Why We Invest in Silicon while Our Bet is on the Cloud¶
When YZ and I talked about this year, one theme we want to participate is the cloud. We've talked about cloud for many years since 2012. But only after 2015, it becomes abundantly clear that the cloud is here, resistance is futile. This is a multi-year theme that involves many stages. Experimentation (2012-2016), adoption and expansion (2016-2021), specialization and vendor lock-in (2021-2030) are the stages that each participant is going to play a different role in this show. In the midst of adoption and expansion stage, we believe the uncertainties of whether cloud will gain adoption in businesses are cleared. But the Wall Street right now is too eager to claim winners and losers while the market is still expanding, this could be presented as opportunities for us during the 2019 to 2021 timeframe.
Starting with the basics. There are 3 types of participants in the cloud business:
- Application service providers (SaaS) such as Workday, Salesforce, or smaller ones (Slack, Okta, Atlassian);
- Infrastructure providers such as AWS, Google Cloud, Microsoft, or smaller ones (DigitalOcean, Paperspace, Backblaze);
- Chips and network equipment providers such as Mellanox, Cisco, SuperMicro, AMD, Intel, NVIDIA.
Other players, such as Oracle or IBM, have businesses awkwardly positioned in both 1. or 2. or 3. and neither are dominate market players. Some, such as Microsoft, have significant presence in SaaS. These companies may fit into more than one type.
Application Service Providers (SaaS)¶
SaaS companies may have a different software delivery mechanism comparing with traditional application software companies. They also have a subscription-based revenue model. However, fundamentally, SaaS companies are competing in the software business. Their competitive advantages lie in how well they serve their customers (how competitive their pricing is, how feature complete to a given customer etc.). They dream to have vendor lock-in but legally they cannot (GDPR).
There are switching costs as any existing software companies have. There will be monopolies on specific technologies (thinking about Adobe). But that is nothing new. SaaS business is not interesting to us because while they may command a premium, the premium is not inherited from the cloud movement, but from other competitive advantages they possess. That is why no wonder Oracle, Microsoft, Intuit (and to certain degree, Adobe) continue to thrive with this new business model.
This is the part YZ and I were most fascinated with. The cloud infrastructure companies eventually will be like PG&E, they provides computing as PG&E provides electricity. While sounds dull, electricity is not a bad business! Most infamously, without regulation, Enron made tons of money on manipulating electricity pricing. As if this is not dramatic enough, all major players in this market have deep pocket.
For now, they are not going to command a premium on the infrastructure they provide. It needs to be a healthy business, yes. But AWS, Google and Microsoft can afford losing money for a very long time. (Probably not Amazon, they need to make their e-commerce balance first. But that is not challenging for someone like Bezos.) At the adoption and expansion stage we are in, they are competing with existing self-managed data centers and individual IT departments. It is not their focus to extract more money. Conquer and assimilation take precedence. They will provide cheap and ops-free infrastructure that are miles better than existing self-managed data centers.
As time goes by, how these cloud infrastructure providers operate will diverge greatly from how to run a self-managed data centers. For example, infrastructure providers will be more aggressive at evaluating the components they use. Unlike self-managed data centers where purchase decisions are driving by relationships, the purchase decisions from these infrastructure providers are more or less merit-based. They have much better statistics to evaluate cost of their components. If financially make sense, switch will be swift.
With a deep pocket, the infrastructure providers are not afraid to go down to the component level. They will invest in chips and hardwares that make sense to them. They are also eager to go up. Middleware services such as database management will be much more integrated with specific hardware. The cost of self-managed database on commodity hardware will stop making sense at that time. This is the specialization and vendor lock-in stage.
In specialization and vendor lock-in stage, infrastructure providers will command significant premium because the price collusion they had with each other. Depending on the political environment at that time, it can be extremely lucrative or morally-corrupt.
We want to participate the above infrastructure provider business. However, it is challenging. Big players (Amazon, Google, Microsoft, AliCloud) in public market traded as a package of complex business conglomerates. If we buy Amazon today, we are buying into cloud infrastructure, e-commerce (both retail and marketplace), entertainment (both original content and streaming), portable hardware (Kindle), even brick-and-mortar (Amazon 4-Star, Books). Same with Google, and to certain extent, Microsoft. YZ and I don't feel comfortable to make investment decisions with bunch of other risks we don't want.
Silicon, so here things become interesting. Let's take an example of Intel. During the adoption and expansion stage, cloud infrastructure providers will purchase more CPUs so that they have enough computation power to compete with self-managed data centers. In shorter timeframe, it drives un-nature growth on CPU manufacturers. But in longer timeframe, this is a displacement effect. Self-managed data centers will die, along with their demand for CPUs. The growth will fallback with the actual demand of computing power from the society. Maybe with the efficiency improvements the cloud infrastructure providers made, the growth will even dampen a bit.
Why we buy silicon if the future is as gloomy as I described?
Infrastructure providers differ from the self-managed data centers in many ways. First, the purchase decisions are driving by the rigorous cost model from statistic data. If Lisa Su was able to deliver a chip as powerful as Intel's, and as power-efficient as Intel's. If the price makes sense, infrastructure providers will deploy these at fast pace. Looking at AWS today, if we launch an instance, the only difference in the first page is on whether x86_64 or Arm_64. AMD or Intel are only differed in the second page with the little 'a' suffix of the instance type.
Unlike usual skeptics from self-managed data centers, if a software-hardware combination can deliver significant efficiency improvements, the cloud infrastructure providers will adopt quickly. That is the story of NVIDIA in the past three years, rising from a niche gaming chip provider to today's giant in machine learning computing.
Looking forward, in the next two years, we believe the dominance of NVIDIA in cloud machine learning, especially in machine learning training will continue. Inference is easy, and many existing chip makers or startups can make fix pipeline inference chips. For training, you need more flexible configuration of either the loss function, or innovative layers design. Because moving data between different computation devices are expensive, the programmable pipeline support in devices is a significant advantage to drive innovative neural network designs. Unfortunately, only few players have the experience of designing a compatible (to CUDA, OpenCL or OpenMP) programmable pipeline (Intel, AMD, NVIDIA). We fail to see how NVIDIA's lead in that area can be challenged in that short timeframe.
In the next 8 to 10 years, cloud infrastructure providers will start to replace their fleet with custom designed chips (around 30% to 50%). Like what AWS already did for their testing water ARM chips. These chips will start their life as massive parallel chips for web servers, or controller chips for massive storage system (See the recent Huawei’s Kunpeng 920 feature list). Slowly moving up to replace chips for database and other computation-intensive tasks. Looking at how Apple handles its mobile chip lead today, if these chips make sense financially, it will be done in a reasonable timeframe.
However, during that stage, even for deep pocket cloud infrastructure players, they cannot simply afford to (or financially make sense to) build their own fab. That leaves them to collaborate with TSMC, Samsung or Global Foundry for design and manufacture. This is a riskier bet because depending on political climate (and geopolitics), today's seemingly dominate fab could not be a viable choice for American companies. On AMD, we are optimistic that Zen 2 architecture will have the right ingredients to be a sensible alternative to cloud infrastructure providers. It seems they made all right decisions on moving to 7nm, having a separate IO die and double the floating-point vector width to 256-bit. It remains to be seen how it performs in real-world, and what's their launch date like.
NVDA sold off massively after Q3 earning and guidance disappointment. While we think in the near term (following quarter or two) the stock might remain weak due to channel and ASP pressure caused by crypto crash, the 3-year thesis remains intact - NVDA can and will continue to grow its data center revenue by at least mid 30 - mid 40 percent annually. Furthermore, with virtually no credible competition on this medium term horizon, we don’t see any problem for them to continue commanding the current gross margin around 70%. Finally, to keep the math simple (and conservative), we give no credit to crypto going forward and assume only modest growth rate for segments other than data center. Modeling it all out, our base case target price for Nvidia comes in at $230, or 70% upside. As for the bear case, the stock will find some support around $110, or 20% downside, if ASP and margin surprise to the downside in the next earnings announcement. That’s not to say this isn’t a real risk, however we feel the risk/reward ratio is really compelling for long term investors who can withstand short term volatility.
AMD was our best idea early last year, and despite it has gone up more than 80% since then, we remain constructive on the name. One conclusion from our NVDA discussion above is that AMD will not pose any real threat to NVDA in high end GPU or data center. However, we liked what we see on the CPU side. Powered by process advantage from TSMC and Zen 2 architecture, for the first time since 2006 it’s not inconceivable that AMD will become a meaningful second player in server CPU market and can take up to 20% of the market share. Given total addressable market is about 20bn, our base case is that AMD will gain 10% market share over the next 12 to 18 months. Assuming 30% net margin and net-net zero growth rate for everything else, this gets you to a target price of $25-$30. But we do see greater downside risk for AMD if either EPYC’s ramp disappoints or current crypto overhang deteriorates. The stock tends to find some support around $15, which according to our calculation corresponds to a low single digit server market share. The best strategy with AMD in our opinion is to be patient and opportunistically accumulate shares when risk/reward ratio makes sense.
In the long run, TSMC will not only benefit from the partnership formed with AMD on 7nm chips but more (we will elaborate in a separate thesis). Though based on our estimation, in near term the contribution from AMD’s 7nm products to TSMC’s top line growth will only be marginal. At 4% yield and with proven pricing power, enviable operating efficiency and best in class management, we continue to view TSMC as a steady compounder even though current valuation is not particularly cheap. Moreover, we are impressed by the accumulative know-how of running a high quantity production state-of-the-art fab and the market seems to under appreciate the process advantage TSMC possesses over Intel right now. Granted, at current stage, it is still something too early to quantify.