TSMC Financial Year 2022
Revenue Up, Cost per Wafer Up
As the world’s biggest semiconductor manufacturer, the health of TSMC is going to be an important metric on how the rest of the world sees contract silicon manufacturing. Not only do we care about TSMC’s roadmaps about how they’re progressing with the latest updates to leading edge technologies, but the financial health of the company spurs a good chunk of that research and development. In the past week, TSMC announced its Q4 2022 and FY 2022 financial performance, and on the face of the numbers, things went well in the last 12 months, even despite the lowering demand for mid-range semiconductor hardware.
So let’s get into the highlights here:
Overall revenue QoQ is up in NTD, but down in USD
In NTD, up 2% $613 billion NTD to $625 billion NTD
In USD, down 1.5%, from $20.23 billion USD to US$19.93 billion USD
Addressing the reason why in New Taiwan Dollars revenue is up, but in US Dollars revenue is down - the exchange rate has shifted significantly even within the quarter due to the weakness of NTD/comparative strength of USD. In 4Q21, the exchange rate was 27.85 NTD to $1, however in 3Q22, that was 30.32 NTD to $1. This is jumped even further in 4Q22, to 31.39 NTD to $1. This means that small 2% gains in QoQ revenue are easily pivoted with 3.5% changes in exchange rates.
However, the reason why revenue is up is mostly due to a shift in 5nm vs 7nm wafer orders, even though as a proportion of revenue, 5nm plus 7nm are the same as last quarter. The shift towards 5nm also results in an average wafer price increase, a slow trend which TSMC has been on for several quarters now after several years of stability. We’ll address that in a bit.
Overall revenue YoY is up 42.8% from Q4 last year
$438 billion NTD to $625 billion NTD (+42.8%)
$15.74 billion USD to $19.93 billion USD (+26.7%, due to exchange rate)
Last year in 4Q21, TSMC cited 50% of its revenue as being on 7nm or 5nm - that has slightly shifted in 4Q22, but not by much - only 54%. But the biggest difference is the weighting - what used to be 23:27 for 5nm:7nm, that ratio is now 32:22, showing the marked increase in 5nm shipments. That being said, revenue for 7nm still increased comparing YoY, from USD$4.2 billion to USD$4.4 billion, but 5nm revenue almost doubled in that time (graph is a few pages down). The overall wafer output stayed the same YoY.
Gross margins are up to record highs
QoQ up +180bps from 60.4% to 62.2%
YoY up +950bps from 52.7% to 62.2%
Other analysts have cited the QoQ change mostly down to exchange rates, however the fact that the longer term GMs have also increased points to increased revenue per wafer. What has a detrimental effect on GMs is utilization rates, especially 7nm, which limited it from being even higher. We’ll go into more detail below.
Full Year 2022
Revenue is up +33.5%, from $56.82 billion USD to $75.88 billion USD
Gross Margin up +800bps from 51.6% to 59.6%
7nm was 26% of 2022 revenue, 5nm was 22% of 2022 revenue
As noted elsewhere, this means TSMC generated more revenue than Intel in 2022. Throughout 2022, this was made by the shift to more leading edge nodes, plus a significant increase in revenue per wafer for those nodes.
Materially, the number of wafers shipped in 4Q actually decreased compared to both the previous quarter and last year:
QoQ down from 3974k to 3702k (down 6.8%)
YoY down from 3725k to 3702k (down 0.6%)
FY22 wafers were 15.23 million, up +7.4% vs FY21
A good chunk of the wafer slow down seems to be attributed to the lowering demand for TSMC’s 7nm. It has been reported that the 7nm tooling that can’t be used for 5nm designs is somewhat underutilized, which is where we come into this crunch in mid-level leading edge semiconductor volumes. Other analysts are reporting a low utilization rate of 7nm - what used to be 100% (or more than 100%) utilization is now down to ~83%, contributing to the reduction in revenue for 7nm. The capacity seems to be there, but as consumer demand is down, it isn’t being used. There are also reports of high-levels of customer inventory based on spending in previous quarters, which also leads to a slow down as orders are finished.
Certain other industries are still seeing shortages, but they don’t use these leading edge nodes - and revenue for some of those nodes are at all time highs for TSMC, but capacity for those larger nodes is somewhat fixed without new investment. It is worth noting here that while TSMC manufactures around 14 million wafers a year (12-inch equivalents), the running theme is that only 2.0-2.5 million of those are on the leading edge 7nm/5nm nodes. Most of the volume is on trailing nodes.
However, average price per wafer went up. A lot.
QoQ average price from $5091 to $5384 (up 5.7%)
YoY average price from $4226 to $5384 (up 27.4%)
(This takes total revenue divided by wafers, and absorbs revenue generated by IP, packaging, and other services)
For average pricing, what we’re seeing here is the mix change to 5nm which commands a premium, but also with those trailing edge nodes still having record demand, TSMC can afford to nudge those prices higher and take orders from everyone willing to pay. We see reports that 5nm wafers are somewhere around $17k average, and 7nm wafers are $10k average (finding real numbers for those nodes is an exercise in guess work and magic given that each customer pays a different amount), which means we can do some extra math on the numbers.
In this graph we have TSMC per-node revenue, which is derived from total company revenue multiplied by the revenue fraction for each node - both numbers are provided in their financials.
For 4Q22,
5nm was $6.4 billion total. At ~$17k/wafer, that equals ~375000 wafers.
7nm was $4.4 billion total. At ~$10k/wafer, that equals ~440000 wafers.
Which means that 5nm plus 7nm is ~815000 wafers out of 3.7 million, or around 22% of TSMC’s total volume is on those leading edge process technologies.
There are also other elements to point out in this graph:
4Q22 was the highest revenue quarter for 28nm, ever
4Q22 was the highest revenue quarter for 65nm, ever
Similarly, 40/45nm nodes are almost at their peak, 90nm+ is almost at the peak, and 16nm is almost the same as in 4Q16 when 16nm was the leading edge. Either way you slice it, almost every node at TSMC has been breaching record revenue, despite overall wafer counts being slightly down compared to last quarter.
The reason I posted the revenue per node graph first over this next graph is because of the distortion effect this graph has on revenue data: this graph of nodes by revenue as a % would suggest that leading edge is holding steady at 48-54% of revenue with the trailing edge nodes declining, however as we know above, some of these trailing edge nodes are both high volume and reaching new highs in revenue.
Long term, TSMC sees this data as somewhat unimportant, especially for leading edge nodes where customer engagement and chip development is more relevant. One example I can think of here is that a large customer might get a good deal on wafers, for example, in exchange for helping development or absorbing some of the yield gains over time and as a result, revenue might be skewed until other customers come on board.
Regardless, this is still a fun graph to plot. The question is when 3nm will start making an appearance, which should be soon. There have been several reports regarding the delay to TSMC’s 3nm node, which uses more EUV and FinFETs, but initial rumours suggested that despite strong desire for a more advanced product, it was difficult to yield to the same degree as 5nm, as well as not performing as intended.
Over time we’ve seen TSMC promote new features regarding N3, such as its new FinFlex design that allows for non-critical path connections to use fewer fins, therefore increasing density and improving speed while also reducing energy consumption. A standard N5 library uses a 2-2 fin design, and One vendor showcased an example of a multi-core Arm mobile phone N3 chip. The high-performance 3-2 FinFlex design in its premium cores (so more fins to help with frequency), the ‘standard’ 2-2 fin design for the medium cores, and the performance and density optimized 2-1 FinFlex for the efficient cores as well as all the IO, the GPU, the AI hardware, the security. Along with N3 general performance enhancements, the idea is that specific blocks within an SoC design can have another degree of optimization.
However, TSMC has obviously advertised several variants of N3. It has been reported that the original version of N3, called N3B, was available for production in 4Q22, however N3E which is meant to be a more volume friendly variant isn’t due for volume until late 2Q23 at the earliest. Originally TSMC had been expecting N3E to start in volume during 2022. This means, famously, that the companies that rely on being first to TSMC’s latest nodes have either had to delay product, or have another generation of N5, such as Apple’s iPhones. Other SoCs are also waiting for N3. It’s going to be interesting to see if Apple takes the jump and relies on N3B over N3E, and takes the hit on yield production to be first out of the gate. It wouldn’t be the first time. Either way, I expect to see 3nm numbers coming in Q2, if not Q3.
Revenue by Platform changes QoQ
4Q22 was 42% HPC, up +10% QoQ
4Q22 was 38% Smartphone, down 4% QoQ
4Q22 was 6% Automotive, up 10% QoQ
Revenue by Platform Full Year
HPC up +59%
Smartphone up +28%
IoT up +47%
Automotive up +74%
For those not familiar with TSMC’s terminology, HPC here means ‘anything above a smartphone’. So all desktop and laptop processors get counted as HPC, even though normally we’d consider HPC the segment above that.
But what we’re seeing here is a shift towards computing at the expense of smartphones. This is despite the drop off in consumer computer sales in the last couple of quarters, the demand for servers and other high-end compute platforms is still strong enough to be an overall net gain for TSMC’s HPC revenue. Everything is up over the full year, especially as we see driving demand for things like automotive, but we wary of inventory levels mixed with 3nm coming later than expected.
Guidance
1Q23 Revenue ~$16.7B to ~$17.5B
Midpoint is $17.1B, a QoQ decrease of ~15%
1Q23 Gross Margin of 53.5% to 55.5%
Midpoint is 54.5%, a QoQ decrease of 7.7%
Long term, 53%+ GM is achievable even with overseas fabs
FY23 tax rate to increase from 11% to 15$
FY23 R&D spend to increase from 7.2% to 8.25% of revenue
FY23 CapEx decrease from $36.3 billion to $32-36 billion
70% for advanced process
20% for specialty process
10% for packaging, masks, etc
FY23 revenue change
1H23 revenue decline mid-to-high single digits vs 1H22 ($35.75B)
2H23 revenue to be greater than 2H22 ($40.16B USD)
Packaging revenue ~7%
These guidance points look really bad. It’s worth noting here that TSMC is assuming a TWD:USD exchange rate of 30.7 NTD, which would mean it would start to creep back to 3Q22 levels. But that doesn’t explain a lot of the decreases here.
Initially, we can look on TSMC’s side. It’s dealing with investing into its N3 process ramp, which will take time to improve yields, inflation, utilization rates, and also all of its investments into new facilities. TSMC also states that certain tax exemptions from the Taiwanese government have expired, then amended, but will lend to an increased expected tax rate to 15%.
However, its financial call showcased that while it expects the market to contract in 2023 by low single digits, specifically bottoming out in the first half. This is reflected by the revenue change predictions, which indicate that TSMC is expecting to decline mid-to-high single digits vs $35.75 billion revenue in 1H22. This means a midpoint of $34.68 billion for the half, or Q1 at $17.1 billion and Q2 at $17.6 billion.
TSMC expects to overall grow in full year based on a very fast N3 ramp in Q2 due to Apple and others. The company states that demand for N3 is strong, stronger than it was for N5 at this stage of the process, and that will drive demand of a premium priced product through the next few years as more N3 capacity is brought online. The company expects N3 will be a mid-single-digit revenue for 2023, and N3 to beat out N5 3-year lagged revenue. However given the price premium, don’t ask what this means for wafer volume production comparisons.
On the point made earlier about 7nm utilization, TSMC stated that during 2H22, 7nm utilization was even lower than predicted, but they believe this to be cyclical rather than structural. Normally with a new leading edge process node, it spends a few years being the best, and then slowly gets usurped by better nodes and volume moves from one to another. At that point, the fab (whether TSMC, Samsung, Intel, GF) then pivots that technology into specialty and differentiated technologies, developing benefits such as high-temp, RF, MEMs, analog, in order to get more out of that investment but also provide value to the markets that typically lag behind leading logic. With all that being said, TSMC is still predicting a weak 1H23 for its 7nm node, especially as R&D filters through to customer designs, and a small pickup in the second half is expected. That being said, having 7nm under-utilized for so long and still not back up to good levels means TSMC is going to take a revenue and GM hit for a while.
Also near the end of the call, not in the prepared remarks, Wendell Huang (CFO) stated in the call that Advanced Packaging is about 7% of the revenue, growing similarly to the corporate rate. It’s hard to partition this out for the graphs above without a Q-by-Q analysis of the revenue rate over time, and I hope TSMC can include that in future.
N2 and Gate-All-Around Progress
On track, better than expected
Risk production in 2024, volume in 2025
One comment from CEO C.C. Wei in the financial call regarding the generation beyond N3, which also introduces a new type of transistor structure called Gate-All-Around. C.C states that the company is still on track, and is making better progress than expected. You might ask if that means the company will pull production in, especially if N3 has issues - the CEO stated ‘the schedule is not changed if we don’t pull it in, but so far so good’. So no real news here, but good to get an update.
Thanks, very helpful to get the perspective. May we infer that the great strength of 5n vs 7n is strongly impacted by Apple taking share in PCs?