AI Networking Stole The Show In Q2 - Part 2
Astera Labs, Celestica & Arista Networks ER analysis
At The Pragmatic Optimist, we help hundreds of investors navigate the evolving AI innovation landscape, identify rock-solid businesses with strong growth trajectories and operational grit, and make long-term investments in the space with proven alpha generating returns.
Become a paid subscriber today
«The 2-minute version»
🏆💰TPO Subscribers Score Big Wins: Earlier this year, we published our deep dive on how the AI Networking & Connectivity space was poised to be a big beneficiary of burgeoning hyperscaler capex budgets and how a technology upgrade was needed in the networking infrastructure of AI data centers. Well guess what? Recent earnings reports from at least three networking vendors strongly indicate that the upgrade cycle in the data center networking infrastructure is robustly in progress, raking in millions of dollars for these vendors.
📈Scale up and scale out are the golden words every networking vendor is singing in earnings calls because those are the biggest billion-dollar opportunities every player is gunning for in the AI data center. Competition is fiercely intense, but so far TPO’s picks in the AI Networking space have delivered an average return of 70%, outperforming the S&P 500 and the Semiconductor Landscape by over 7x and 3x, respectively.
⏩Starting with Astera Labs, to say that business is booming would be an understatement. While its retimer devices continue to be rapidly used by hyperscalers in the data center, Astera Labs is carving out a new revenue base for itself with the recently launched custom silicon Scorpio switches, which are poised to double its market share this year.
🛜💪When it comes to Arista Networks and Celestica, both companies are benefiting from elevated demand in their networking switches, which are deployed in AI data center networks, but these companies are also beginning to see strong signals of inflection in the enterprise campus segment of their respective businesses.
Moving forward, the question is, how much upside is still left for these three companies, and what does the risk-reward look like?
You can find our updated price targets, conviction scores, and ratings for all three companies on the AI Stock Rec Tracker that all our premium members have access to.
🎥 Let’s Set The Stage
If there is anything that the first half of 2025 has made clear, it is that capital is flowing where conviction lies.
And right now conviction is sizzling with AI as the gold rush for chips and networking infrastructure underlines the need of the hour: data center capacity.
The Q2 earnings season started off on a strong note with hyperscalers seeing strong demand at their end for their cloud products and services. This strong demand in turn has stimulated hyperscalers to throw even more capex dollars at resolving capacity constraints they’re facing in the data center to meet the strong demand for their products.
Naturally, makers of accelerator chips (GPUs and XPUs) have and will continue to benefit from the burgeoning capex dollar budgets of their deep-pocketed hyperscaler & neocloud clients. But there is one segment of the data center infrastructure ecosystem that significantly surprised to the upside, deepening our conviction in these companies.
We started the year with a deep dive as to why we believed AI Networking and Connectivity companies would become the new cohort of capex beneficiaries.
We followed it up with bullish coverage on Celestica CLS 0.00%↑ in March, where we picked the company as our favorite AI Networking play for 2025. In May, we also covered Broadcom AVGO 0.00%↑, Marvell MRVL 0.00%↑ and Astera Labs ALAB 0.00%↑ and issued our first “buy” rating on Astera Labs at the time.
Since the beginning of the year, we have issued 7 total “buy” ratings on Celestica, Astera Labs, and Arista Networks ANET 0.00%↑ , delivering an average return of 70%, compared to 8.58% for the S&P 500 and 21% for the Semiconductor industry.
In this post, we break down the blitzkrieg of ERs from Celestica, Astera Labs, and Arista Networks that reported this week, along with updated price targets, conviction scores, and ratings that can be found on the AI Stock Rec Tracker.
Quick Primer on the AI Networking Market
Before delving into the ERs of these companies, we decided to quickly do a quick primer on the AI Networking segment which is a highly summarized version of the primer we published at the start of the year.
There are two kinds of networking architectures deployed in the data centers today: scale-up and scale-out networks.
The goal of scale-up networks is to increase the efficiency (high bandwidth, low latency) and density (the max number of GPUs/XPUs) of the network and is often deployed in a server node or pod or cluster where all the GPUs and resources (like memory and power) are connected together on a single fabric.
Scale-up networks usually rely on a combination of networking protocols such as Nvidia’s NVDA 0.00%↑ NVLink along with PCI Express and CXL, which is the de facto standard. Although UALink and SUE (Scale Up Ethernet) are also being increasingly used in place of NVLink.
Conversely, the goal of scale-out networks is to increase the number of nodes in an AIDC (AI data center) so that AI workloads can be efficiently distributed across server nodes. Scale-out networks are typically deployed using Ethernet only and interconnect all the server nodes together so each server node can communicate with the other server node. These kinds of scale-out networks are called back-end networks.
Back-end scale-out networks, along with scale-up networks, represent some of the most rapidly growing, high-margin revenue-generating parts of the AIDC at the moment. The networking companies that reported earnings in the past two weeks address significant market opportunities in scale-up and scale-out, as we noted in our March primer, which is where we had placed our highest conviction in the Semiconductor industry landscape.
Let’s begin our analysis of recent ERs with a tiny little company called Astera Labs vying to pick a fight with the two big bros of AI Networking.
Astera Labs Is Defiantly Taking On Marvell & Broadcom
Astera’s Q2 ER was a sight to behold. The company crushed expectations, reporting revenues of $192M, 2.5 times bigger than Q2 revenues last year, galloping past consensus revenue estimates of $173M. Gross margins came in at 75%, above management’s LT gross margin target of ~70%, while adj. operating margins expanded by an eye-popping 550 bp to 39.2%. This resulted in the company reporting GAAP earnings of $51.2M, or 29 cents per share. On an adjusted basis, Q2 EPS was 44 cents, also ahead of expectations of 32 cents a share.
The company’s Aries retimer devices, tiny little components that cost under a