so years, the N5 will no longer be the cutting edge process. It will be superseded by the N5P and the N3.
The chips for Apple’s latest iOS devices as well as the M1 MacBook are made in Fab 18 in Tainan. Fab 18 is the cutting edge of cutting edge facilities, costing an estimated $17 billion. TSMC broke ground on Fab 18 in Jan 2018. Initial production run of N5 chips began some time in Q2 2020, but the Fab won't hit full volume targets until 2021 when its Phase 3 is completed.
It does not appear that TSMC is going to be building any more N5 facilities on Taiwan. The Arizona facility seems to be what they are focusing on right now (they have recently begun hiring for it).
TSMC has a two year half-step cycle for its flagship processes. The first year, you get a N7, N5, or N3 process advancement. The second year, TSMC unveils a half-step improvement, a "plus" process that refines the previous year's step. So that means a N7P or N5P process. The foundry does this to accommodate Apple's annual iPhone release cycle. 2020 is for the N5 process, so 2021 will see N5P and 2022 the N3.
N5P is not expected to require a new fab so it will still come out of Tainan's Fab 18 facility. The N3 Fab, also in Tainan, is scheduled to start coming online two years from now. The new fab (Fab 19, I guess?) is estimated to cost $19.5 billion in total, double of what it cost TSMC to build Fab 15 in Taichung 10 years ago. Cost inflation for semiconductor processes is very real.
So since we can’t expect more foundries being built in the immediate future, I think the key to TSMC getting more chips out to meet customer demand is by improving the yield rate. That means getting the highest possible number of “good” chips per wafer. Per the documentation, N5 is trending in the right direction - following its predecessor N7’s path to a 90%+ yield. So that’s good news.
Scale and the Intel Question
Capacity means more than some consumers not getting their stuff immediately. It also means real business. Nvidia and Qualcomm defected from TSMC to Samsung Foundry recently for some of their top new products. Nvidia's exciting new graphics cards RTX series 3000 for example are fabbed on Samsung's 8nm process.
Part of the reason for this has to be price, as Samsung reportedly offers more favorable terms than TSMC, but I also wonder if TSMC's production capacity has something to do with it too.
And I think this is one thing that Intel still has over AMD, Apple and many other semi companies in the market. Intel might not have the performance crown anymore but they do still have the production capacity. Intel covers the entire market from top to bottom and have built enough foundries to meet that supply. Intel's 2020 chip production numbers are estimated to be over 60% higher than TSMC's and twice as much as AMD, Nvidia and Apple combined.
It’s more of a competitive advantage than you think. Half of winning the game is showing up. The customers might love the products, but if they can’t get them it doesn’t matter at all.
The sheer size of Intel's capacity makes me wonder about the likelihood of the rumors of the company tapping TSMC's N5 process for a future chip. Intel has been a TSMC customer for a very long time, ever since Intel bought Altera for $16.7 billion in 2015. But it's never used the foundry for its most advanced CPUs. It might need to though for a critical project that industry people are watching.
There's an American government project that Intel has been commissioned for - the $500 million Aurora supercomputer. It's due to be delivered in 2021, but its chips (named Ponte Vecchio after a bridge in Europe) need to be made by Intel using its latest 7nm process (roughly equivalent to TSMC's N5).
Intel 7nm is delayed and now scheduled to go into high volume by late 2022. So either Aurora is delayed (for the second time it seems) or Intel taps TSMC for the early Ponte Vecchio run. The interesting thing is that a lot of signs seem to be pointing to it. Not only the comments made by Intel's CEO in recent earnings call, but technology-architecture-wise.
Anandtech recently pointed out in an article that Intel has been rolling out these new technologies that allow it to essentially mix and match dies of different process technologies for a final product. So theoretically it can put together a chip with dies from its own 14nm foundry, its own 10nm, and maybe a die from TSMC's 5nm foundry. It gives them more flexibility and further opens the door to third-party dies. Fascinating stuff.
The thing is capacity. Can TSMC provide the capacity considering Intel’s sheer scale and the current demand they are already facing? Where’s that capacity going to come from? Are they building towards possibly meeting that demand?
Conclusion
TSMC is expected to spend some $17 billion in capital expenditures for 2020. It had raised that budget from $16 billion to $17 billion at Q2 2020 earnings call. In 2019, the company spent some $15.3 billion.
The fact it didn't change that forecast at the Q3 2020 earnings call makes me wonder if they either think they have enough or that they can't get any more of what they want to spend it on. TSMC gets supplier hardware from a company in the Netherlands, ASML. ASML only makes a few of these machines every year. Does the flat capital expenditure number means that there’s no machines available to buy?
TSMC has said that it hopes to maintain 50% market share in the foundry industry going forward. This is a hyper competitive market. You are competing against Samsung as well as the unlimited funds of the Chinese government. Capacity is going to be a critical factor in whether or not it can hold that market position. TSMC is likely going to be able to sell every chip they can make. For me I wonder if they will be able to make enough.
If you know someone who might be interested in the stuff that Asianometry newsletter is about, feel free to share it with them using the button below