The Semiconductor Industry from 10,000 Feet

Semiconductors are mission-critical to modern life. They power everything from smartphones and laptops to cars, AI, industrial automation, and the internet of things. There are few areas of life semiconductors that have been left untouched. And yet, per capita consumption of semiconductors looks virtually certain to increase.

Moore's Law has democratized semiconductors. The "law" is actually an observation and extrapolation of a historical trend rather than an iron rule of physics. Nonetheless, it's been about right. It states that every 12-24 months, we get roughly double the computing power (transistor density) for the same price and at the same margin to the manufacturers. Win-win-win. Semiconductors are now affordable enough to embed in virtually any product.

Given the significance of semiconductors and their increasing importance to modern life, I thought I'd take some time to write about the current landscape, how we got here, and where we might go. Be forewarned that I consider semiconductors outside of my circle of competence. You're unlikely to see any of the companies discussed in our portfolio anytime soon.

Historically I've shied away from semiconductors because they are:

Among the most complicated machines mankind has ever built;
Cyclical;
Capital intensive; and
Subject to rapid technological change.

These are the polar opposite characteristics of the simple, predictable, profitable, and replicable businesses that Dan and I prefer.

Two observations initially stoked my curiosity. First, semiconductor companies are consistently among the world's largest by market cap. Second, there is an unusually wide valuation spread between the industry leaders. Intel trades for just 11x earnings, while AMD, its closest competitor, trades for 129x earnings. What's going on?

Source: Author, Data from Google Finance and Value Line

History

AT&T's Bell Labs invented the first transistor in 1947 and in 1954 Texas Instruments produced the first commercial silicon transistor. Texas Instruments went on to build the first integrated circuit in 1958, which won Jack Kilby the Nobel Prize.

In the early years, semiconductor companies like Texas Instruments used a soup-to-nuts vertically integrated model. TI designed and built their own chips and the end products too, like their famous calculators. Intel, which also came of age during these early years, continues to use a similar strategy. Intel is an IDM (integrated design manufacturer) because it designs and manufactures chips.

Morris Chang worked for TI between 1958 and 1983, ultimately becoming VP of Semiconductors. Chang pioneered the strategy of aggressively pricing chips ahead of TI's cost curve. He was willing to sacrifice initial profits to gain market share that would yield greater long-term profits. Controversial at the time, this is now standard practice.

In 1983 TI passed over Chang for the company's top spot. Disillusioned, Chang left TI and went home to China. Taiwan recruited him to head a non-profit promoting industrial and technological development. In 1987 Chang founded Taiwan Semiconductor Manufacturing Company (TSMC), which pioneered the fabless foundry model. One of the macro trends in semiconductors is the battle for dominance between the IDM model and the fabless foundry model.

In the late 80s, American companies began outsourcing their manufacturing to lower-cost Asian countries. TSMC hopped on this train by offering to build semiconductors to any company's specification. TSMC aggregated demand from several firms and built unmatched scale. Scale gave TSMC a cost advantage and enough R&D dollars to give it a good shot at staying on the cutting edge.

The fabless foundry model also allowed other firms to focus completely on designing chips without worrying about making them. These companies, like NVIDIA, operate more like software companies. Engineers design chips and then basically email the file to TSMC for production.

Below is a map industry map of the industry’s largest players.

Geopolitics

TSMC is the technological leader in semiconductor manufacturing and one of the world's most valuable companies (currently #11). Beyond market value, TSMC carries significant geopolitical value. TSMC anchors a vast semiconductor supply chain in Taiwan that doesn't exist anywhere else in the world. Approximately 70% of all semiconductors pass through Taiwan at some point in their life. If a nuclear bomb were to hit Taiwan, the world's technology would revert something like 10 or 20 years into the past. It's hard to say that about any other individual geography or company.

This is particularly notable because Taiwan is a disputed territory. The U.S. and Taiwan think it is sovereign, while China thinks it is its own.

Although cutting-edge manufacturing occurs in Taiwan, it is virtually impossible to design and build cutting-edge chips without relying on US equipment and software. This puts China in a tight spot, and the US knows it. Semiconductors are a choke point in China's rise.

The U.S.exercised control when it put ZTE — China's Motorola — on the Entity List (a US trade blacklist). The US alleged that ZTE sold semiconductor equipment to Iran in violation of sanctions. Without access to US equipment, ZTE was crippled. It can't build anything anymore.

More recently, the US put Huawei on the Entity List. Huawei is a Chinese national champion akin to America's Cisco and Apple, combined. The US even banned TSMC from building anything for Huawei using American equipment. This was a big deal: Huawei was 20% of TSMC's sales, equal only to Apple. With a stroke of a pen, the US could completely stop China's technological rise.

China isn't taking this lying down. It's working as fast as possible to achieve semiconductor self-sufficiency. SMIC, a Chinese-based fab, has been in the works for over 20 years. They expect their 28nm process will be operational in three years. By then, that technology will be a decade old.

The only way any company has ever reached semiconductor manufacturing's cutting edge is through progressive improvement. No one has ever skipped a generation of manufacturing technology because the knowledge is cumulative. Further, semiconductors are virtually impossible to reverse engineer. Without US equipment, experts think SMIC is decades away from semiconductor parity.

America faces similar challenges re-shoring cutting-edge semiconductor manufacturing. It's not as simple as cutting a $12 billion check and buying some land in Arizona (as TSMC recently did). An entire supply chain exists in Taiwan that would need to migrate to the US. This is doable but takes money and time.

This week Europe decided to throw its hat into the ring too. Concerned about their dependency on the US and Asia for chips, 13 countries agreed to allocate EUR 145 billion to closing the technological gap (Reuters).

EUV Photolithography

To understand why replicating the current cutting edge technology is so difficult, it's worth understanding a bit about how semiconductors are made. They are, far and away, mankind's most complicated inventions.

To fab a chip, a company needs to shine light onto a silicon wafer to harden it. Then, they etch lines into the silicon to connect the transistors on it. Moore's Law calls for ever-increasing transistor density, which means these lines have to get closer and closer together.

TSMC currently leads the world with its 5nm process (1 nanometer is 1×10-9 meters). That's just 10-15 silicon atoms wide! Only Samsung and Intel are even close to TSMC's capabilities. Samsung has reached 7nm while Intel has reached 10nm. Over the summer, Intel surprised the market when they announced they're having trouble with their 7nm process.

A 5-7nm process is even more remarkable considering that the wavelength of light is 193 nanometers. How is this possible? EUV (extreme ultraviolet) photolithography.

There's only one company in the world that makes EUV photolithography machines. It's a Dutch company called ASML. Its machines shoot a piece of molten tin with a laser in a vacuum to create plasma. Only plasma radiates EUV light, which has a wavelength of just 13.5 nanometers. However, EUV is absorbed by virtually everything, including air and mirrors. That's why this process is done in a vacuum. ASML had to invent special mirrors that could reflect EUV. This is their secret sauce and the reason they have no competition.

ASML's process requires about 20 mirrors before the EUV is properly aimed and focused on the silicon. To generate enough EUV, ASML must shoot the molten tin 50,000 times per second to produce enough plasma.

For a while, there were two competitors in the photolithography market (ASML and Nikon). Several years ago, Nikon gave up on developing EUV. It was too hard, too costly, and too uncertain if they would succeed. Now ASML is the most valuable tech company in Europe.

Getting to its monopoly position wasn't easy. It took tens of billions of dollars. ASML funded this partially through an innovative co-investment program, which Intel, Samsung, and TSMC participated in. Its machines cost upwards of $200 million and require 4 fully-loaded 747s to transport them. They weigh enough to crush the foundations of most buildings.

This is just a taste of the technology used to produce the chips in your iPhones, laptop, and datacenter hosting this blog. To say fabbing is difficult is an understatement.

IDMs

Building state-of-the-art fabs is not merely a matter of throwing money at the problem. Intel spends more on R&D in a year than AMD earns in revenue, and yet AMD continues to take share.

That's one reason Intel trades for 11x earnings and AMD trades for over 100x. The market is much more enthused about AMD, NVIDIA, and fabless foundries working with TSMC than IDMs like Intel that are eating TSMC’s dust.

Intel argues that IDMs have the advantage of tightly coupling the design and manufacturing process to maximize product quality. This is true. But a problem arises when an IDM’s chip designs advance faster than their manufacturing ability. Either one can become a bottleneck and hold back growth. Knowledge compounds, and it's virtually impossible to skip a generation of manufacturing. So, once a company falls behind the leader, they're behind for good. No one has ever leapfrogged a leader.

This summer, Intel said that it is considering abandoning its 7nm fab and outsourcing to TSMC. This would be like Apple abandoning iOS for Android. The IDM model is in Intel's bones.

Switching from an IDM to a fabless foundry isn’t as easy as turning off the lights at the factory and emailing the design to TSMC. Chips need to be designed with the manufacturing process in mind, so Intel would need to go back to the drawing board and redesign its cutting edge chips for TSMC's process. This would take a few years and push Intel further behind. The semiconductor space is littered with dropouts - like Nikon abandoning EUV lithography and ceding the monopoly to ASM. If Intel drops out of cutting-edge fabs, it will only strengthen TSMC and Samsung's position. It would also be a major blow to the US's semiconductor self-sufficiency dream.

AMD has proven that it's possible to shift from an IDM to a fabless foundry model. It is one of the most remarkable US stocks of the last decade. In 2006 AMD levered up to buy ATI Technologies for $5.4 billion — 50% of its market cap at the time. Interest payments and lack of demand strained the company during the financial crisis. They raised capital, sold their foundry (GlobalFoundries), and took a $2.65 billion write-down on their ATI acquisition. A series of product misses and glitches caused operating losses to pile up between 2012 and 2016, as shown below.

In a last-ditch attempt, they decided to design a new architecture called Zen and have TSMC fab it. This Hail Mary hit, and the stock has been off to the races ever since. The stock bottomed out at $2 per share in 2016 and now trades north of $90 — a 45x return in 4 years (160% CAGR)!

Now AMD has the upper hand on Intel. AMD has what some regard as a superior architecture and superior manufacturing (by way of TSMC). The longer Intel struggles with its 7nm process, the further it risks falling permanently behind. The company is due to provide investors an update on the situation in January. If you're interested in AMD's full story, read this.

If Intel decides to outsource, its stock is likely to sell off. However, it’s conceivable that outsourcing will maximize future free cash flows. Building fabs requires billions of capex, which Intel wouldn’t need to spend. It’s margins would decrease, to account for TSMC’s take, but a decreased capital intensity would more than offset that and lead to higher returns on capital.

Fabless Foundries

AMD and manufacturing woes aren't the full extent of Intel's troubles, however. The company also faces competition from ARM and NVIDIA. NVIDIA is actually in the process of buying ARM from SoftBank for $40 billion. NVIDIA will most likely add its own IP to ARM's licensing "menu," which will create an even more compelling value proposition. That said, acquisitions, especially large ones, have a low base rate of success.

NVIDIA specializes in GPUs (graphical processing units), while Intel and AMD specialize in CPUs (central processing units). GPUs were traditionally used for graphics and targeted at gamers. However, designers have recently discovered that they can increase power efficiency and computing power by attaching several GPUs to a CPU rather than adding more CPUs—the CPU farms out specialty tasks to the GPUs. This combination makes GPUs great for parallel computing and ideally suited for cloud data centers.

Like companies increasingly specializing in designing or manufacturing, chips are increasingly designed for specific applications. Apple’s new M1 chip is a great example.

Apple licensed IP from ARM, designed the M1 in house, and used TSMC for production. By designing in-house, they tailored the chips precisely to their software to optimize performance and minimize power consumption. The results seem to have shocked even Apple's own engineers. Bottom-of-the-line M1 chips are blowing away the top of the line Intel Macs. For a deep-dive into how they work, read this.

This latest evolution is both an extreme version of specialization and a return towards vertical integration. Apple cherry-picked the best IP and best manufacturer and added value by designing chips specifically for its own software. Apple is in a unique position to do this because they sell the end product and write their own software. No one else is in this enviable position.

The cloud hyperscalers are also designing custom silicon. Their software is so advanced that hardware is their chokepoint. Few companies have the scale to even contemplate this, which is a competitive advantage for the "haves" and could be a permanent problem for the "have nots." Of course, none of the hyperscalers would be considering custom silicon if not for TSMC and the fabless foundry model.

Dell and HP look like potential losers here. They're dependent on off-the-shelf hardware from Intel and AMD. And they don't write their own software, so they can't compete with Apple by optimizing their chips for their software. Intel and AMD will undoubtedly produce more specialized chips for them with specific end-products in mind. Still, they will never be able to approach Apple's intimate relationship between its software and hardware.

That said, I wouldn't count any of these players out, least of all Intel. Intel is still very near the cutting edge of fabbing. The US will want them to succeed to preserve our technological independence. The company still generates tens of billions of annual free cash flow and has over 100,000 employees working on getting back on track. Money isn't a guarantee of success in this industry, but it does help.

While Intel's valuation is attractive at just about 10x, it's too hard to predict what its business will look like in five years, let alone 10 or 20. The same goes for AMD and NVIDIA. Businesses that depend on constant innovation are bound to stumble occasionally. AMD, TSMC, AMSL, NVIDIA, and ARM are priced like they will remain perpetual winners. Yet, AMD, Intel, and Nikon show that nothing is forever.

Analog

So far, I've been discussing digital semiconductors, like the kind you're using to read this. There's an entire other class of semiconductors called analog. Digital semiconductors only work with 1s and 0s. Digital signals are their input and output.

Analog chips interface with the real world. They take input from something like a pressure sensor or actuator, or thermometer and convert it into a digital output that a digital semiconductor can handle.

Digital chips cost hundreds to thousands of dollars. Analog chips cost less than a buck. Texas Instruments, the world's largest analog chip manufacturer, has an average price of 30 cents across its portfolio of tens of thousands of products.

Analog chips are being embedded in everything these days: cars, smart homes, industrial automation, and anything else related to the internet of things. 5G and AI will only accelerate the per-capita consumption of analog chips.

TI's business is like TransDigm's. They make mission-critical parts with a low cost to value ratio. TI’s chips are deeply embedded in its customer’s products, which produces high switching costs. When a company like John Deere designs a new tractor, they contract with someone like TI to design chips. The design process can take up to three years. Once done, the chip gets embedded into the product and will produce a recurring revenue stream for that product's life, which could be 7-10 or more years.

TI has a 20% share of the global analog chip market. It's next-largest competitors, Analog Devices and Maxim Integrated Products, are currently merging. Together, these three still only have about a third of the market. Both companies will grow by taking share from the remaining 70% held by sub-scale competitors rather than by fighting each other, like AutoZone and O'Reilly.

As in digital chip manufacturing, scale is critical. Like Intel, TI uses the IDM model. Unlike Intel, it is the clear technological leader with its 300mm process. TI's large market share, and large share relative to any single competitor, gives it the revenue base to justify a state-of-the-art fab. It also helps that TI has been at this game since day one.

TI's scale advantage and technological advantage give it a cost advantage over rivals. The company boasts 70% gross margins and 40% free cash flow margins. These are among the best in the S&P 500. The company is now nearing the end of a multi-year push to internalize distribution. Virtually all of TI's 40,000 products are now available on TI.com. By vertically integrating distribution, TI will grow even closer to customers, helping manage the demand cycle. They'll also capture even more margin.

Because analog chips are so cheap, no one will win a part from TI by bidding a penny less. Customers value customer service, quality, and availability above all else. In light of this, TI's move to self-distribute makes sense.

As you can probably tell, TI is my favorite business of semiconductor designers and manufacturers. It is not subject to nearly as much technological change as the companies in the digital space are. It has a recurring element to sales, which makes them more predictable than Intel or NVIDIA. And it has a cost advantage that appears relatively durable.

To boot, TI is also run by efficient capital allocators. The company provides investors with an annual capital allocation update presentation, a shockingly rare practice. TI aims to distribute 100% of free cash flow to investors through a 50/50 mix of dividends and buybacks. They're also one of the few companies that repurchase shares countercyclically. In 2009 TI seized upon low prices by buying a couple wafer fabs and used equipment from Qimonda. In 2011 they bought National Semiconductor, cementing their analog leadership. Since then, acquisitions have been scarce because prices are higher.

TI’s managers are excellent strategic thinkers. In 2009 20% of TI's revenue was from digital chips for smartphones. Management saw that the business was getting competitive, and margins were shrinking. They tried to sell the business but to no avail. So, they decided to simply wind it down. They stopped investing in it and allowed the business to run off. Instead, they redoubled their efforts on analog chips for automotive and industrial applications. This was an unsexy market but one that had clear tailwinds. Shareholders have benefitted immensely from that contrarian insight.

In 2010 TI traded for just 10x earnings because analysts doubted whether the company would replace the revenue it was leaving on the table. The market was also broadly cheap in the shadow of the financial crisis. Today TI trades for 25x earnings. This price fully recognizes TI's competitive advantages. While this price doesn't seem totally silly, it also doesn't offer much margin of safety. It's hard to foresee today's valuation could allow for meaningful outperformance.

Picks and Shovels

The sure way to get rich in the California gold rush wasn't prospecting. It was selling picks and shovels to the prospectors. In a similar vein, some of the most attractive and predictable semiconductor businesses might be the suppliers.

ASML, which I've already discussed, is a good example. They currently have a monopoly. It's hard to say how long that will last. My guess is that so long as EUV lithography is the cutting edge, ASML will dominate it. But I have no idea how long it will be before EUV is replaced by something better.

Another interesting niche is Electronic Design Automation (EDA). Without EDA software, engineers can't design chips. There are only two companies that offer cutting-edge (<22nm) tools: Synopsys and Cadence Designs. Both firms are American, which gives the US government leverage over China.

These companies have engineers on-premises at customers' offices working to tailor the software to their needs and the latest technological advancement. However, both firms use interoperable data (GDSII files) that allow customers to switch between them. Almost all semiconductor firms use both software. They're not locked into one or the other. This limits switching costs and pricing power, but has worked out as retention at both is about 100%.

Scuttleblurb sums of their competitive position well when he wrote:

EDA is a duopoly with huge entry barriers, but it is also a very competitive one with powerful customers who set the terms of product development and can shift their spending among interoperable tools. This limits differentiation and pricing power, even as it keeps Synopsys and Cadence running on the innovation treadmill. Synopsys and Cadence consistently devote 30% of their revenue to R&D and spend another ~10%-15% on acquisitions to maintain their leadership positions.

Both of these companies trade at 35x+ earnings, so there's likely quite a bit of time to learn more about them before they're near an investable multiple.

There are plenty more pick and shovel companies to explore, including equipment manufacturers like LAM Research, KLA Tencor, Applied Materials, and Tokyo Electron. There are also several testing and packaging companies that dominate niches. I'm still exploring these and will discuss them another day.

Broad Themes

When I think about the semiconductor industry, several broad themes emerge.

Consolidation

The semiconductor industry has been consolidating for decades. Today, many parts of the value chain have just one or two cutting-edge competitors. When NVIDIA started, there were about 100 competitors in the GPU space. Now, there are two. Companies have consolidated through M&A and attrition. M&A has been in the pursuit of scale (e.g., the Analog Devices-Maxim merger), which has clear advantages in a high-fixed-cost industry. Attrition resulted from the frantic pace of technological change Moore's Law demands. This was why Nikon stopped working on EUV lithography, AMD dumped GlobalFoundries, and could be why Intel quits pursuing 7nm process manufacturing.

If M&A and attrition lead to monopolies, investors will likely prosper. If TSMC and ASML are the only companies left in their fields, their incumbent advantage may mean that their monopolies endure. However, given the monetary and geopolitical stakes, I wouldn't necessarily bet on that at today’s prices. Incentives that big are immensely powerful.

2. Specialization

As technology advances, companies can only stay on the cutting edge if they focus and specialize in one area. This gave rise to the fabless foundry model and explains why Intel is struggling as an IDM.

Companies aren't the only ones specializing. Products are becoming specialized too. Apple's M1 chips may set a new standard for consumer products. Just as Apple looked to custom silicon as a means to improve its software, the cloud hyperscalers are also designing their own chips. Increasingly, Microsoft and Google may compete head-on with Intel, AMD, and NVIDIA.

3. Cooperation

Increasing specialization has made cooperation critical. Chip designers need to work with manufacturers to ensure what they dream up can actually be built. And the engineers behind the software have to understand where their customers want to take their designs.

Cooperation can embed companies into others, create switching costs and act as a barrier to entry. Like consolidation, this trend may lead to interesting investment opportunities down the road.

As the capital required for next-generation techniques escalates, we'll probably see more co-investments like ASML did with TSMC, Samsung, and Intel. No single company can afford to invest, and all benefit from it, so it's best to cooperate and pool resources. This creates a strong barrier to entry that protects the incumbents.

Semiconductors are mission-critical to the modern way of life. There's been astounding innovation in the past decades, which has winnowed the number of players in each niche down to just a handful. This has allowed the survivors to earn attractive margins and explains why they are among the world's most valuable companies.

The pace of innovation appears as fast as ever which makes picking the future winners and losers hard. How many predicted AMD's Hail Mary would hit in 2016? AMD at $2 only looks like a once-in-a-lifetime bargain in retrospect. I’m not aware of any investors who’ve succeeded at consistently predicting technological change. It’s certainly not my game. Fortunately, consolidation through M&A and attrition could create structural winners (and losers), which might be investable at the right price.

Disclosure: The author, Eagle Point Capital, or their affiliates may own the securities discussed. This blog is for informational purposes only. Nothing should be construed as investment advice. Please read our Terms and Conditions for further details.

Subscribe | Contact Us