Today we are reaching the limitations of Moore’s Law where semiconductor efficiency is at a point where it can be scaled all the way from the world’s most powerful supercomputer (Fugaku) down to the phenomenal capabilities exhibited in products as diverse as Apple Airpods and autonomous vehicles. To get to this point though, it’s taken an entire ecosystem of innovators both upstream (chip fabs, designers) and downstream (CPS, GPU, SoCs) all the way through to the Original Equipment Manufacturers (OEMs). This piece breaks down what is quite a complicated piece of technology into simple explanations so you can understand what semiconductors are, how we’ve arrived here and where we’re going. I’ll conclude with a reasonably thorough overview of key (public) market participants across the ecosystem; with a particular focus on cloud/data centre technologies so you can see where opportunities may lay.
There’s a lot to squeeze into this article. I’m going to start by providing an overview of cloud and edge computing and the companies, at a high level, who provide the infrastructure that ties all of that together (data centres, infrastructure owners and network providers). Then I’ll take a step back and look at the semiconductor landscape that powers everything from the cloud to the edge - looking at little at the history and the progress that has driven us to where we’re at today. And along the way you’ll see why the industry is centre stage of a heated China / US trade battle, who is stuck in the middle (i.e. TSMC!) and how to navigate yourself through the investment landscape.
To help you understand this, let’s start with a typical ‘use case’ of how someone would engage with cloud and edge devices in their day-to-day lives.
Let’s assume you’re not in hotel quarantine and instead you’re walking down the local high street to the local cafe for your morning take away extra hot wheat milk decaf latte. As you’re doing this you’re listening to the latest Cardi B single WAP (#1 in the US as we speak) on Spotify from an Apple Watch with Apple AirPods.
This journey (ex-the music and coffee choice) is a typical day-to-day experience for hundreds of millions of people around the world. But what is behind the scenes of this typical experience is something quite incredible.
In this example, the watch and AirPods are what would be called ‘Edge Devices’. These devices ‘speak’ to edge nodes (iPhone, telco tower) which in turn pulls data from Spotify (or Apple Music’s) servers in the US. All of this happens seamlessly and is outlined below in a very high-level overview of the cloud and edge landscape.
Another common example of cloud-to-edge is autonomous vehicles (or what we have today — semi-autonomous vehicles). These cars are embedded with processors, graphics cards, servers etc all using advanced semiconductor technology. They pull data from traffic lights (edge devices), continually scan the environment with a series of integrated LiDAR and radar cameras for cars, pedestrians and stop signs which then seamlessly overlays with onboard (and cloud) memory of the streets/landscape. In this instance, the car needs on-board processors and memory so it can react (with near-zero latency) to obstacles. Pulling data directly from the cloud (like in the Spotify/Apple use-case) would be fatal in this instance.
The one common denominator across this ecosystem — and the only reason this technology exists is because of semiconductors. Advancement in semiconductor technology power everything from the tiniest processor (in an AirPod) to the world’s fastest processors (in the data centre).
The Evolution of Semiconductors
As we know, the brain is made up of memory (temporal lobe), vision processing (occipital lobe), logic processing (frontal lobe) and so on (see below). What drives the processing between all of these various lobes are the neurons (or nerve cells) whose job it is to process all of the information. On average we have around 86 billion neurons in our brains.
Now, at a high level, this isn’t too dissimilar to what’s in a computer/server/SoC (more on that below).
On a server, we have areas that control vision (GPU), logic (CPU) and memory (RAM). What’s behind all of those are individual conductors that process all the information and switch instructions on/off. These are the semiconductors (this is all extremely dumbed down so apologies to the purists out there!)
Now, let me take this analogy a step further.
Through evolution, we’ve gone from being a single-cell / quite basic organism to one which has billions and billions of cells and neurons; giving us the advanced ability to gather rocks on Mars (or take a selfie).
Just as we’ve evolved to be able to do more complex tasks, so have computers.
This advancement has happened because of Moore’s law, the premise of which, is that the number of transistors you can get on an integrated circuit board doubles every two years (see below); thus increasing efficiency. Think of that transistor count as a neuron count and those blue dots as different species on an evolutionary path.
The law’s namesake, Gordon Moore (now 91), went on to found Intel back in 1968 and it was there where he helped lay the foundations for the X86 chip architecture which was the first commercial CPU (highlighted on the left-hand side of the below ‘evolution’ diagram).
Intel were the pioneers and leaders for along time.
However, this is where the plot thickens! In the 90s, England’s Acorn Computers (founded 1978) spun their Acorn RISC Machine (ARM) unit into a separate business — with Apple buying a 30% share. The aim of this business was not to manufacture and sell their chip (as Intel do) but to license their design to Original Equipment Manufacturers (OEMs) like Apple.
Almost every single semiconductor out there today is based on one of these two architectures.
So what’s the difference between Intel’s X86 architecture and ARM?
Put simply, Intel’s architecture is complex and ARM’s is simple.
Intel’s X86 architecture is based on a Complex Instruction Set Computer (CISC) which churns through considerable power but generate better performance. ARM, on the other hand, created an architecture based on the Reduced Instruction Set Computer (RISC) — better energy efficiency but less performance.
This made ARM ideal for smartphones and smart-watches where watts need to be used more efficiently. X86, on the other hand, was better suited to laptops, desktops and servers which required more processing power (and more watts).
It’s really all a trade-off of processing power per watt.
So, with one chip being more powerful / less energy efficient and the other being less powerful / more energy-efficient — they respectively gained monopolies in key market segments. X86 (Intel & AMD) took a monopoly of the data centre and server market whilst ARM took a monopoly of the mobile and tablet market.
Now let’s take the lid off at the server end to see what’s under the hood.
Data Centre Servers
Every day we’re demanding considerably more computing power, and over the next decade, this is only going to accelerate as AI, AR, automation and various other technology gains prominence.
This requires semiconductor innovation at the edge and at the server.
When you lift the lid on a server, it will generally be broken into a few core components — predominantly:
- Processor (dominated by Intel and AMD)
- Graphics card (predominantly led by NVIDIA and AMD)
- Accelerators (Intel, NVIDIA, Marvell, Xilinx)
- Memory (Micron, Samsung, SK Hynix); and
- Other: Network adapters, switches, pre-configured rack servers etc
There are a few organisations — at the data centre level — who dominate the landscape; and this is outlined below:
Now, NVIDIA, who were the GPU pioneers in gaming, found that those same processing units were equally (if not more) important in servers (as more and more computing power was required for the advanced tasks demanded of them).
Their latest DGX A100 data centre platform (which I’ve loosely categorised as a ‘rack server’ ) is a good example of how this (CPU, GPU, memory) all comes together into a comprehensive (and exciting) product.
Nb. NVIDIA beat Intel, Microsoft and Xilinx to buy Israeli networking company Mellanox in 2019 which forms part of this platform. The key component within the platform that NVIDIA does not own is the CPU (which in this case is 2x AMD EPYC processors).
As mentioned, Intel (and AMD) semiconductors dominate the data centre space. They’re in a vast majority (if not all) major data centre servers today.
Now, this is where it all starts to get a bit shaky for Intel.
Intel are the only semiconductor company that is truly vertically integrated.
They design, manufacture (“fab”) and market their X86 processors. On the other hand, ARM ‘only’ own the IP/design and license that IP to over 500 partners (OEMs) — Marvell Technology, Qualcomm, Nvidia, Fujitsu, Samsung, Tesla, Amazon and Apple (everyone but Intel and AMD). Those companies then design their own chips (often with outside help) and outsource to either TSMC or Samsung to manufacture/fab those chips. But it’s not just ARM chips those fabs produce; they also churn out a superior (vs Intel) X86 for AMD.
In fact, TSMC and Samsung are killing Intel on the fab side and this is seriously impacting the company.
TSMC and Samsung are currently producing 7nm chips (both ARM and X86) whilst Intel are still on the 10–14nm chip (less efficient). When Intel’s 7nm chip hits the market (in 2022), TSMC and Samsung will already be rolling out 5nm chips. So Intel — the pioneers — are now an entire generation behind the curve.
How does this play out commercially?
On the server level, Amazon Web Services, who dominate cloud (and ultimately depend on Intel and AMD X86 processors for their data centres) have developed their own processors (Graviton2) based on ARM which, according to Amazon, “provide up to 40% better price performance over comparable current-generation x86-based instances”.
This is a 7nm ARM chip produced by TSMC (remember Intel don’t have the tech to build 7nm yet).
If this isn’t concerning enough for Intel, we’re also seeing the world’s fastest computer (Japan’s Fugaku) using a Fujitsu ARM processor (again, 7nm) produced by TSMC.
So Intel are under threat at the top end of town where processing power is critical.
They’re also continuing to see their market share being eroded as you go beyond the cloud to the edge.
Apple’s watches and iPhones (as well as all other wireless devices) are already on ARM but only recently the company announced that the Apple Mac is, for the first time, shifting from an Intel-processor to an in-house designed ARM ARM SoCs (System-on-a-Chip). This follows the lead of Tesla who also designed their own ARM SoCs (for self-driving cars), manufactured by Samsung.
The best way to think of a System on a Chip is to get everything you have in a server (GPU, CPU) and shrink it into one chip (instead of multiple chips). Below is a simple breakdown of Qualcomm’s Snapdragon chip that’s used in cell phones (their latest version has 5G capability squeezed in as well).
None of this really bodes too well for Intel.
They still completely dominate the server space (i.e. their data centre revenue for 2Q20 was up 43% YoY) but it’s quite clear that there are challenges for the business with increased competition and their lagging fab capabilities.
If you don’t believe me, take a look at how the market has rewarded TSMC (the leading fab) vs Intel (the lagging fab). TSMC is up 31% vs a 20% decline for Intel.
So, that’s X86……..what about ARM?
The Future of ARM
It is widely reported that ARM’s owner (Softbank and their Vision Fund) are looking at options for the business.
In a perfect world, Softbank would likely wait a few years to consider these options, however, given a number of factors (i.e. market strength, Mayoshi-son’s bruised ego and the WeWork debacle) it appears a 2020 deal may be likely. These options would include three possible avenues:
- Private sale to NVIDIA
- Public sale (IPO); or
- Public sale with NVIDIA (and others) as a cornerstone investor
There’s a lot already written on this but in summary, the suggest valuation is around US$40-$45bn. At the upper end; that’s ~25x fwd P/S. By comparison (on an estimated forward P/S basis) Nvidia is trading at 20x, AMD 9x, Marvell 8x, TSMC 7x and Intel at 3.5x. So, despite its dominance — the valuation is a bit rich.
Breaking all of the avenues down, I’d say option #1 is the least likely outcome. If regulators approve NVIDIA owning 100% of the IP for the world’s most expansive chip architecture then they’ve probably failed in their job as regulators (unless there are huge restrictions, which would see the deal stall at the starting grid).
Then there’s the option of going to the markets with an IPO. In this instance, Softbank (and the Vision Fund) would likely retain a stake (i.e. 30%) and cash out the rest (likely resulting in a special dividend, with a lot of the cash staying on the balance sheet for the next big deal).
The final option is a hybrid with two-classes of shares; Class A with NVIDIA and Softbank (with more voting rights) and Class B going to the public. Not an option that’s been thrown around but certainly something which is within the realms of possibility.
Obviously NVIDIA coming on as a cornerstone shareholder or outright owner would, at the right price, be a significant milestone for NVIDIA because, as you saw with their latest server, there are a couple of gaping holes in their business (namely their lack of CPU).
The Ecosystem (End-to-End)
We now know (hopefully) a little of the history of semiconductors and roughly where the two main architectures, ARM and X86 sit within the edge and cloud landscape.
I’d like to wrap this up by diving in a little into the broader ecosystem and highlight some of the leading corporations in that ecosystem, starting with the manufacturers (fabs).
There’s a vast chasm between the capabilities of the market leaders in semiconductor manufacturing (fab) and the rest.
In essence (at the pointy end of 7nm/5nm chips) there are only really two companies — Taiwan Semi (TSMC) and Samsung who are worth mentioning.
Taiwan Semiconductor Manufacturing Co (TSMC) generate ~US$40 billion in fab revenues (annualised off 2Q20). In fact their 2Q revs were up a massive 34% YoY with roughly one-third of that revenue came off the back of the advanced 7nm process and most of that growth attributable to high-performance computing (i.e. data centres/graphics cards). In fact, quarter-on-quarter, this was the only segment in which they saw growth (smartphones, IoT, auto, networking/DCE) were all off. They’re also world leaders in 5nm technology, having completed the industry’s first design infrastructure for the technology which is 15% faster and 30% more power-efficient than the current 7nm technology.
Next best, for high-end processors, are Samsung who are really the only company I know of in the world who are 100% vertically integrated (i.e. they make the chips that go in their branded phones/TVs/white goods etc). Despite fab revenue not being broken down, it is estimated to be around US$13 billion (which is ~25% of overall revenue). Like TSMC they are also producing the 7nm chips and have plans to roll out production of 5nm in the next 12 months with (like TSMC) a dedicated 5nm fab facility.
So, as TSMC and Samsung go about lapping the field of competitors there’s, of course, the elephant in the room, which is China.
This is where it starts to heat up.
It’s been reported that China are hacking corporations, stealing IP and “pillaging Taiwan's Semiconductor industry”. I won’t weigh in on this here but I’d expect such reports will only escalate leading up to the November’s US elections (along with further antagonism between China, US and their key allies).
However, casting aspersions aside, what’s clearly evident is that China is dependent on Taiwan‘s semiconductor industry. In fact, unlike any other technology or capability in the world — every single OEM across the globe that develops cutting edge technology needs TSMC (or Samsung).
So, if they (China) are to maintain their mantle (or aspiration) as a technology leader, they either need to keep a strong supply pipeline coming from TSMC and Samsung or build their own capabilities.
In terms of domestic capability, the Government are investing huge sums of money (including a recent ~US$30 billion semiconductor fund) to create an edge domestically. Much of this centres around a few companies such as Tsinghua Unigroup (private), Semiconductor Manufacturing International (SMIC) and Advanced Micro Fab Equipment (AMEC).
However, despite all of the money thrown at these companies and the local industry as a whole, it will still take some time China to get anywhere near parity vs TSMC and others. Various reports indicate that SMIC may start producing 7nm chips in 2021 which would put it, best case, at parity with Intel and 3–4 years behind TSMC and Samsung. Then they’ll need to close the gap on 5nm. Even if they hire the very best engineers and develop or ‘procure’ advanced semiconductor IP they will still be way behind the curve. Despite this, SMIC stock is absolutely surging; trading over 125x forward P/E and 25x forward sales. Compare that to the only other pure-play fab, TSMC, which trades at ~25x fwd earnings and 8x fwd sales.
It’s worth noting that a lot of that premium comes down to domestic hype of SMICs local shares (or ‘A’ Shares) by Chinese retail investors after the stock listed in Shanghai (in July 2020). The stock jumped 245% when it listed on China’s NASDAQ equivalent, STAR, in July 2020. The (a) reason for this listing is that Chinese companies who are ‘dual-listed’ in the US are under more strict regulations. Section 945 (Holding Foreign Companies Accountable Act) now requires Chinese companies listed in the US to (amongst other things) certify that they are not under the control of foreign governments, and have their books audited in the US (a rule reinforced by the Lukin coffee debacle).
Despite TSMC being the only real ‘pure-play’ fab peer for high-end semiconductors, I’ve outlined where SMIC are trading vs a pool of companies that have various in-house fab capabilities (just so you can see how incredibly expensive the stock is!). Note — this doesn’t mean the others are cheap! It’s purely an illustration.
So that’s Chinese domestic capabilities and challenges.
Then there are the external roadblocks around trade. On this front we’ve seen the US successfully negotiated for TSMC to establish a high-end chip production facility in Arizona — which would be the most advanced (i.e. 5nm production) outside of Taiwan when it opens in 2024. China won’t have a TSMC 5nm foundry (n.b. there’s also some spurious concern that China may ‘acquire’ or nationalise such plants!). But here’s the bigger point — as well as influencing TSMC to establish manufacturing capability in the US, the current US administration have also put rules in place whereby companies who use US technology (such as TSMC!) need to apply for a license to export products to Huawei! As a result, TSMC cut off the company from their supply chain!
Earnings / Valuation Overview
So, you can see now where we’ve come from in the semiconductor industry and where we seem to be heading — a China v USA trade war which is broadening into a China v Everyone Else trade war. Regardless of who wins the election, I wouldn’t expect the US to take the foot off the pedal when it comes to IP and the tech trade war.
I’ll leave you with a couple of overviews charts so you can see who is doing what across cloud and edge ecosystem as well as who is worth what (again, on a fwd P/E and P/S basis).
Firstly, a snapshot of upstream semiconductor companies — designers, equipment providers, testers, fabs etc. This isn’t exhaustive but shows you a good cross-section of some of the key players in the upstream chip business.
Outside of Intel, the US is hugely dominant in downstream chip ‘sub-sectors’. For example, companies like Synopsys provide TSMC designers with leading IP, tools and design flows which are critical to TSMC (and others) ongoing success. Applied Materials fab process technology is widely adopted whilst other US-domiciled organisations like KLA-Tencor, LAM Research and Cabo Microelectronics provide various other value-add products and services helping to advance upstream fab technology.
Outside of the US, you can look to Netherland’s ASML who are the world’s sole supplier of EUV lithography machines; critical to the production of leading semiconductors (50% of their sales went to TSMC!).
Rounding out the list you have Japan’s Tokyo Electron (semi equipment), South Korea’s Samsung and SK Hynix (not pure-play downstream though!) and China’s Advanced Micro-Fabrication Equipment (AMEC) — developers of MOVPE equipment necessary for creating semiconductor structures.
Side-by-side (left) you can see how they all stack up on a forward (next financial year) price/earnings basis. I won’t dive into detailed forecasts here, but what you can see here is that quite clear chasm (highlighted above) with those Chinese semi companies. There’s 2% of me that thinks — if China procures the vast majority of their chips from SMIC into perpetuity — that that premium may be justified. However, 98% of me says no. Priced into that is considerable hype for a company that is significantly behind leading chipmakers. Then there’s the whole issue of what happens to chip progress when we get to the outer realms of Moore’s Law — 3nm and beyond (most of which is yet to be solved for!).
I’ll leave you with a super quick snapshot of some of the downstream / OEM companies in the semiconductor industry.
You’ll see below, this is also broken down at a high level based on forward P/E and is further sub-categorised (quite broadly) into:
- Chips (primary focus on SoEs and similar chips — commonly seen in edge applications such as smart homes, AVs and similar applications)
- Diversified (mostly multinationals producing CPUs, GPUs, SoEs, accelerators etc)
- Memory (primary focus on memory devices/products); and
- Other (switches, controllers, amplifiers, PCB etc)
What this piece aimed to do was provide a high-level overview of quite a complicated ecosystem.
I believe this is important as the investment case for semiconductors (upstream and downstream) and as well as their various applications areas such as autonomous vehicles, 5G, data centres, headphones, AI, AR/VR etc will only ever be front and centre.
No doubt, many of the companies above will provide considerable investment opportunity (but also considerable risk). Again, some of these opportunities and risks have been presented at a high level and of course, the landscape we see now — although dominated by a few — may be very different in a few years, a decade or a couple of decade as technology accelerates at such a rapid rate.
As usual, invest wisely and feel free to connect if you have any burning questions or need more details on specific technologies (or companies).