Since I covered Arista Networks (NYSE:ANET) in November 2023, the stock has surged by over 40%, while I had a neutral position mainly based on the normalization of supply chains which could result in pricing pressures. Also, as charted below, the stock has progressed by more than 79% during the last year to reach $303.14, outperforming the Invesco QQQ ETF (QQQ) by more than 35%.
At that time, I had not identified any Generative AI (Gen AI) related opportunities, but, things have changed recently and this thesis aims to provide an update on $750 million of potential sales. However, amid all the hype, it is important to differentiate between the artificial intelligence the company already uses for enhancing its products versus sales opportunities that can be derived from the networking needs of AI models.
First, analyzing gross margins and comparing them with a competitor, shows that after recovering from supply chain woes Arista now faces an uncertain demand outlook for a company that has more than doubled sales to $5.8 billion in its latest fiscal year 2023, from $2.3 billion only three years ago.
Rapidly Adapting to Take Advantage of Digital Transformation
This is a networking equipment and services company that primarily designs and builds core and edge infrastructure, and also develops the accompanying software required for managing the internet infrastructure of some of the largest CSPs (cloud service providers) and enterprises, mainly telcos. As shown by its superior revenue growth which is four times competitor Cisco (NASDAQ:CSCO) and exceeds the IT sector median by a whopping 687%, it has been benefiting from the underlying need to move data around from corporate databases to public clouds as part of the digital transformation secular trend.
To explain this success lies the way it has rapidly evolved from a manufacturer of routers and switches, into more of a platform-oriented company that boasts only one operating system, something which simplifies life for network admins faced with more complex IT infrastructures.
Furthermore, it provides a broad range of services around the security theme including threat detection and response, identity management, and observability. Another of its differentiators is that it was an early adopter of software-driven architectures whereby the switch is treated as a platform and services are gradually added instead of having to change the hardware every time. This approach resulted in higher gross margins, of 64.87%, exceeding the sector median by 27.5%.
Uncertainty as to Demand Outlook
However, historical data as per the blue chart below shows that gross margins fell below 60% in early 2023, caused by additional component costs due to supply chain disruptions. Margins eventually rose as the supply chain normalized.
The above chart also shows Cisco’s margins which have improved too thanks to its software and subscription strategies together with a higher revenue base on which to spread fixed costs. Noteworthily, the networking giant also started suffering from supply chain constraints in 2022 resulting in higher shipping costs as I explained in a previous thesis. Thus, its grey chart dipped before eventually recovering, even faster than Arista whose CFO stated during its fourth quarter 2023 (Q4) earnings call that margin progression was partly due to “better supply chain costs”.
Another reason for higher margins was a better enterprise mix or more business from telcos versus big CSPs or hyperscalers. This is explained by the company squeezing better pricing out of telcos compared to hyperscalers who buy larger volumes and benefit from relative product discounts. Therefore, the increase in gross margins is due more to a variation in customer dynamics than demand-led higher product pricing as is the case with Nvidia with its H100 GPUs.
Along the same lines, management mentioned that they are having to keep additional inventory as customers “refined their forecast product mix” which points to probable demand softness going into 2024. This may in turn be due to inventory digestion or customers first consuming what was hastily procured in a supply-constrained environment before placing orders. There is also “limited visibility” which reinforces the case for demand uncertainty with the revenue for 2024 expected at $6.5 billion or growing at 10%-12% YoY which is nearly a third of 2023.
By comparison, Cisco is also facing demand headwinds and is expected to suffer from a revenue decline for its next fiscal year, which may be the reason it has acquired Splunk (SPLK) as I recently elaborated upon. On the other hand, Arista is expected to continue growing organically due to its ability to build an ecosystem of products and services that use a single platform in turn facilitating deployment and operations. But the question is whether it will be able to adapt rapidly to the realities of interconnecting Generative AI workloads.
Differentiating between Older AI and Generative AI to Identify Potential Opportunities
In this respect, it is important to differentiate between traditional flavors of artificial intelligence like ML or machine learning which has been around for years, and Generative AI, which drives sophisticated language processing and image generation applications like ChatGPT. For this purpose, Arista’s solutions (picture below) make use of AIOps (Artificial Intelligence for IT Operations), and predictive analytics which are both driven by ML to extract insights out of data with the intent of automating and enhance certain platform functions.
However, despite its innovative and AI-enabled platforms for managing complex network traffic, a different approach is required to support the workload necessitated by Gen AI.
In such a scenario, in addition to the accelerated computing power provided by Nvidia’s GPUs, Gen AI’s software algorithms also need to access massive amounts of corporate data, the more of it the better. The challenge here is data does not always reside in one place as companies have their IT distributed with part of it in public clouds and the rest in legacy systems sitting somewhere in private data centers. To address such a challenge, the solution is fast interconnects to connect data wherever are located to GPUs.
Now, one of the quick ways to achieve this is by buying Nvidia’s proprietary InfiniBand switches which are packaged with its DGX H100 supercomputer, but the problem is that it is expensive at around $270K each, which puts it beyond the means of smaller players. For some larger players, InfiniBand may simply not conform to the standards used in their data centers.
In consequence, those wishing to opt for more open interconnect back ends can choose Ethernet-based AI as proposed by Arista, and as per its CEO during Q4’s earnings call last month, this could represent a $750 million pipeline. Looking across the industry, Cisco could obtain more, or $3 billion and this excludes additional opportunities as part of its partnership with Nvidia for rapidly and securely deploying AI infrastructures at scale.
No Opportunity at Present
However, these are only pipelines or opportunities, and these two companies have to work closely with Nvidia and potential customers to design and test AI networking clusters for Ethernet versus InfiniBand. As for production, both Cisco and Arista plan it for 2025.
In the meantime, Broadcom (NASDAQ:AVGO), which is also a networking switch manufacturer in addition to a semiconductor and software has already harvested $2.3 billion of AI-related gains during its last fiscal year and is projecting about $10 billion during the current one as per its CEO. Therefore it looks like it is benefiting from hyperscalers who are trying to figure out how to accommodate Nvidia’s GPUs into their data centers without purchasing its InfiniBand technology.
In contrast, Arista has seen the percentage of total revenues it collectively receives from hyperscalers Meta Platforms (NASDAQ:META) and Microsoft (NASDAQ:MSFT) fall from 42% in 2022 to 39% in 2023 while at the same time, these two companies have been spending more on Capex. This may suggest that they are reorienting capital allocation to networking equipment from companies other than Arista.
In these circumstances, unless one has a very long-term investment horizon, it is better not to invest in Arista solely because of AI-related revenues, which will not flow during fiscal year 2024. Also, the demand outlook remains uncertain.
Instead, this year is dedicated more to testing as it adapts its 7800R AI spine and 7060X AI leaf switches to work with Nvidia’s GPUs for the training AI models. Thinking aloud, with its ability to adapt rapidly, Arista could benefit. Also, given its history of consistently beating topline and bottomline estimates, the CEO may have displayed a cautious approach when talking about AI training networking opportunities earlier this month.
Still, the stock is not a buying opportunity given that, at a forward price to sales of 14.5x, is already overpriced relative to the sector median by over 400%, possibly driven by the AI hype.
This said the stock could still gain because of its high momentum score especially after the Federal Reserve seems to have turned to a more dovish stance during its March 19-20 meeting which implies that interest rates could be cut soon. Interestingly, the prospect of looser monetary policy which is generally good for equities has provided a catalyst for Arista’s stock to rise by nearly 15% since the Fed meeting.
Therefore, there could be volatility risks in case the Fed does not cut rates rapidly and also because of high expectations as to the AI opportunities, which explains my cautious position.
Finally, after most of the AI training networking opportunities have been captured by Nvidia’s InfiniBand switches in 2023 mostly as a result of bundling them with its H100 GPUs, 2024 will be a crucial year for Arista as it adapts to the specific requirements of Gen AI. This implies that any positive update as to Ethernet AI testing progress could result in the stock surging, while any advances made by competitors could induce volatility risks. Thus, it is better to wait.