2023-11-27 16:21:43 ET
Summary
- Arista Networks' stock is currently trading at an all-time high, and its prospects for beating its financial targets are limited.
- The reason is that more demand for generative AI-specific versus general-purpose networking could impact sales in the short term.
- Arista may also face pricing pressures and requests for discounts from customers, as the supply chain normalizes.
- Therefore, there are uncertainties.
- On the other hand, this remains a highly profitable company and is likely to be a longer-term beneficiary of Generative AI, which makes it a hold.
Already trading at an all-time high of $218.39 and at 12x to sales which is 350% above the median for the IT sector, Arista Networks ( ANET ) is richly valued. Still, after Nvidia's (NASDAQ: NVDA ) record high orders for its AI-enabling GPUs, one may think that it is now the turn of networking companies to benefit, as the focus shifts to installation within the data centers of CSPs (cloud service providers) or enterprise customers.
The Price Performance (www.seekingalpha.com)
However, as this thesis will elaborate upon, this is not the case. Also, in a field where it is easy to be confused, it makes sense to differentiate between “general purpose” and “Generative AI-specific” networking. Thus, by precisely framing the potential benefits that Arista can derive out of this new flavor of artificial intelligence as well as factoring in the normalization of the supply chain, the aim is to highlight the limited prospects for the stock to beat its topline or bottom line during the fourth quarter 2023 (Q4) financial results. This makes the stock more of a hold.
General Purpose Vs Generative AI-Specific
Computing networks within data centers are characterized by two parts, the front end for linking up to servers and making IT applications available to users, Then, there is the back end, which interconnects the servers and storage devices. One of the architectural changes to accommodate new AI workloads is customers (CSPs and enterprises) now reinforcing their back-end networks for data to be crunched and processed at much faster rates than before.
In other words, back ends that support AI-specific networking are becoming very data-intensive now including GPUs supplied by Nvidia, namely to support the large language models used by applications like OpenAI's ChatGPT. This application is in turn driven by Generative AI, whose training and inference (utilization by users) phases not only require faster servers but also a tremendous amount of network bandwidth since the data needs to be transferred over and over again.
However, as a cloud networking company, Arista’s growth has been driven mostly by the front end using general-purpose networking based on super fast 400G, which, in addition to transporting normal IT workloads also supports artificial intelligence. However, this is more traditional flavors of the technology like machine learning, not Generative AI.
On the other hand, GPUs are mostly being added to the back-end network. For this purpose, Nvidia neither sells individual GPUs nor racks full of them. Instead, it sells entire clusters of GPUs called the HGX supercomputing platforms which include their own networking, using NVLink over the short haul and InfiniBand for the long haul. Thus, the semiconductor giant's data center sales were up by 279% YoY for its fiscal third quarter of 2024 thanks not only to Compute (GPUs) increasing by 324% but also Networking (which includes InfiniBand) up by 152%.
Consequently, in the short term, one should be open to the possibility of less demand from CSPs to impact Arista's sales in case they reorient their capital spending towards NVLink and InfiniBand. In this respect, after Meta Platforms' ( META ), one of Arista's main customers reduced capex guidance, the networking company's shares dropped by around 5% on October 26.
Normalizing Supply Chain and Margins
However, the stock more than recouped its losses a few days later after its third-quarter 2023 (Q3) financial results revealed that revenues of $1.51 billion constituted a 29.1% YoY surge and beat estimates by $30 million. Noteworthily, this was not due to Generative AI-related demand, but a normalization in the supply chain.
Furthermore, during the earnings call, the management commented about a decline in visibility amid improved lead times thanks to improving supply chain conditions. In fact, the company is currently selling more general-purpose networking equipment to enterprises that previously suffered from extended lead times due to the supply crunch for components which earlier impacted Arista Networks' ability to ship products in time. Thus, the higher topline was driven primarily by orders that had not been fulfilled earlier.
In the same breath, the revenue mix shifting more towards enterprise compared to CSPs helps to partly explain why the GAAP gross margin of 62.4% in Q3 was 2.1% higher than for the third quarter of 2022. Another reason for the higher profitability was the company benefiting from lower component pricing as a result of coming back to a normalized supply chain environment.
Going forward, the non-GAAP gross margin of approximately 63% guided for Q4 is about the same as in Q3, and for 2024, the expectation is for margins to decrease (stabilize) as the revenue mix reverts back to the normal enterprise vs cloud mix. Thinking aloud, there are also competitors to contend with, one of them being Cisco ( CSCO ). As such, there may be pricing pressures, especially after customers become aware of Arista's recovering gross margins. Also, in a normalizing supply environment, customers, especially larger ones may be tempted to request discounts.
In these conditions, it appears difficult for the consensus EPS estimate of $1.70 for the fourth quarter which represents a substantial 20% YoY increase, to be beaten.
In consequence, it is better not to put money on the stock, a position which I further support using top-line growth expectations.
Growth Forecast and Valuations
According to the management, the top line should continue to benefit as a result of supply chain improvement together with higher enterprise contributions to satisfy "previously committed deployment plans for some time". Thus, the company has guided Q4's revenue estimates of $1.5 billion to $1.55 billion which are higher than the average analyst estimate of $1.47B by $55 million (mid-point).
Now, given that this already constitutes a high benchmark when compared to Q3 with no impetus likely to be obtained from Generative AI-specific sales, it is unlikely for the company to experience a revenue beat. On the contrary, any adverse news pertaining to customers refining their forecast product mix in favor of Generative AI-specific networking could impact the sales forecast for fiscal 2024, which is projected at 33% over this year.
Talking price action, since one reason for the stock to surge by over 25% in October was better-than-expected revenue guidance, just think of the volatility in case the management was to guide lower figures for 2024. Worse, any decrease in sales may manifest itself in unused inventory of finished goods or unused components. For this matter, Arista's inventory amounted to $2 billion at the end of September, higher than the same period in 2021 by 233%.
In these conditions, given that the stock is already overvalued , it does not constitute an opportunistic buy in my view.
ANET valuation metrics (seekingalpha.com)
On the other hand, given its superior profitability grade of A+ due to its ability to leverage software as part of its technology mix, Arista is more of a longer-term hold for those who own the stock or to be put on the watchlist of those contemplating an investment.
Longer term Opportunity
In this respect, analysts at Morgan Stanley ( MS ) believe Arista is the "best way to participate in the AI networking market", but this is subject to connectivity moving towards Ethernet (from InfiniBand). I agree with them but believe that this is more of a longer-term opportunity and to justify my stance, I use data from Arista's 17th Annual Needham Virtual Security, Networking, and Communications Conference which was held on November 14.
Thus, as per Arista's Chief Platform Officer and Vice President-Product Management, not all CSPs or enterprises are spending heavily on Nvidia’s products. Instead, some are closely watching the developments while others are prioritizing a multi-vendor strategy that calls for the presence of another GPU supplier, possibly Advanced Micro Devices (AMD). In this scenario, these companies are not likely to choose only Nvidia’s proprietary InfiniBand technology but instead, they may opt for the conventional Ethernet supplied by Arista and others to avoid vendors being locked-in.
This thought process is aligned with the evolution of computer networks, where customers generally choose technologies that are more open (instead of proprietary) while also opting for supplier diversification. In this connection, the cost factor favors Ethernet given that the price of a GPU cluster (including InfiniBand) is pricier than the cost of a networking interconnect.
To benefit from the Generative AI wave, Arista is currently trialing Ethernet for use in GPU clusters and plans to reach the proof of concept phase in 2024, with the production phase for Ethernet-based clusters scheduled only for 2025. This means that, unless the company accomplishes these milestones ahead of expectation, it is positioned more as a long-term beneficiary of Generative AI-specific networking. Still, it is important to mention that the company is a beneficiary of more traditional flavors of AI which have been the subject of interest in the wake of ChatGPT and which its networking products support.
In conclusion, with the emergence of Generative AI, the back-end network requirements in data centers are evolving given much higher volumes of data have to be processed and transferred. As a solution, Nvidia supplies the networking gear required in addition to the compute part. This not only does NOT create windfall gains for companies like Arista in my view, but on the contrary, can result in revenue shortfalls in case customers were to reduce Capex pertaining to front-end network infrastructures. Instead, the company is benefiting from a normalization in supply chains.
Finally, there are long-term opportunities as the deployment of Nvidia's GPUs reaches an advanced stage which would call for more front-end capacity as more users access the large language models being developed. In this respect, larger CSPs adopting a multi-vendor strategy should help as the company adapts its product to be Generative AI-ready.
For further details see:
Arista Networks: Benefiting From The Supply Chain, Not Generative AI