Discover more from The Asianometry Newsletter
TSMC and AWS: Comparing Two Infrastructure Giants
Author’s note: If you want to watch the video, here it is below. If you want an audio version, it is linked here.
I personally think this is a good one. The video did not get all that much play, but I found it a really fun exercise to compare these two massive infrastructure companies. One underpins the software world, and the other hardware world.
A lot of people got angry at me in the comments for calling TSMC “state-backed”. Like, legit angry. Why would people get mad over this? A big portion of the company’s shares are held by the Taiwan government. I would be more concerned if TSMC did not have the backing of the state, because Samsung and GlobalFoundries sure are. I have removed it here, but I am bemused to think of someone watching a 15 minute video and be moved to anger by a single phrase in the first ten seconds.
If you want to support the channel, I encourage you to take a look at the Patreon. Try out the Early Access tier, which grants you access to a number of unlisted videos that are waiting to be released. I highly recommend the video on Gogoro’s battery swap network build-out.
What do Amazon Web Services and TSMC have in common? One is the cloud computing division inside the biggest e-commerce company in the world. The other is a Taiwanese semiconductor giant.
While these companies might seem quite different, one thing that I notice unifies them is that they provide a platform that enables other companies to succeed and exist.
TSMC's focus on manufacturing and commitment not to compete with its customers has spawned a generation of "fabless" semiconductor companies. AWS has allowed Silicon Valley to build pure internet service companies without needing to build a data center of their own.
For this video, I want to look at and draw a few parallels between these two special companies. The plumbing providers of the modern internet.
I have a few other videos on this channel about Taiwan Semiconductor Manufacturing Company or TSMC. If you want to get the overview, you can check out the linked playlist. But I haven't introduced Amazon Web Services (AWS for short), so let’s take a few minutes.
AWS is the dominant cloud platform in the western technology industry. It is in the business of building out and renting to customers computing services on-demand. These services can range from online storage (AWS S3) to virtual computing (EC2). It has leading share in the industry with 33%, competing with Google and Microsoft, and made around $35 billion in revenue in 2019.
End users like you or me rarely use these services directly. Instead, AWS serves other businesses - making it a business-to-business company, or B2B. AWS customers can connect to AWS's services using internet protocols and use those services to build or host a consumer service for their customers.
The division is immensely profitable and a big driver for Amazon's share price. It contributes over half of the company's total profit. When Amazon broke out AWS's revenue and profit as an individual line item in 2015, it caused Amazon's market capitalization to rise $25 billion to $207 billion. Or in essence, a $25 billion IPO.
Amazon first began as a book store on the internet, selling wares that it had bought itself. It evolved from there to offering its services to small vendors looking to sell their own wares on the Amazon website.
It expanded on this idea of selling picks and shovels to miners rather than itself digging for gold - and expanded that so to enter into a variety of other industries. For example, logistics with the "Fulfilled By Amazon" program. And Internet services with Amazon Web Services.
In doing so, they expand their market and deeply embed themselves into the lives of both buyers and merchants. The buyers are trained to visit Amazon.com for whatever it is they need. Merchants know that they need to go to Amazon to find those customers. And so on.
TSMC's founding story is different. It had been a part of the Taiwan's government initiative to develop its semiconductor capabilities. Founder Morris Chang had worked in the American semi industry for several decades at Texas Instruments. After retiring, he was recruited to help the Taiwanese. He knew the complexity and cost of making semiconductors at scale.
He also knew that what Taiwan had at the time was not competitive with the market. Its most advanced process node, 3 micrometers, was two generations behind the leading edge. UMC a few years earlier had attempted to export its wares to the global market but could not get their chip products into anything more complicated than a few toys. The design was not good enough and the marketing was poor.
As Chang tells it, the pure-play foundry model was created out of necessity. Chang decided to focus on Taiwan's traditional manufacturing strengths and entirely forego design and marketing. They told their customers that they would not compete with them as a rule. TSMC leaned into its weakness and turned those deficiencies into a selling point.
The idea of renting computing power out to customers has been around for a while, but certain challenges needed to be overcome to strengthen the business model. A key one is that in order to help its customers serve their customers, infrastructure companies needed to offer a simple, easily understood method to "abstract away" the complexity of something really hard. Otherwise, the customers are not going to see the value.
For AWS, they offer the API, the interface from which the client can interact with AWS. It turns this complicated task that the customer needed to do into a simple black box with inputs and outputs.
To illustrate, let us look at the example of a web application that needs to know what objects are in a photo. Rather than building and maintaining this entire other separate system for this minor task, it interfaces with this AWS service Amazon Rekognition and lets them do it.
To do so, the application uses the API to send Rekognition the image and receive the results. There is no need for the application to know anything more than how to call on the interface and how to receive and interpret the results.
That the interface is simple and unambiguous is important. The generalized use of open web standards like XML and JSON to communicate with Amazon's databases helped AWS demonstrate its value to developers. The integration costs would probably have otherwise been too high.
TSMC likewise has had to do the same for its partners. That interface with TSMC's service needs to be as simple as it can be. So TSMC offers to its customers what is called a Process Design Kit or PDK. It is essentially an API for making semiconductors.
The PDK is a set of design files with a library of "blocks" that can be customized per rules listed in the attached documentation. It lets a designer layout a design and test it right in their computer using an emulator.
TSMC's process specialty, CMOS, fits the building blocks model well. With CMOS, much can be standardized while still having a great deal of customization and variation. This building blocks approach helps abstract away a lot of the complexity of making a 5 nanometer chip while at the same time helping TSMC standardize what it needs to fab.
Of course as with anything that boasts billions of transistors, the learning curve for developing a chip even with all of the tools TSMC offers is quite steep. And at cutting edge nodes like N5 and N3, the development costs can get prohibitively expensive. At N3, the design cost is expected to fall between $500 million and $1.5 billion.
Unlike with XML and other open source web standards, the PDK is proprietary and exposes many trade secrets. They are very crucial to the company's overall success.
Both AWS and TSMC focus on enabling their customers create products for their audience. Who are these customers?
For TSMC, these are semiconductor sellers without their own foundries or fabless semi firms. These were often medium sized companies with hundreds of people rather than thousands. They often make their living selling products into a specific niche, different from that made by Intel or other established market leaders.
Fabless firms had been around since the mid-1980s, long before TSMC's founding. A few examples are Altera, Xilinx, Nvidia, and Qualcomm. But they were small fish in a big pond. In order to get their product made, they had to partner with a large established semiconductor maker to do so.
These manufacturers extracted a high cost for their aid. First, they required the transfer of trade secrets and IP over to themselves. Second, they gave themselves the option to essentially make their own version of the product if it sold well. Thirdly, foundry capacity was not guaranteed - priority always went to the manufacturer and the fabless maker could be booted whenever required.
Nvidia for example, originally partnered with SGS Thomson Microelectronics to build their chips. When co-founder Jensen Huang heard about TSMC, he wrote Morris Chang letters to become a customer.
I previously did a video about MediaTek, another fabless semiconductor maker and the biggest in Taiwan. If you are interested in the way fabless semiconductors work, you can check that out in the description.
Both AWS and TSMC grew with their customers. TSMC clients like Nvidia and AMD are now themselves multi-billion dollar giants. AWS likewise provides critical services for mega-corporations like AirBnb, GE, and Netflix.
But the vast majority of AWS's million strong customers are small or medium sized businesses. Developer shops who appreciate the pay-as-you-go flexibility, the ease of starting something, and the vast product options. Much like the fabless semiconductor companies, they don't want to have to commit a lot of money up front.
Something that I think I should mention AWS that makes it different from TSMC is that there is no pledge that AWS won't eventually compete with its customers. As Bezos once said during a Congressional appearance:
> "I think there may be, uh, categories, uh, databases of different kinds where we see it's an important product for customers and we make our own product offering in that arena ... But it doesn't mean we stopped servicing the other companies making those products. We have competitors using AWS, and we work very hard to make them successful."
Amazon has a reputation for bringing out their own offerings of products made by their platform users. And using its own customers' data to compete with them. Perhaps it is for that reason that Wal-Mart does not want its vendors to use AWS.
Developers continue working with AWS despite this reputation. Perhaps because the cloud service offers the best pricing, or they have no feasible competing choice, or because they feel AWS won't compete with them for whatever reason. So the pledge is implicit, rather than explicit.
Lots of Capital
In a recent video, I talked about the scale of capital expenditure that TSMC is projected to spend in building out new facilities and capacity for its customers. For 2021, the company estimates that it would need to spend $25-28 billion.
$28 billion is a staggering amount of money, and that takes TSMC's capital expenditure spending into the same stratosphere as AWS and other tech giants. AWS is building data centers all over the world - they serve 24 geographic regions around the world. Their capital expenditure budget for 2021 is $32 billion.
These expenses include the cost of buying land, building on top of it, and installing it with expensive equipment. All of this requires people with experience to run and do this. And management to manage them. That also costs money.
Customers outsource these capital intensive activities to TSMC and AWS because it makes financial sense to do so. At low scale, operating a data center or a foundry is extremely unprofitable. Such facilities cost billions to create but are only useful for basically just a few years. The equipment inside gets old, breaks down and needs to be replaced. Like a new car, the second it drives off the sales lot it starts to lose value in depreciation.
To compensate for that depreciation, TSMC needs to be using that foundry at its maximum output. AWS needs to make sure that its servers are being used. Both make it economical for themselves by aggregating demand so that the facility is almost always running.
With the right accounting treatment, the business looks very profitable. You are turning a bunch of expenses into assets that can be stored on the balance sheet. Cash flow-wise, though, might not be the same story, depending on the nature of the depreciation.
Conclusion: Friendly Little Helpers?
AWS and TSMC together became fantastically successful as enablers of other people's success. Their focus on servicing other businesses, their abstracting away of the complexity of something inherently complex, and their aggregation of demand. They really are getting rich by selling pickaxes and shovels during a gold rush.
Things have developed such that companies without fabs and data centers are seen by investors as having a better business model than those with such expensive items. Look at the price multiple of Intel as opposed to AMD for example. Wal-Mart versus Amazon.
It is great when it works. But what happens when things break down? What happens when AWS boots you from their network? Or when the government bans you from using TSMC's technology? Or when suddenly Stripe closes your account with them? Then what happens? What can you do next?
Being a customer of an infrastructure B2B company is about living in a state of constant nervous tension. Believe me, I know. Knowing that your business depends on this other party for this very important service. And they are the best at it. You love it. 100%, you would not be where you are without it.
But you also know in the back of your mind that one day you can wake up to a really bad email and then it is all gone. Everything you worked for gets pulled out from under you. And that would really suck.