Taiwan Semiconductor: No 3nm Soon (NYSE:TSM)

Investment Thesis

Due to the macro environment (and previously worries about a Taiwan invasion), TMSC (NYSE:TSM) stock has been in decline. Nevertheless, the business itself is actually still growing and is expected to continue to do so with above-market growth. As such, TSMC could be a compelling and relatively low risk investment at the current levels, since semiconductors will remain important (and only become more so) as the world continues to become more digital, with TSMC manufacturing chips for pretty much all important chip companies.

However, while the leading edge foundry market for the last decade has been a duopoly between Samsung and TSMC, in the next decade a third competitor will enter this market, Intel (INTC), who has the outspoken aim to become the world’s leading fab yet again (after losing this title last decade).

Background

I have covered TSMC’s progress with 3nm from the beginning:

• In spring 2020 TSMC first revealed that N3 would deliver a quite conservative 1.7x logic shrink, as well as that it was targeted at a ~2.5 year cadence with volume production in H2’22: Taiwan Semiconductor Manufacturing Company Losing Its Process Leadership To Intel (NYSE:(TSM)).
• In mid-2021, TSMC used the word “delay” to describe N3. While the schedule was unchanged (with management indicating a July-August ramp), management did imply that development hadn’t progressed in time for the 2022 iPhone: TSMC Confirms 3nm Delay: What This Means For TSM Stock.
• In late 2021, reports indicated that TSMC had started N3 risk production in Q4’21. Since it takes another 12 months from risk to volume production, this implied that TSMC would miss its previously stated July-August target: Taiwan Semiconductor Stock: 3nm Further Delayed (NYSE:(TSM)).

No 3nm soon

While TSMC’s financials have been discussed elsewhere, the most crucial piece of information from the recent earnings call was a correction of what has previously been posted in the media. To wit, it was widely reported in August that TSMC planned to start 3nm production last month: TSMC to Begin 3nm Chip Production Next Month.

TSMC will begin mass producing chips using its leading-edge N3 (3nm-class) manufacturing process this September, according to a Commercial Times report that cites equipment manufacturers.

However, TSMC directly contradicted this information no less than two times, leaving no room for ambiguity:

Looking ahead to 2023 with the successful ramp up of N5, N4P, N4X and the upcoming ramp of N3, we’re continue to expand our customer product portfolio and increase our addressable market.

(…) Now I will talk about our N3 and N3E status. Our N3 is on track for volume production later this quarter with good year.

I was actually a bit surprised with the August news, since in my previous coverage of TMSC’s N3, I informed investors late last year about how N3 had moved to risk production around November. Since it usually takes a year (12 months) from risk to volume production, the August report of production starting in September already (10 months after risk production) seemed a bit suspicious, or remarkable at best.

However, the new TSMC comments indicate that the reason it was remarkable was simply because it wasn’t true. On Twitter, some people argued with semantics that the August rumor was about production (as opposed to volume production), but the actual article (as opposed to just its title) was clearly about volume production too, as quoted above.

Overall, the takeaway is that the latest management comments have confirmed my thesis from a year ago, which was a confirmation of my thesis from back in 2020 already: TSMC is steadily losing its process leadership due to its slower cadence as a late Q4 production start implies a cadence well over 2.5 years. Indeed, my thesis last year was that TSMC was actually doing a bit worse than my original 2.5-year cadence thesis, now trending towards 2.75 years.

Low volume too

Despite TSMC’s comments that there are twice as many tape-outs for N3 as for N5, there is another signal that indicates N3 isn’t ramping as fast the previous nodes, which were lined up with the iPhone schedule. TSMC stated following:

Now I will talk about our N3 and N3E status. Our N3 is on track for volume production later this quarter with good year. We expect a smooth ramp in 2023, driven by both HPC and smartphone applications. Our customers’ demand for N3 exceeds our ability to supply partially due to the ongoing tool delivery issues, and we expect N3 to be fully utilized in 2023.

We expect N3 revenue in 2023 to be higher than N5 revenue in its first year in 2020 and for N3E to contribute mid-single-digit percentage of our wafer revenue in 2023, as our overall revenue base is much larger today than in 2020. N3E will further extend our N3 family with the enhanced performance, power and yield, and offer complete platform support for both smartphone and HPC applications. N3E development is progressing ahead of plan, and volume production is now scheduled for second half 2023.

Despite the ongoing inventory correction, we observe a high level of customer engagement at both N3 and N3E with a number of tape-outs more than 2x than that of N5 in its first and second year. Thus, we are working closely with our tool supplier to address towards delivery challenges and prepare more 3-nanometer capacity to support our customers’ strong demand in 2023, 2024 and beyond.

Our 3-nanometer technology will be the most advanced semiconductor technology in both PPA and transistor technology when it is introduced. We are confident that N3 family will be another large and long-lasting node for TSMC.

The key piece here is the “mid-single-digit” contribution to 2023 revenue. While TSMC also compared the revenue to N5 in 2020, TSMC only started recording meaningful N5 revenue in Q3’20, whereas TMSC will be recording N3 revenue from late Q1 or early Q2 onwards. More to the point, since the 2023 iPhone will obviously use N3, this alone should push 2023 N3 revenue on par to 2020 N5 revenue, with any other products pushing it higher.

However, given the comment for only mid-single-digit revenue, this actually implies that there may not be much if any other products in 2023 using N3, besides Apple (AAPL) silicon. Previously, it was expected that Intel would also be an early N3 adopter, but there do not appear to be any products on Intel’s roadmap in 2023 that use N3 – which suggests these may have been delayed at best or canceled at worst.

In numbers, assuming $90B revenue in 2023, then TSMC is expecting N3 to contribute around $4.5B. Using a $30k wafer price, this suggests around 150k wafers in 2023. For comparison, TSMC’s N5 Arizona fab was somewhat ridiculed for being just a small 20k wafers per months fab.

These conclusions are in line with previous reports that indicated the original N3 will be a low-volume node, to be followed by the high-volume N3E a year later.

Implications

Intel is still widely seen as being far behind TSMC. However, with seemingly no customers using N3, besides Apple, until 2024, the practical impact of TSMC’s N3 node is becoming very slim.

For comparison, Intel will be starting its Intel 4 production at the same time as TSMC N3, but Intel 4 delivers a 2.0x scaling compared to TSMC’s 1.6x (the scaling target has been further reduced compared to the original already conservative 1.7x goal from 2020). Then, in the second half of 2023 when the high volume TSMC N3E node starts ramping, Intel will start ramping its own Intel 3 node.

The magic act, however, remains that Intel will then quickly follow this up with the 20A node just half a year later. TSMC has no response to this, as its own N2 node is schedule for late 2025 production. Since N2 is on an even slower 3-year cadence and delivers an even lower ~1.3x scaling at best, TSMC will basically remain stagnant for the next half a decade (until it has something beyond N2 in ~2028).

Indeed, previously I have shown the math that N3 and N2 combined are scaling slower than Intel’s very infamous forever-delayed 10nm node: Taiwan Semiconductor Is Likely Finished (NYSE:(TSM)).

As a reminder, since Intel started its IFS foundry business in 2021, the ramp of this business lines up with it overtaking process leadership from TSMC in 2025 with 18A. Intel has already signed up MediaTek (a Taiwanese company no less) for Intel 16, as well as garnering interest from Cisco (CSCO), QUALCOMM (QCOM), and likely NVIDIA (NVDA) for the leading edge nodes.

Besides Intel, Samsung has already started its own 3nm production. While no specific details are available, Samsung’s node is expected to be noticeably less dense than TSMC, although it may nevertheless be quite competitive in terms of performance and power given that it is the industry’s first gate-all-around based process technology.

Financial implications

Since the chip shortages started, TMSC’s gross margin has ballooned to 60%, which is in line with Intel’s gross margins during the heydays of IDM. Its operating margin is out of this world around 50%. So clearly, combined with the strong growth TSMC is still reporting, investors are arguably discrediting TSMC financial performance, as has probably been the case ever since the Taiwan invasion worries gained steam a few years ago.

Nevertheless, TSMC is not immune to the downturn, and it has adjusted its capex to $36B, although some of the reduction is due to equipment shortages. Previously, TSMC had been expecting up to $44B spending this year, but has been gradually reducing this spending. Note that since only around 70-80% of TSMC’s capex spending is towards leading edge capacity, TSMC’s gross leading edge capex spending is now in line with Intel’s 2022 gross capex. Although admittedly for now the profitability and growth achieved by said capex spending is vastly different for Intel and TSMC.

Looking forward, TSMC said it expects a semiconductor contraction in 2023, but claimed that it itself would still deliver growth. One of the drivers for continued growth in the near-term is its increased partnership with Intel. Intel is currently primarily using TSMC for its AXG (graphics) business, and there will be multiple tiles from TSMC (on N6 and N5) in the upcoming Meteor Lake chiplet-based CPU.

Investor Takeaway

With N3 yet to begin production, this new process node will be later than people were led to believe by a widely reported rumor in late August which had claimed HVM would start in September. However, the September schedule would have been far less than 12 months after risk production, which made the rumor not credible for those who had actually followed the node’s progress.

In any case, with Intel seemingly bailing as a primary N3 customer, this leaves literally just Apple as an N3 customer in 2023, with the obvious candidate being the M3 for iPad and Macs as well as the 2023 iPhones.

This low adoption means that TSMC will only be extending its process leadership on paper, with few real industry implications as Apple is basically its own ecosystem. As argued, this provides opportunities for both Samsung and Intel, who will both be moving to gate-all-around (GaaFET) over the next two years while TSMC has been struggling to squeeze another node out of the ageing FinFET.

So overall, this confirms my late 2020 thesis that TMSC would fall behind both Samsung and Intel by 2024: TSMC To Fall Behind Both Intel, Samsung By 2024 (NYSE:(TSM)).

While financially the near-term still looks bright for TSMC (with for example Intel significantly increasing its reliance on TSMC with Meteor and Arrow Lake in the next two years), and its valuation too has become very investible again over the last few quarters (after a few years of having a premium valuation), investors should remain aware of how the competitive-technological landscape is evolving and the risk this could (and likely will) pose in the mid to long-term.