文︱Trading Places Research
Source — seekingalpha
Compile | editorial office
When I think of Pat Gelsinger, energy is the first word that comes to my mind. Since taking office, Kissinger has been Intel's CEO for a year. He brought tremendous energy to a company that was in full swing. In the past year:
Intel's new semiconductor manufacturing business will provide chip foundry services to external customers, including Intel x86 and CPU chips centered around arm and RISC-V instruction sets. It's something Intel has never done before.
Intel has already broken ground in Arizona with two new manufacturing facilities (FABS) with about $20 billion in capital, which are expected to open to customers by mid-2023. Ohio has announced a similar investment, but I think the investment is more dependent on federal tax credits. At present, the House and Senate have passed the U.S. Chip Act, and they must eliminate substantial differences in cost and scope. The European Union is also considering a $50 billion chip fab subsidy, which Intel hopes to benefit from.
Kissinger has devoted a great deal of time in the media to restoring Intel's image.
He also invested a lot of time in consolidating relationships with semiconductor manufacturing equipment manufacturers, especially the most important ASML. Kissinger met with the ASML CEO at least three times. This is a welcome change for a company that tends to say no to its suppliers.
Intel employs a large number of talented people.
Automotive chip division Mobileye is about to raise capital in an initial public offering and sell much of its NVRAM business to SK Hynix.
Recently, Intel also announced the acquisition of Tower Semiconductor. Tower Semiconductor produces a range of low- and mid-range chips, such as image sensors and power management chips, which are among the most out-of-stock chips in the current semiconductor market. This acquisition is an option with complementary strengths. Intel is making inroads into the high-end market, which gives them more access to the entire fab business from top to bottom. Tower Semiconductor also brings experience working with these external customers, who will be new customers to Intel, while also bringing the types of industry-standard design software that customers need. The software side is the key to attracting customers.
Intel has formed an "Ecosystem Alliance" with several companies in the ARM and RISC-V IP space. Again, the problem here is that most of the work Intel does is design x86 CPU chips and put them into mass production in its own fabs. Now they are opening up to new customers who need an IP and design software menu to offer these products.
And so on, Pat Kissinger brought tremendous energy to Intel. It's no exaggeration to say that in just one year, he has upended Intel's culture. This is an amazing achievement.
But the question remains: Is he moving in the right direction, or is he just moving forward?
The strategy can be summarized as follows: In 2025-2030, Intel wants to maintain its current share of the PC and data center business while seeking to strengthen its cooperation with TSMC.
Source: YCharts
This is a very ambitious goal and, in my opinion, is likely to end in failure. The most likely outcome is that Intel's foundries will still rank third after TSMC and Samsung, so Intel will not have the head pricing power that TSMC enjoys. Meanwhile, the x86 CPU business, Intel's primary business since 1981, will fade away as Intel and AMD hardware are replaced by cheaper, more efficient, and more flexible ARM and RISC-V hardware. Ironically, Intel's new foundry business will produce the latter two chip products, and its average profit from each laptop and server will also be much less than before.
But again, I was quite shocked by what Kissinger had accomplished in just one year. He set ambitious goals for Intel, and the first step went well. But I still think failure is a more likely outcome.
Intel's legacy business
First, let's look at Intel's x86 and other businesses. The two main sections of Intel's annual report are "client computing business" or PC chips, and "data center business" data center chips. These businesses account for 84% of Intel's revenue and 111% of operating income.
Source: Intel Annual Report
The blue part is the x86 business, which can be seen dominating. Since the red part of the NVRAM and Mobileye are already planned to be sold, the x86 business will further dominate, with 7% of revenue and 9% of operating income will be excluded. The sale of NVRAM is largely complete, and Mobileye's IPO is expected to take place in mid-2022.
But these blue sections also contain "proximity" revenue, products sold with x86 chips, such as server motherboards, network chips, memory, and storage. These adjacent markets are highly dependent on x86 CPU chips, but let's take them out in order to focus on the share of x86 chips.
As a result, 76% of Intel's revenue in 2021 will come from the x86 chip itself, and another 7.9% of adjacent revenue is highly dependent on x86 sales. If you leave aside the NVRAM and Mobileye related businesses, its revenue is more concentrated in the blue x86 segment:
To sum up, since 1981, although Intel has gradually begun to lay out other businesses, its main business is still to design, manufacture and sell x86 CPU chips.
Problems with Intel x86 chips
With the release of the latest 12th generation Core i9 chip designed for laptops, the issue of Intel x86 chips has gradually become the focus of attention. Similarly, the laptop segment accounts for about a third of Intel's revenue. In the past, Intel started at the bottom of the i3 or i5 and went all the way up. But this time they opted for gorgeous high-end products that wouldn't ship high because laptops with high-end chips are expensive, more than $3,500, which would be impractical for any consumer other than gamers.
They're not only doing this because, over the years, Intel has been beaten by its main competitor, AMD. Apple is now making the M1 Pro and M1 Max, not x86 chips, but based on the ARM instruction set, and beat Intel's 11th generation chips in performance comparisons. According to Bloomberg, Apple was previously Intel's fourth-largest customer, accounting for about 10 percent of Intel's revenue, and about $7 billion in 2019.
As a result, Intel believes they need a compelling, high-performance laptop chip that has garnered a lot of buzz from these headlines.
Image source: PCWorld
But there are also headlines that think it's hype.
Image source: The Verge
Back in January, Intel sent a $4,000 17-inch laptop to reviewers, the MSI GE76 Raider. They even sent out low-resolution models, so the base frame rate would be higher in comparisons with AMD's games. Monica Chin's second review in The Verge goes to the heart of Intel's problems, as well as pointing out AMD's problems:
In terms of raw power consumption, it outperforms any Intel machine we tested last year. It outperforms the M1 Pro and is close to the M1 Max in running real-world tasks. But this comes with some serious problems. First, raiders are more expensive than similar Apple machines (it also pairs these specs with stunning high-resolution screens that make this screen look like it belongs to Fisher-Price children's toys). Second, to achieve the same results that the M1 Max could easily achieve, the device consumes a lot of desktop-class power. Not only does our 16-inch MacBook Pro last more than twice as long on a battery as Raider, but it can also reproduce these benchmark results while running on batteries. Again, this comparison by Apple may not be important to Raider's target audience (gamers), but it is important to place these results in the broader context of Intel's path forward.
As Chin points out, this is a bit unfair because MSI laptops are geared toward AAA gamers who are used to always plugging in, while the new 16-inch MacBook Pro is more widely available to professional consumers who require easy portability. But it still hits all the x86 chip issues. They sacrifice too much power consumption in exchange for performance, especially for the marginal gains of the head. Of course, this is very important in non-gaming laptops, where battery life is the most important feature. Moreover, power and cooling expenses are important drivers of data center operating expenses. After we eliminated two business units, laptops and data centers accounted for two-thirds of Intel's revenue.
In summary, power consumption is the main problem of the entire x86 business being threatened.
Threat from ARM: Laptops
The threat of the ARM architecture to x86 is gradually becoming prominent. RISC-V is a new set of open source instructions that is freely available and used by enterprises. But RISC-V is still in its early stages, and it is almost entirely concentrated in the field of IoT chips. But never underestimate the power of time and cash flow, as this is the key factor in ARM's evolution to be able to replace x86. When you look back five to ten years from now, risc-V is likely to drive more industrial development.
It was a slow revolution that took place in three phases:
Phase 1: Intel and AMD missed a huge opportunity due to the inability to make energy-efficient chips, while ARM CPU chips dominate the smartphone and tablet markets. This market situation has been overwhelmed.
Phase 2: Companies like Apple and Qualcomm take cash flow from it and build ARM-based chips that replace x86 chips in PCs. This phase has already begun, and Intel has lost 10% of its revenue.
Phase 3: The power efficiency of ARM chips allows hyperscale data center enterprises to create denser data centers in the same place. Amazon's AWS Graviton3 chip is a forerunner in this phase.
As I said, Phase 1 is complete. ARM customers have shipped 220 billion chips, compared to 21 billion in the last nine months of 2021. Kissinger's biggest blunder of the previous CEO was when Steve Jobs contacted Intel in 2005 to customize chips for the iPhone, but Intel turned down the opportunity.
The best example of Phase 2 is Apple's current M1 chip. Starting with the CPU, the only product we're bullish on comes from Anandtech's adventure that began in October, when the 16-inch MacBook Pro first hit the market. So, these are comparisons with the previous generation of MSI laptops and their internal Intel chips. I still haven't seen a similar comparison of the new 12th-generation Core i9. But we have a great idea based on how they compare to the previous generation. (Benchmark score from Cinebench, a good test for high-end design applications.) ）
Source: Anandtech
The 12th-generation Core i9 may be about the same vertical level as the M1 Max, but pushing further to the right along the horizontal axis could potentially use about three times as much power as the M1 Max. In either case, Intel chips sacrifice too much power to get performance.
Move to multi-threaded performance and power consumption:
Similarly, we're likely to see the 12th-generation Core i9 rise vertically to Apple's level, but push it to the right horizontally, consuming more than twice as much power.
But performance and efficiency from the CPU is just the beginning. Laptops and mobile chips are often referred to as "system-on-chip" or SoCs because these chips don't just have a CPU on them. Intel and AMD's laptop chips integrate CPU and underpowered integrated graphics. High-performance laptops also require a separate GPU chip, which is also a lot of power consumption.
Not only do the M1 Pro and Max have more efficient CPUs that can be equal to or better than the performance of x86 laptop chips, but they package more features on the chip, speeding up common user tasks with lower power consumption:
More powerful integrated graphics than Intel or AMD, it's comparable to the standalone power-hungry GPU chips that come with high-end Windows laptops.
Provide a powerful machine learning core for AI tasks on the device.
Secure enclaves and encryption accelerators.
All I/O, including Intel Thunderbolt, which intel has never installed on the SoC.
Dedicated units for audio, video, and pictures.
So, on laptops, we see that Apple's ARM-based chips have a greater advantage than any x86 chip from Intel and AMD, and that they are built into brand new products, increasing Mac sales. After years of flat sales, the past 6 quarters have been the best-selling 6 quarters for the Mac ever (although the pandemic has a lot to do with that). Intel still has a long way to go to catch up.
But Macs make up a small percentage of laptop sales. The real question is whether anyone could create a Windows/Chromebook ARM platform to replicate what Apple has done. Qualcomm and Microsoft are working closely together on this. Before acquiring Nuvia, its first product, the Surface Pro X, ended in failure.
Nuvia was founded by Gerard Williams, who transformed Apple's chip design division from pretty good to the best in the world. He left at the end of 2019 and brought with him several other former Apple colleagues who were at Google at the time to found Nuvia. What they're looking into is a scalable CPU core with a performance/power profile similar to Apple's.
Now, Gerard Williams, a senior vice president of engineering at Qualcomm, brings in Nuvia's entire all-star team. Qualcomm's entire Snapdragon SoC line for smartphones/tablets, PCs and cars will begin to incorporate Nuvia's technology, with customer samples available starting in the second half of this year. We should start seeing products in mid-2023. If all goes well, I think they will also work on data center chips.
In my opinion, this is the biggest threat to Intel and AMD's laptop businesses, accounting for a third of Intel's revenue. Intel has already lost the opportunity to get 10% of its revenue through Apple, and now they may lose another third to Qualcomm.
Threat from ARM: Data Centers
Semianalysis' Dylan Patel expressed a positive view of Amazon's newly announced Graviton3 data center CPU chip:
While x86 CPU vendors will maintain their peak performance per CPU, Intel and AMD overlook the more important challenges. That is, the rack level with respect to the total cost of ownership of each compute unit on the server and the generic CPU. The CPU market has entered mass commercialization, and even if the individual core designs of Intel and AMD are significantly better, the equation will not change. Intel and AMD are highly focused on certain aspects that make them miss key factors in system-level design, such as too high peak power, too low density, and clock speed being pushed too far.
Let's break down superscale computations like AWS. The limited issue that hyperscale enterprises are dealing with is the number of their data center buildings, or more accurately, the number of server rack slots they can accommodate in those buildings.
What they want to do:
CAPEX: A calculated metric that maximizes each rack slot within its capex budget.
OPEX: Minimize maximum variable operating expense costs, i.e. power for running servers and cooling.
The disadvantage of ARM data center chips now is that each core is not as fast as Intel or AMD's x86 chips. However, the cost of x86 performance is the power they consume, and the heat they generate that must be dissipated. In a two-way competition, Intel and AMD have always assumed a maximum of two CPU chips per rack slot, and have designed power and heat dissipation around this.
As a result, ARM CPU chips also consume much lower power and generate much less heat. This allows hyperscale enterprises to make data centers much denser than when using x86 chips. This can be done with densely packed chips, such as ampere's privately held 128-core ARM data center chips (used by Oracle Cloud), which feature a standard 2-chip per-server slot design, or 64-core chips like graviton3, with a 3-chip configuration per socket.
ARM chips are much cheaper upfront, so capex will drop even in high-density configurations where the chip is increased by 50%. Even in these denser configurations, they use less power and generate less heat per slot, which means lower power budgets, as well as higher margins and lower prices for the impact on customers.
Amazon's claim in the demo is that with the x86 chip, the power consumption exceeds the power distribution of the standard 42-slot rack before filling the rack with the x86 chip. They couldn't fully fill the rack with x86 chips. The predecessor, the Graviton2, featured a design of 2 chips per socket, and when fully filled, they ended up using only a fraction of the rack's power distribution. Now, with a denser Gravicon3 design, Amazon is maximizing the computing power of each rack and still far below the power distribution per rack. Judging by their descriptions, it sounds like they may have enough power to move the four chips to each rack slot at some point.
That's the math Intel and AMD's biggest customers are working on, and more and more of the answer is arm chips of their own designs. Keep in mind that both Intel and AMD have new data center chips that will be available soon, but not included here:
Amazon has made a huge leap forward from Gravion2 here. In addition to a 25% increase in performance per core and a 60% increase in floating-point numbers over the previous generation in integer computing, they now encapsulate more than 50% of floating-point numbers in each rack. Just as Apple did with smartphones and now PC chips, Amazon prioritizes power consumption and heat dissipation. The result is Graviton3 points, all in the upper left corner, with higher performance per rack slot and much lower power consumption.
Translate that into a performance ratio per watt per rack slot and their lead is clear:
Similarly, both AMD and Intel will launch new chips in data centers within a few months, but that's a huge lead on this important metric.
Intel is already hit by AMD in the data center. You can see how bad their 2021 chips compare in the last two charts, and it doesn't look like it's going to get any better for a while. But the bigger threat for both is that their biggest data center customers want to roll out their own ARM chips because that makes so much sense.
Intel foundry business
So, now that we've solved the difficulty that Intel will be maintaining its legacy x86 business, let's talk about Intel's vision of catching up with TSMC and becoming a leader in wafer foundry.
Remember that the words "nano" and "nm" no longer have anything to do with the unit of measurement. They have been a marketing term for years. Now Intel has changed their naming conventions, and these numbers are more comparable than in 2020, which is the focus of the name change. The lower the number, the better.
TSMC Manufacturing Nodes by Revenue (TSMC Q4 2021 Presentation and Author's Notes)
This pie chart shows TSMC's manufacturing "nodes". At present, the most advanced in the world is TSMC's 5nm, followed by Samsung's 5nm. Intel's 12th Generation Core i9 is manufactured on its new 7nm node. I have noticed which Apple chips are manufactured on advanced nodes for reference. The A and M series run iPhones, iPads and Macs, the S series is in the Watch, and the U1 chip supports AirTags and other features. 5nm accounts for almost a quarter of TSMC's revenue, almost all of which are occupied by Apple.
The line I drew from the middle is the line that now separates the high-end process node from all the other nodes. This is what I call an ASML line, because in the case of TSMC, the 5nm and 7nm nodes are the nodes of the polar ultraviolet or EUV lithography machine that uses ASML. Any fab that hasn't invested heavily in this machine will stay in 2018 forever. They cost more than $100 million each, and the fourth-generation machines that will be delivered in 2025 will cost $300 million. Huge capital investment and export restrictions to China are why there are only three EUV customers: TSMC, Samsung and Intel. Intel won't have a process to use UV until later this year.
However, there is also a lot of action in the middle and low layers of the pie chart, especially now, because this is where the chip shortage is. It is also the first part of the business downturn to collapse. The 28nm node is now unusually busy, which is a crowded space for all foundries. One of the biggest drivers is the digital camera chip that converts light into digital data. But in 2019, due to weak demand, the foundry capacity was idle at 28nm.
The advanced nodes (5 and 7 nm) are smaller in size but have a much higher margin. Once these nodes mature (18-24 months after release), they will receive a gross margin of close to 60% until they are replaced by the top node of a new full ramp. This is where TSMC is now 5nm, and the ramp has been completed. When you browse the pie chart clockwise from 5nm, the profit margins go down as they face more competition and price pressure. That's why it's so important to be at the top.
So that's why it's so important that Intel climbs all the way to the top, both to save the x86 business and to make the foundry business more durable. At any given time, high end means high profits.
UMC is a very good Taiwanese fab, but they only work in low- to mid-range processes below the ASML line. As can be seen from the chart, the mid-2018 to the beginning of the epidemic is the last semiconductor downward cycle. Even in this case, TSMC did not lose a lot of gross margin, and UMC did lose.
Similarly, in the fab business, competition at the high end is key, otherwise gross margins will be seriously threatened cyclically.
But if you look at what's going on with UMC's gross margin right now, we're going to talk about chip shortages next.
Here's Intel's timeline for catching up with TSMC:
Intel and TSMC Investor Presentation
There's a big caveat about all of this, and that's that these timelines tend to be ambitious. In Intel's case, under the control of the previous management, they missed many node releases, which is why it was the former management. TSMC has postponed their 3nm node. However, beyond the usual challenges, Intel is talking about introducing 3 new process nodes in 18 months, as well as the two most important nodes on that timeline, 3nm and 2nm. Everything after 7nm uses an EUV. These were hard to use, and it took TSMC and Samsung years to figure it out. It's very ambitious, to say the least. The complete process of a node takes up to two years, and if they are to catch up with those dates proposed on Investor Day, Intel will have 4 processes at different stages in early 2024. They set themselves a daunting task.
I mentioned some of Intel's big capital expenditure figures earlier — $20 billion in Arizona, similar in Ohio, and similar in Europe if subsidies go through. In addition, they have installed all new equipment in the existing facility, including dozens of new EUV units. Subtract the capital expenditures that leave Intel's segment, they spent $18.7 billion on capital expenditures in 2021, and there will be more in 2022-2024. The final figures are highly dependent on subsidies from the United States and the European Union.
By comparison, TSMC spent $31 billion on capital expenditures in 2021 and another $40 billion to $44 billion by 2022, with similar figures likely in both 2023 and 2024. The spending midpoint is $125 billion over three years, probably at least double Intel's final spending. But TSMC has only added 2 new nodes, and its 4nm is actually an extension of the 5nm node for specific customers. Of the $125 billion, there's a message to Intel and Samsung: Do you think you can catch up? Try it.
But for Intel, that's not that grim. They have a big tailwind here, and they re-establish relationships with device manufacturers. This is especially true for ASML, which will be able to help Intel shorten the learning curve for TSMC and Samsung in using EUV by several years. In the past, Intel has acted as if they know more about how to use these tools than the device manufacturers themselves. It's a very welcome change and one of the major cultural shifts Kissinger has influenced over the past year.
Intel wants to make this compressed timeline work by running on both tracks at the same time. Each project has a specific end product or external customer. For example, the 4nm and 3nm processes will be superimposed on each other when the new foundry opens in Arizona in 2023:
Source: Intel
Intel needs to catch up with competitors in a shorter period of time, which is a tough road, and its main competitors have hardly stood still.
Chip cycle
At this stage, we are in the midst of the largest semiconductor upswing cycle ever recorded. There are three trends driving this now.
First, during the pandemic, demand for durable goods and equipment has surged. Semiconductor demand is closely related to the demand for durable goods such as PCs, phones, electronics, appliances, vehicles, and commercial/industrial equipment.
Consumer durables and commercial/industrial equipment as % of GDP (BEA)
This is the percentage of consumer durables and commercial/industrial equipment in GDP. After two decades of weak demand, the pandemic has led to a surge in demand for consumer durables, followed by a surge in equipment investment, a trend that continues.
The second trend is that consumer durables have a much higher silicon content than before. This is most evident in vehicles. For decades, each new feature, like a trunk release or heated seat, had its own cheap microcontroller to run it, and now there are dozens of new cars. Electric vehicles and assisted/autonomous driving add new chips to the combination of power management and high-end SoCs. Silicon is worth more than twice as much in electric vehicles as in internal combustion engine cars.
The last trend is non-durable goods, and we're starting to see that goods that never contained silicon are now equipped with very cheap ARM or RISC-V microcontrollers, such as single-use chips from the $25 COVID test kit.
The chip shortage is mainly due to the low-end process. The two types of sensors I've been hearing about are power management sensors and image sensors for digital cameras. The latter is the mid-range we're discussing, usually 28nm nodes, but the former used to be cheap commoditized chips on every board. These cheap chips have been at the heart of the shortage.
The reason is simple, and we've already discussed it: 20 years of poor demand, which is the low-margin part of the business. Investments in new capacity match demand, and new investments go primarily to advanced process nodes with higher profit margins.
Of course, this has changed dramatically now with the pandemic and the surge in demand for durables. Apple's Tim Cook very cleverly summed up what happened:
The pandemic ensued. Some people in the industry and some people outside the industry believe that the epidemic will reduce demand. They gave the order. Things reset. What really happened was that demand rose, even exceeding the prediction of a straight-line trend.
Now, we're seeing a surge in capex for foundries. From Intel and TSMC alone, they spent a total of $50 billion in 2021 and could spend about $175 billion to $200 billion in 2022-2024. Many of these billions of dollars are spent on ASML.
The danger for everyone, especially Intel, is the sharp decline in demand for consumer durables. This happens a lot.
Durable goods as a percentage of total consumption (BEA)
Demand for durable goods has a rapid downward trend, which is often a precursor to a recession. The reason durable goods tend to decline rapidly is in the name: durability. I've owned my last car for 14 years. At some point, everyone who wants a new 4K TV to watch the pandemic revelry has already got one.
So the biggest question for everyone, especially Intel, is "How long can this last?" "There's a scenario where Intel opens a new fab in Arizona in 2023, leading to overcapacity rather than a shortage like we have now. This is the biggest threat to the program beyond their control.
Summary and outlook
Intel's plan is to maintain its x86 CPU business while seeking to strengthen its partnership with TSMC. Both of these parts are daunting tasks.
In the short term, the threat to Intel's x86 business is competition from AMD.
But in the long run, the bigger threat is more efficient ARM chips in the laptop and data center space, both of which account for a third of Intel's revenue.
Apple was once Intel's fourth-largest customer, accounting for about 10 percent of Intel's revenue, but it has lost that customer.
Qualcomm's partnership with Microsoft is the biggest threat in the Windows/Chromebook space.
AWS has shown everyone that ARM chips can bring higher density in data center designs, as well as a combination of lower prices and higher AWS margins for customers.
Intel's foundry plans are ambitious and the planned catch-up time is very tight. It's unclear whether its supplier's help and dual-track approach will support Intel to add five new process nodes between 2022 and 2025.
Declining demand for durable goods is the biggest threat to Intel, just as they are putting all of these new capacities into use.
I still think the most likely outcome is that Intel x86 business shrinks, foundries still come in third place, have no pricing power, and are more sensitive to cyclical recessions.
In Long View Capital's view, the entire premise of ARM's portfolio is that 82% ($60 billion) of Intel's x86 chip revenue in 2021 will go to other companies. It has already started with Apple paying about $7 billion a year. Most of that $7 billion is now shared between Apple and TSMC. This could be a preview of Intel's future.
That sounds terrible! But my assessment of Intel is not as negative as it was a year ago. At the beginning of the article, the energy and cultural transformation brought about by Pat Kissinger is introduced, and now this concludes. I still think the plan is too ambitious and a lot of things are beyond its control. But a year ago, I didn't think Intel would go as far as it does today, and the biggest shift so far has been cultural. This is a multi-year turnaround project and the next steps will be the most difficult.
Regarding the HOLD rating, although it is negative, I think a lot of what I am talking about is already built into Intel's price: