We’ve been long Nvidia (NVDA) stock since it was trading at $90/share. However, being long doesn’t prevent us from raising questions regarding Nvidia’s long-term standing in the autonomous driving segment. In this research, we will explain why we believe that Nvidia’s long-term position in autonomous driving may not be as guaranteed as it seems. Actually, there are many questions that need to be answered before arriving at any conclusion regarding Nvidia’s long-term AD advantage.
At this point, with its Drive PX2 platform, Nvidia is the leader in the AD industry. Before explaining our thesis, it would be better to explain what Nvidia’s platform exactly does.
The Drive PX platform connects to it the steering wheel, multiple cameras, a radar/LIDAR( if any), and an SSD for data storage. By serving as a motherboard, the software in the platform connects what the vehicle sees (through cameras or radar) to how the driver reacted (turning the steering wheel right or left). It then compares the driver’s reaction to what the software would have predicted if there was no driver. The difference between the predicted reaction by the software and the real reaction by the driver is the “error”. This error is then used to correct the weights of the software so that it would be able to predict with higher accuracy the correct reaction in the future. This process is called “training” and it can be only done by GPUs, which are connected to the Drive PX platform.
After having trillions of megabytes of data that have the correct weights, the vehicle can drive completely autonomous. However, there is a major concern; it takes a lot of time for the vehicle to search for all the operations it did during training to get the real output. Think of that as a student studying for a math exam by studying 100 books. It takes a lot of time for the student to remember the needed equation for the problem they are solving during the exam. What’s the best thing to do? Make a list while studying, writing the equations expected to be needed during the exam.
Well, this is called “inference”. While training, the autonomous vehicle takes notes that include what part of the data it expects to be needed when facing a similar situation.
All GPUs have the capability to do both: training and inference. While Nvidia’s GPUs can do that at much higher speeds than other GPUs, Nvidia has no edge in inference as it focuses more on the training side.
At this point, Nvidia is partnering with multiple auto manufacturers such as Volvo, Volkswagen, Toyota (TM), Tesla (TSLA), Audi, and Daimler to train their vehicles. However, these partnerships are not exclusive and not long-term. These manufacturers need Nvidia’s Drive PX in their early days to collect and train data (training requires high-speed processing). But once they have the needed trained data, the importance of having a PX platform on every vehicle diminishes. While these vehicles will need a platform in case of an update on the traffic/roads, the speed of this platform won’t be crucial at this point. Once they have the scale, auto manufacturers won’t need every vehicle to train data as they would have millions of duplicated data that don’t give any benefit. At this point, auto manufacturers would be seeking a better cost/performance platform, which will be connected to the trained data in the data centers, and not for a faster platform. Here, Nvidia’s edge will diminish as there are many other companies that have these platforms but are not as fast as the Drive PX.
While these vehicles will need a platform in case of an update on the traffic/roads, the speed of this platform won’t be crucial at this point. Once they have the scale, auto manufacturers don’t need every vehicle to train data as they would have millions of duplicated data that don’t give any benefit. At this point, auto manufacturers would be seeking for a better cost/performance platform, which will be connected to the trained data in the data centers, and not for a faster platform. Also, they would be seeking platforms that have better “inference” capabilities than Nvidia’s GPUs as “inference” would be more important than “training” in the mass-production phase.
As a result, Nvidia’s edge will diminish as there are many other companies that have cheaper “training” platforms but are not as fast as the Drive PX, and Nvidia doesn’t have an edge in the “inference” side.
For instance, Xilinx (XLNX), a semiconductor company, is confident that it can beat Nvidia in the “inference” segment. Actually, Xilinx’s CEO shares our same view regarding the long-term importance of “inference” over “training”.
Also, NXP’s (NXPI) BlueBox (a platform like Drive PX but has a lower processing speed than Drive PX) is selling at $5,000 per unit at a retail price while Nvidia’s Drive PX2 is selling at $15,000 per unit at a retail price (auto manufacturers get a very high discount). And NXP has better connections with auto manufacturers as it has been selling them ADAS systems for decades.
Besides Xilinx and NXP, another threat can be from Alphabet (GOOG)(GOOGL) as the company’s inference machine, the Tensor Processing Unit (TPU), is used in data centers and can send commands from there to the vehicle using the gathered trained data.
Last but not least, Mobileye (MBLY) will be regarded as a significant threat to Nvidia in both “training” and “inference” as the company will transform its EyeQ chips next year from ones operating on a proprietary architecture to ones that have an open architecture. This played a significant role in Nvidia’s outperformance over the last few years as Mobileye’s EyeQ chips used to operate in a closed system, a not-welcome feature for manufacturers and developers. As a result, Mobileye may take market share from Nvidia in the mass-production phase as auto manufacturers will start caring about power consumption, an area where Mobileye and Nvidia compete closely.
Nvidia needs to be a leader in the mass-production phase of autonomous vehicles to get a decent revenue boost. By being a leader in the training phase, Nvidia is not getting any significant revenue. That’s why the automotive segment recorded just $140 million in revenues, barely higher than Q4-2016 and Q3-2016 numbers while having partnerships with many auto manufacturers. Most of the current revenue from the automotive segment is generated from sales to Tesla Inc. and not from sales to other manufacturers (this can be seen in the low sequential sales growth in the automotive segment). Also, there is a decent possibility that all other auto manufacturers won’t be involved in producing a sufficient number of vehicles carrying the Drive PX, which will affect Nvidia’s automotive revenues.
Nvidia’s Drive PX edge can be temporary and limited to a relatively small number of vehicles as the platform is mostly used for training data. Once the potential training data reaches its limit, Nvidia’s edge will be diminished as processing speed won’t remain a priority. Mobileye and NXP both offer platforms that have lower speed but at much lower cost. These alternatives can be more competitive once the training phase ends. Also, Alphabet’s TPUs are a threat for Nvidia’s Volta as future models of the former will carry high speeds for data centers, which might decrease the gap between the two.
We believe that at a $80 billion market cap, Nvidia needs to be a leader in the mass-production phase of autonomous driving and not only in the training phase as the former is where the real money is. Also, after pricing in most of the gaming and the data-center expected growth, the automotive segment will play a major role in determining the value of Nvidia’s shares.
Until we see real answers to the questions mentioned below, we prefer taking the sidelines.
In the meantime, we will be selling two-thirds of our Nvidia holdings as we see Nvidia’s upside is limited and dependent on a sector that is still unclear.
Now Read:“Why Nvidia’s Valuation Is Still Reasonable“
Disclosure: I am/we are long NVDA.
I wrote this article myself, and it expresses my own opinions. I am not receiving compensation for it (other than from Seeking Alpha). I have no business relationship with any company whose stock is mentioned in this article.