Intel: A Year Into The Gelsinger Plan – Seeking Alpha

President Biden Speaks On Rebuilding Country"s Supply Chains In U.S.

Chip Somodevilla/Getty Images News

Chip Somodevilla/Getty Images News
That is the first word that comes to mind when I think of Pat Gelsinger, who is rounding out his first year as Intel (INTC) CEO. He has brought a tremendous amount of energy to what looked like a moribund company. In the past year:
That’s a lot! And I am leaving some smaller stuff out, too. Pat Gelsinger has brought a tremendous amount of energy to Intel, and it is not too much to say that just a year later, he had flipped the culture at Intel on its head. It’s quite an achievement.
But the question remains: is he running in the right direction, or just running?
The strategy can be summed up as: current Intel plus TSM (TSM) in 2025-2030. They would like to hold on to their current PC and data center business, and at the same time add on TSM’s operating statement.

Intel vs TSMC market cap and revenue
Data by YCharts

This is a very ambitious goal, most likely in my opinion to end in failure. The most likely outcome is that Intel’s foundries will remain in third place behind TSM and Samsung, and they will not have the pricing power that TSM enjoys at the top. At the same time, their x86 CPU business, their main business since 1981, will whither away with Intel and AMD hardware replaced by cheaper, more efficient and more flexible ARM and RISC-V hardware. Ironically, Intel’s new foundries will fabricate some of these, and they will be pulling far less average margin out of every laptop and server than they used to.
But again, I am fairly astounded at how much Gelsinger has accomplished in just a year. He has set forth highly ambitious goals for Intel, and the first steps have been good ones. I still see failure as a more likely outcome.
First, let’s look at Intel’s x86 and other businesses. The two main segments in Intel’s reporting are the “Client Computing Group,” or PC chips, and the “Data Center Group” for data center chips. These groups account for 84% of revenue, and 111% of operating income.

Intel revenue and operating income

Intel annual report

Intel annual report
The blue blocks are the x86 segments, and you see how they dominate. They will dominate more soon, because the red blocks, NVRAM and Mobileye, are being sold, so 7% of revenue and 9% of operating income will be leaving with them. The NVRAM sale is mostly complete, and the Mobileye IPO is slated for mid-2022.
But those blue blocks also contain “adjacent” revenue — stuff that is sold with the x86 chips like server boards, networking chips, memory and storage. These adjacent markets are highly dependent on the x86 CPU chips they are sold with, but let’s split that off so we are only looking at the x86 chips.

Intel revenue splits

Intel annual report

Intel annual report
So 76% of Intel’s top line came from the x86 chips themselves in 2021, and another 7.9% in adjacent revenue, highly dependent on the x86 sales. If we pull out the two units being unloaded, it is even more concentrated in the blue x86 wedges:

Intel revenue splits ex-NVRAM/Mobileye

Intel annual report

Intel annual report
The point here is that despite the other things they have going on, since 1981, Intel’s main business remains designing, fabricating and selling x86 CPU chips.
The problem with Intel’s x86 chips came into sharp focus with the release of their latest 12th Generation Core i9 chip for laptops. Again, laptops are about a third of Intel’s top line. In the past, Intel has started from the bottom of the line, the i3 or i5, and worked their way up, but this time they went for the flashy top end that will not sell many units because the laptops made around them cost over $3500, and are wildly impractical for anyone but gamers.
The reason they did that is because they were not just getting beat in performance by their main rival for years, AMD (AMD). Apple (AAPL) is now making the M1 Pro and M1 Max, not x86 chps but rather based around the ARM instruction set, and they beat up Intel’s 11th generation chips in performance comparisons. Apple was formerly Intel’s 4th biggest customer at around 10% of revenue, roughly $7 billion in 2019, according to Bloomberg.
So what Intel thought they needed was a splashy highly performant laptop chip that got these sorts of headlines.

Intel ‘Alder Lake’ laptops crush the competition


But there were also headlines that saw through the hype.


The Verge

The Verge
Back in January, Intel sent reviewers a $4000 17” laptop, the MSI GE76 Raider. They even sent out the low-resolution model, so the benchmark frame rates would be higher in the gaming comparisons with AMD. That second review from Monica Chin at The Verge got to the heart of the issue for Intel, and really for AMD as well:
When it comes to raw power, it’s outperforming every Intel machine we tested last year. It’s outperforming the M1 Pro, and it’s close to the M1 Max on the real-world tasks we ran. But that comes with some heavy caveats. First, the Raider is more expensive than a comparable Apple machine (which also pairs those specs with a stunning high-resolution screen that makes this one look like it belongs on a Fisher-Price Laugh and Learn). And second, this device is guzzling desktop-level wattage in order to achieve the same results that the M1 Max doesn’t even really need to try for. Not only did our 16-inch MacBook Pro last over twice as long on battery as the Raider did, but it can also reproduce those benchmark results while it’s running on battery. Again, this Apple comparison likely won’t matter to the Raider’s target audience [gamers], but it matters in terms of placing these results in the broader context of Intel’s path forward. [emphasis added]
As Chin points out, it’s a bit of an unfair comparison, because the MSI laptop is geared towards AAA gamers who are used to being plugged in all the time, while the new 16” MacBook Pros are for a wider range of professional customers who want more mobility. But it still hits at the problem for all x86 chips. They use too much power to get their performance, especially for marginal gains at the top. This is of course hugely important in non-gaming laptops where battery life is a top feature. But also, the power and cooling bill are big drivers of OpEx in the data center. Laptops and data center are two-thirds of Intel’s top line, after we subtract the two units that are leaving.
This is why the entire x86 project is under threat — power consumption.
I’m going to focus on the threat from ARM, as it is more present. RISC-V is a new open source instruction set, so it is free to use. It is still early days with RISC-V, and it is being used almost entirely for IoT chips. But never underestimate the power of time and cash flows, because that is exactly how ARM got to the point where they are trying to unseat x86. We may look back 5 or 10 years from now, and RISC-V may be pulling a lot more weight.
This is a slow revolution that is happening in three stages:
Stage 1, as I said, is complete. ARM customers have shipped a cumulative total of 220 billion chips, 21 billion in the last 9 months of 2021. That’s 2.7 ARM chips for every man, woman and child in the world, in just 9 months. Those could have been Intel or AMD chips. The biggest whiff from a Gelsinger predecessor was in 2005 when Steve Jobs approached Intel to make custom silicon for iPhone and they passed.
Stage 2 is best exemplified by Apple’s M1 chips right now. Starting with the CPU, the only good look we have here comes from Anandtech’s frisk from back in October when the 16” MacBook Pros first hit the streets. So these are comparisons to the previous generation of that MSI laptop and the Intel chip inside it. I have still not seen similar comparisons of the new 12th generation Core i9. But we have a pretty good idea, based on how they stack up to the previous generation. (The benchmark scores are from Cinebench, a good test for high end design applications.)

Performance-power consumption matrix: Apple vs Intel


The 12th Generation Core i9 is probably up at roughly the same vertical level as the M1 Max, but pushed much farther to the right along the horizontal axis, probably using around three times the power of the M1 Max. In either case, the Intel chips use too much power to get the performance they get.
Turning to multi threaded performance and power consumption:

Performance-power consumption matrix: multi threaded


Again, we probably would see the 12th Generation Core i9 up around Apple’s level vertically, but pushed way out to the right on the horizontal, using well over twice the power.
But the performance and efficiency coming from the CPU is just the start. Laptop and mobile chips are often referred to as a “system-on-a-chip,” or SoC, because they have more than just the CPU on the chip die. Laptop chips from Intel and AMD have a CPU and underpowered integrated graphics. Any high powered laptop also requires a separate GPU chip, which is also a big power draw.
The M1 Pro and Max not only have far more efficient CPUs that can equal or better the performance of x86 laptop chips, but they pack a lot more onto the die to speed up common user tasks at even lower power consumption:
So on the laptop, we see that Apple’s ARM-based chip has huge advantages over any x86 coming from Intel and AMD, and that is showing up in great new products, and increased Mac sales. After years of flat sales, Mac’s best 6 quarters ever are the last 6 quarters (though the pandemic has a lot to do with that). Intel has a very long way to go to catch up here.
But Macs are a small percentage of laptop sales. The real problem comes if someone can create a Windows/Chromebook ARM platform that copies what Apple has done. Qualcomm and Microsoft (MSFT) are working together closely on this. Their first shot, the Surface Pro X, was a misfire, but that was from before they got a lot of help in the form of Qualcomm’s acquisition of Nuvia.
Nuvia was founded by Gerard Williams, who turned Apple’s chip design unit from pretty good into the best in the world. He left in late 2019, and brought a couple of other former Apple colleagues who were then at Google (GOOGL) to found Nuvia. What they were working on were scalable CPU cores with a performance/power consumption profile similar to Apple’s.
Now, Williams is SVP Engineering at Qualcomm, and he brought the whole all-star team from Nuvia with him. Qualcomm’s entire Snapdragon SoC line for smartphones/tablets, PCs and vehicles will begin incorporating Nuvia’s technology, with customer sampling beginning H2 this year. We should start seeing products in mid-2023. If things go well, I think they will also work on a data center chip as well.
In my opinion, this is the biggest threat to Intel and AMD’s laptop business, a third of Intel’s top line. Intel already lost 10% of their revenue via Apple, and they could lose another third to Qualcomm.
Dylan Patel of Semianalysis put it nicely in reaction to Amazon’s (AMZN) newly announced Graviton3 data center CPU chips:
While x86 CPU vendors will maintain their peak performance per CPU lead, Intel and AMD are ignoring the more important battle. That battle is over total cost of ownership per unit of compute on a server and rack level for generalized CPUs. Commoditization is here for the CPU market, and even if Intel’s and AMD’s individual core design is markedly better, it won’t change the equation. Intel and AMD are hyper focusing on certain aspects, which make them miss crucial factors in system level design such as peak power being too high, density being too low, and clock speeds being pushed too far.
Let’s break down the calculations for hyperscalers like AWS. The finite thing hyperscalers are dealing with is the volume of their data center buildings, or more precisely, the number of server rack slots they can fit in those buildings.
What they want to do:
The disadvantage that ARM data center chips have right now is that, per core, they are not as fast as the x86 chips from Intel or AMD. But the cost to that x86 performance is the power they consume, and the heat they produce that has to be dissipated. In the two-way competition, Intel and AMD have always assumed a maximum of two CPU chips per rack slot, and they have designed power consumption and heat dissipation around that.
So ARM CPU chips’ also use far less power, and produce far less heat. This lets hyperscalers make the data center much more dense than if they were using x86 chips. This can be done with densely packed chips like privately-held Ampere’s 128-core ARM data center chip (used by Oracle Cloud) in a standard 2-chip per server slot design, or 64-core chips like Graviton3 in a 3-chip per slot configuration.
The ARM chips are much less expensive up front, so CAPEX goes down, even in a high density configuration with 50% more chips. Even in these more dense configurations, they use less power and produce less heat per slot, so that means a lower electricity budget, and some combination of fatter margin and lower prices to customers.
The way Amazon put it in their presentation is that with x86 chips, power consumption exceeds their power allocation for a standard 42-slot rack before they can fill up the rack with x86 chips. They cannot fully populate a rack with x86 chips. The predecessor Graviton2 is in a 2 chip per slot design, and they wound up using only a small fraction of the rack’s power allocation when fully populated. Now, with the denser Graviton3 design, they are maximizing the compute power of each rack, and still staying well below the power allocation for each rack. From the way they described it, it sounds like they may have power overhead to move to four chips per rack slot at some point.
This is the math that Intel and AMD’s biggest customers are looking at, and the answer they are increasingly coming up with is ARM chips that they design themselves. Keep in mind that Intel and AMD both have new data center chips that will be available very soon and are not included here:

Data center performance/power matrix

Anandtech; AWS; author estimates

Anandtech; AWS; author estimates
Amazon has made a huge jump here from the Graviton2. Besides a 25% per core performance improvement over the previous generation in integer computing and 60% higher in floating point, they are now packing 50% more of them into each rack. Like Apple has done with smartphone and now PC chips, Amazon has prioritized power consumption and heat dissipation. The result is that Graviton3 dot, all alone at the top-left, getting more performance per rack slot with far less power consumption.
Converting that to a ratio of performance per Watt per rack slot, their lead is stark:

Data center performance per watt.per rack slot

Anandtech; AWS, author estimate

Anandtech; AWS, author estimate
Again, AMD and Intel will both have new chips in data centers in a few months, but this is a dramatic lead right now in this all important metric.
Intel is already getting hammered in the data center by AMD. You see how poorly their 2021 chips compare in the last two charts, and that does not look to get better for some time. But the bigger threat to both is that their largest data center customers want to roll their own ARM chips, because it makes too much sense.
So now that we’ve dealt with the difficulty Intel is going to have maintaining their legacy x86 business, let’s talk about their new business that’s they’d like to add, TSM’s.
Please keep in mind that the words “nanometer” and “nm” no longer have any relation to the unit of measurement. They have been marketing terms for many years. Now that Intel has changed their naming convention, the numbers are much more comparable than they were in 2020, which was the point of the name change. Lower numbers are better.

TSM manufacturing nodes by revenue

TSM manufacturing nodes by revenue (TSM 4Q21 presentation and author annotations)

TSM manufacturing nodes by revenue (TSM 4Q21 presentation and author annotations)
That pie chart shows TSM’s manufacturing “nodes.” The most advanced in the world right now is TSM’s 5nm, followed by Samsung’s 5nm. Intel’s 12th Generation Core i9 is made on their new 7nm node. I’ve noted which Apple chips are made on the top nodes for reference. The A and M-series run iPhone, iPad and Mac, the S-series is in Watch, and the U1 chip enables AirTags and other features. The 5nm, almost a quarter of TSM’s revenue, is almost all Apple.
The line that I drew down the middle is the line now separating the high end process nodes from everything else. This is what I call the ASML line, because in TSM’s case, the 5 and 7 nodes are the ones that use ASML’s extreme ultraviolet, or EUV machine. Any fab that doesn’t invest heavily in this machine is stuck in 2018 forever. They cost over $100 million apiece and the 4th generation machine to be delivered in 2025 will cost $300 million. That, and export restrictions to China are why there are only three EUV customers: TSM, Samsung, and Intel. Intel will not have a process using EUV until later this year.
But there is also plenty of action in the low and mid slices of the pie, especially now, because that’s where the chip shortage is. It is also the part of the business that collapses first in a down cycle. The 28nm node is exceptionally busy now, and this is a crowded space at all foundries. The big driver of that is digital camera chips that turn light into digital data. But in 2019, foundries were idling capacity in 28nm because of slack demand.
The high end nodes, 5 and 7 nm, are lower volume, but much higher margin. Once those nodes mature (18-24 months after launch), they get something approaching a 60% gross margin until they are replaced with a new fully ramped top node. That’s where TSM is at with 5nm right now — ramp completed. As you go around that pie chart clockwise from 5nm, the margin declines because they face more competition and price pressure. That’s why being at the top is so important.
So this is why it is so important for Intel to claw their way to the top — both to save the x86 business and to make the foundry business more durable. The high end is where the high margin is, in both good times and bad.

TSM vs UMC gross profit margin
Data by YCharts

UMC (UMC) is a very good Taiwanese fab, but they only work in the medium and low end processes, below the ASML line. I backed out the chart so you could see the last semiconductor down cycle, mid-2018 through the beginning of the pandemic. Even under those circumstances, TSM didn’t lose a lot of gross margin, though UMC did.
Again, in the fab business, it is key to compete at the top end, or gross margin comes under severe cyclical threat.
But you see what is happening now to UMC’s gross margin, and we’ll talk about the chip shortage next.
Here’s Intel’s schedule for catching up to TSM:

Intel and TSM node timelines

Intel and TSM investor presentations

Intel and TSM investor presentations
There is a big caveat with all this, which is that these schedules tend to be aspirational. Sometimes they hit them, but sometimes they don’t. In Intel’s case, they missed numerous launches under previous management, which is why they are previous management. TSM has pushed back their 3nm node already. But aside from the usual challenges, Intel is talking about introducing 3 new process nodes in 18 months, and the two most important on that timeline, 3 and 2, within a year of each other, and then another one, not pictured, right after that. Everything after 7nm uses the EUV machines. These are very hard to use, and it took TSM and Samsung years to figure it out. This is, to say the least, very ambitious. A full ramp of a node takes up to two years, and Intel will have 4 ramps at different stages in early 2024 if they were to hit those dates they presented at their investor day. They have set before themselves a monumental task.
I mentioned some big CapEx numbers for Intel earlier — $20 billion in Arizona, a similar amount in Ohio, and something like that in Europe if subsidies come through. And then there’s also all the new equipment they are putting in existing facilities, including dozens of new EUV machines. Subtracting CapEx for segments that are leaving Intel, they spent $18.7 billion on CapEx in 2021, with a lot more to come in 2022-2024. The final figure is highly dependent on US and EU subsidies.
By contrast, TSM spent $31 billion on CapEx in 2021, will spend another $40-$44 billion in 2022, and probably a similar number for each of 2023 and 2024. The midpoint of their guidance was $125 billion over the three years, likely at least double what Intel will probably wind up spending. And they are only ramping 2 new nodes — their 4nm is really an extension of their 5nm node for specific customers. There is a message to Intel and Samsung in that $125 billion: “You think you can catch up? Try it.”
But it is not all that grim for Intel. They have a big tailwind here — their renewed relationships with the equipment makers. This is especially true regarding ASML, where they will be able to help Intel cut years off the learning curve TSM and Samsung had in using EUV. In the past, Intel had always acted like they knew how to use these tools better than the equipment manufacturers themselves. This is a very welcome change, and one of the big cultural shifts that Gelsinger has affected in the past year.
Intel hopes to make this compressed timeline work by going on two tracks simultaneously. Each track will have a specific end product or external customer. For example, the 4nm and 3nm ramps, which will be right on top of each other in 2023 as the new Arizona foundries are opening:

Intel 3 nm and 4 nm nodes

Intel presentation

Intel presentation
They have a tough road to catch up on a compressed timeline, and their main competitor is hardly standing still.
Right now we are in the middle of the biggest semi upcycle of all time. There are three trends pushing this right now.
The first is that demand for durable goods and equipment has surged during the pandemic. Semiconductor demand is very much tied to demand of durables — PCs, phones, electronics, appliances, vehicles and commercial/industrial equipment.

Consumer durables and commercial/industrial equipment as a % of GDP

Consumer durables and commercial/industrial equipment as a % of GDP (BEA)

Consumer durables and commercial/industrial equipment as a % of GDP (BEA)
That’s consumer durables and commercial/industrial equipment as a percentage of GDP. After two decades of slack demand, the pandemic saw a surge in consumer durables demand that was then followed by an investment surge in equipment, which is still going on.
The second trend is that consumer durables come with much higher silicon content than they used to. This is most evident in vehicles. For decades, every new feature like trunk release or heated seats has gotten its own cheap microcontroller to run it, and now new vehicles have dozens of them. EVs and assisted/self-driving have added new chips to the mix in power management and high end SoCs. The value of the silicon in an EV is over double the value in an internal combustion car.
The final trend is in nondurables, where we are starting to see goods that never had silicon in them now coming with very cheap ARM or RISC-V microcontrollers, like this single use disposable chip from a $25 COVID test kit.

The chip shortage is largely in the low end processes. The two types of chips that I keep hearing called out are power management chips and image sensors for digital cameras. The later is in the mid-range we discussed, often the 28nm nodes, but the former is what used to be cheap commoditized silicon that goes on every circuit board. These cheap chips have been the at the center of the shortage.
The reasons are simple, and we’ve already discussed them: 2 decades of poor demand, and this is the low margin part of the business. Investment in new capacity matched demand, and what new investment there was largely went into new advanced nodes with higher margin.
That has now changed dramatically of course with the pandemic and the surge in consumer durables demand. Tim Cook of Apple summed up what happened very neatly:

The pandemic came along. Some people in the industry and some people outside the industry thought that the pandemic would reduce demand. They pulled their orders down. Things reset. And what really happened was demand went up, and went up even more than a straight trend would predict.
-Tim Cook, October 28, 2021
Now we see a flood of CapEx in foundries. Just looking at Intel and TSM, they spent a combined $50 billion in 2021, and will probably spend about $175-$200 billion in 2022-2024. Many of those billions go to ASML (ASML).
The danger here for everyone, but especially Intel, is a crash in consumer durables demand. This happens frequently.

Durable goods as a % of all consumption

Durable goods as a % of all consumption (BEA)

Durable goods as a % of all consumption (BEA)
Durable goods demand has a tendency to drop off quickly, often as a precursor to a recession. The reason durables tends to drop off quickly is right there in the name: durable. I owned my last car for 14 years. At some point, everyone who wanted a new 4K TV for pandemic binge watching has already gotten one.
So the big question for everyone, especially Intel is “how long does this last?” There is a scenario where Intel opens up those new Arizona fabs in 2023 into a glut of capacity, not a shortage like we have now. This is the biggest threat to their plans that is out of their control.
You just made it through 4800 words. I wish I had some sort of medal for you, but in lieu of that, a bulleted summary:
At Long View Capital, the entire premise of the ARM Portfolio is that the 82% of Intel’s 2021 revenue that is x86 chips, $60 billion, will be going to other companies. It has already begun with the roughly $7 billion a year Apple used to pay them. Most of that $7 billion is now split between Apple and TSM. This is likely a preview of Intel’s future.
Boy, that sounds bad! But I am not nearly as negative on Intel as I was a year ago. I started this article talking about the energy and culture shift that Pat Gelsinger has brought, and that’s how I’ll end it. I still think the plan is too ambitious, and much is also out of his control. But a year ago, I did not think Intel would be as far as they are today, and the biggest shift has been cultural so far. This is a multiyear turnaround project, and the next steps will be the hardest.
Regarding the HOLD rating despite the negative take, I think a lot of what I am talking about has already been built into Intel’s price:

Intel vs AMD valuation
Data by YCharts

At Long View Capital we follow the trends that are forging the future of business and society, and how investors can take advantage of those trends. Long View Capital provides deep dives written in plain English, looking into the most important issues in tech, regulation, and macroeconomics, with targeted portfolios to inform investor decision-making.
Risk is a fact of life, but not here. You can try Long View Capital free for two weeks. It’s like Costco free samples, except with deep dives and targeted portfolios instead of frozen pizza.

This article was written by
Confirmation Bias Is Your Enemy.
Tech and macro. Deep analysis of long term sectoral trends, and the opportunities arising from them. I promise not to bore you. Author of Long View Capital, a Marketplace service for long-term investors. Risk Factors: I am also wrong sometimes.

Disclosure: I/we have a beneficial long position in the shares of AAPL AMZN ASML GOOGL MSFT QCOM TSM either through stock ownership, options, or other derivatives. I wrote this article myself, and it expresses my own opinions. I am not receiving compensation for it (other than from Seeking Alpha). I have no business relationship with any company whose stock is mentioned in this article.


By Kwetu Buzz

Leave a Reply

Your email address will not be published.