Xilinx – Investing in FPGAs For AI Hardware
The year was 2012 and Apple (NASDAQ:AAPL) had just become the world’s most valuable company reaching a market cap of over 600 billion. If exactly 5 years ago you bought shares in Apple, you’d be up about +103% on your investment vs a NASDAQ return of +109%. If in fact you decided to invest in NVIDIA (NASDAQ:NVDA) instead, you’d be up over +1,000% on that investment today. You’d be even happier this morning to see that Softbank (TYO:9984) has been quietly amassing a position in NVIDIA of $4 billion making them the 4th largest shareholder.
We recently wrote about NVIDIA’s market share of GPUs vs their biggest competitor, AMD (NASDAQ:AMD). A number of
nerds critics chastised us for being so blunt about our findings, but our conclusion was that AMD poses little threat to NVIDA dominating the market for GPUs as AI hardware. We then took a look at 12 startups building new AI chips and noted that while some of these startups looked exciting, they would have an incredibly difficult time displacing NVIDIA’s strong position.
If you’re an investor in NVIDIA at the moment, what you should probably worry about the most is something called a Field-Programmable Gate Array (FPGA). You see, both Intel (NASDAQ:INTC) and Microsoft (NASDAQ:MSFT) are betting that FPGAs will be the dominant AI hardware in the future. A paper recently released by Intel engineers titled “Can FPGAs Beat GPUs in Accelerating Next-Generation Deep Neural Networks?” provides some of the technical reasoning behind this speculation. As investors though, this made us ask a very basic question. Who builds FPGAs today? Here’s a breakdown of market share according to a recent article on EE Times by Paul Dillien:
We don’t care if the percentages are precise or not, but what we do care about is knowing who the players are. Of the companies listed above, let’s look at the ones that are publicly traded:
|Ticker||Market Cap||1-Year Ret.||5-Year Ret.|
At this point in time, what we could do is create a motif and invest in all of these stocks, or we could just focus on the market share leaders. Let’s start with Intel first.
While Microsoft is presently using FPGAs in their data centers, they don’t actually build the chips. Intel builds the FPGAs that Microsoft uses, but they didn’t always. In 2015, Intel announced that they would acquire Altera Corporation for $16.7 billion in cash, the largest acquisition that Intel has made to date. Just a few weeks ago, Intel put up an “Artificial Intelligence Investor Tour” with the below chart showing just how they plan to address AI hardware:
That’s all fine and dandy, but if we look at the contributions that the Altera acquisition had to revenues, it’s minimal at best. The below pie chart shows the introduction of the “Programmable Solutions Group” which represent the FPGAs from the Altera acquisition:
The revenues that can be attributed to FPGAs are just 3%. We cannot discount Intel’s capabilities when it comes to executing on an AI hardware strategy given their size and capabilities, however, this article is about FPGAs. Let’s move on to discussing Xilinx which is the market leader in FPGAs.
Xilinx may have an advantage considering the fact that they invented the FPGA. They were the first “fabless semiconductor manufacturer” which means they don’t have to worry about running a capital intensive business or worry about the risks that come with being a manufacturer. We can see that they’re quite profitable and have steady revenues over the past 5 years:
The breakdown of where these revenues are coming from can be seen below:
- Communications & Data Center – 42%
- Industrial, Aerospace & Defense – 41%
- Broadcast, Consumer & Automotive – 17%
Just a few days ago, Xilinx had an analyst meeting where they presented a deck which talks extensively about the plans they have for their business and of course, is peppered with references to AI and machine learning. Here’s where they see their growth coming from:
The above slide is a bit confusing to us and here’s why. The “data center” segment seen above has a very low forecasted growth rate (1-4%). These “data centers” are the infrastructure that powers cloud computing and these “data centers” are where all the growth is coming from for NVIDIA GPUs. Here’s an excerpt from an article on The Next Platform which talks about NVIDIA’s growth:
The star of the final quarter of the fiscal 2017 year was without question the Datacenter division, with revenues up by 205 percent to $296 million. That’s a more than tripling of this business over the fourth quarter of fiscal 2016, yielding a run rate of this business of $1.2 billion.
NVIDIA’s growth is clearly coming from deploying GPUs for data centers that are used for all the cloud-based artificial intelligence services being offered today. Microsoft chose not to use GPUs for their data centers and uses Intel’s FPGAs instead. So who has Xilinx lined up with? Only the biggest provider of cloud computing by a long shot, Amazon:
So Xilinx is getting into bed with Amazon and yet they’re not expecting tremendous growth as a result? Maybe it’s just a matter of semantics. If we look at the full analyst presentation, we see a common theme is talking about how Xilinx plans to target a large range of machine learning applications like these:
You can have a look at the deck yourself, but it seems that the expectations for growth are quite conservative considering the potential of the Amazon relationship. It’s good to see Xilinx has identified machine learning as a potential growth opportunity and they also appear to have a lead on the competition in regards to their technology:
Just to keep it simple, the smaller the number, the better the technology. Xilinx claims to have an 18-month lead on the competition when it comes to their FPGAs. This means that if retail investors are going to make a bet on FPGAs being the way forward for AI hardware, it seems like Xilinx might be the only viable pure-play. Here’s what Morgan Stanley analyst Joseph Moore had to say about this in December of 2016:
2016 was a year where machine learning ‘training’ became a major driver, primarily benefiting graphics leaders NVIDIA and AMD. In 2017, we should see the beginnings of specialized hardware for “inference” – using those graphics-trained neural networks in a real time environment. Somewhat to our surprise, our checks show that Xilinx is emerging as a leader in this segment. The announced win with Amazon Web Services a couple of weeks ago was largely ignored by the market, yet was clear evidence that Xilinx was able to win a socket competing head to head with Altera at one of Intel’s most important customers. It’s a small revenue driver in 2017, but one that opens up a window to a large potential opportunity.
While this same analyst downgraded NVIDIA back in 2013 which was about the worst call ever, he makes some good points above. Even if he’s wrong about Xilinx emerging as a leader in inference, Xilinx is not just a play on artificial intelligence, but also on other exciting emerging technology themes like virtual reality and autonomous driving.
NVIDIA isn't the only way to invest in AI Chips. Become a Nanalyze Premium annual member and we'll tell you about a company that sells their software to all AI chipmakers.