Arista Networks ($ANET) Deep Dive
A high-quality compounder, now positioned to outperform
Arista Networks ($ANET) has always been one of my favorite companies in the AI buildout.
And it isn’t hard to see why:
Revenue has grown at a +30% CAGR since 2014, while net margins have expanded from 10% to an incredible 39-40% today.
A perfect playbook for a high-quality compounder, exactly the type of business we like.
The problem is that the stock has almost always traded at 35-45x forward P/E, well above the rest of the sector, so the valuation was never exactly inviting.
The point is that, after the recent rally in semiconductors - and given that Arista is exposed to many of the same underlying dynamics - the stock has been left behind. For investors who have followed the company for a long time, that creates an interesting opportunity…
So, is it finally time to buy Arista Networks ($ANET)?
That is what we will cover in today’s deep dive.
We will walk through how the industry actually works, where Arista came from, how it really makes money, what it sells and why customers stay, its competitive position, financial profile, long-term outlook, and finally, the valuation (with target price for YE2027) - including what would need to change for the stock to outperform from here.
This is the longest and most in-depth piece we have ever published at Jimmy’s Journal, so here is a table of contents to help you navigate it:
Industry Overview
Company History
Business Model
Products and the Value Proposition
Unit Economics
Competitive Advantages
Competitive Landscape
Microsoft, Meta, and the Rack-Scale Pivot
Supply Constraints
Corporate Governance and Founder DNA
M&A and Capital Allocation
What Is Needed to Outperform?
Financials and Long-Term Targets
Valuation
Investment Thesis
Main Risks
Are We Buying $ANET?
1. Industry Overview:
For most of computing history, the network was plumbing. It moved data between machines, and as long as it was fast enough and rarely failed, nobody in a boardroom thought about it.
AI broke that assumption.
If before, networking was mostly about connecting a computer to a few external cloud resources, now, it is about hundreds or thousands of accelerators, the GPUs, TPUs, and custom XPUs that do the matrix math - all working on pieces of the same problem and constantly exchanging intermediate results.
The arithmetic is sparse and enormous, and the chips spend a meaningful share of their time waiting for each other to finish and synchronize. When that exchange stalls, expensive silicon sits idle.
At the scale of a frontier training cluster, even small inefficiencies in how data moves translate into millions of dollars of wasted compute and weeks of additional time to finish a model.
The network stopped being plumbing and became the thing that determines whether the most expensive hardware in the building actually earns its keep.
In simple terms, networking is the speed and efficiency of communication between chips.
In the context of data centers, these are the three metrics that matter most - and the ones companies in the industry are constantly trying to improve:
Bandwidth: how much data can move at the same time.
Latency: how long it takes for data to travel from one chip to another.
Synchronization: how efficiently all the chips can work together without waiting on each other.
The industry has organized this challenge into three layers:
Scale-up is the connection between accelerators inside a single rack or a single tightly coupled system. This is the densest, lowest-latency, highest-bandwidth domain, and it has historically been owned by proprietary technology, most visibly Nvidia’s NVLink, which ties its GPUs together inside a rack.
This is the domain Arista has not yet meaningfully addressed, and it is the one with the most strategic tension, because whoever owns the rack interconnect owns a large share of the architecture decision.
Scale-out connects accelerators across many racks to form the full training or inference cluster, anywhere from thousands to a million-plus chips. This is the heart of the back-end AI network, and it is where the most consequential shift in networking is happening: the migration away from InfiniBand, the proprietary interconnect long sold alongside Nvidia GPUs, toward Ethernet.
The Ultra Ethernet Consortium, of which Arista is a founding member, exists precisely to make Ethernet competitive with InfiniBand on the metrics that matter for AI, including lossless behavior and congestion handling.
Scale-across connects clusters that are physically separated, sometimes by long distances, because power and space constraints mean a single building can no longer hold everything a hyperscaler wants to deploy.
This involves routing data over distance while managing packet loss, latency, and on-the-wire security. It is a domain where Arista’s deep-buffer routing heritage gives it a genuine head start.
The switches that move this traffic are built around merchant silicon, switching chips designed and sold by a handful of vendors, with Broadcom the dominant supplier.
Broadcom's Tomahawk family handles the high-radix, low-latency switching that AI fabrics demand, and its Jericho family handles the deep-buffer routing used for scale-across and the broader data center interconnect.
The generational cadence of this silicon, from 400G to 800G to 1.6T ports, sets the rhythm of the entire industry. When Broadcom ships a new generation, the companies that turn that chip into a working, deployable, software-managed system fastest are the ones that win the design.
This is exactly the layer where Arista (and its peers) plays: the networking fabric that ties the entire AI cluster together.
The data center Ethernet switch market is forecast to grow at +27% between 2026-2030E, with the 1.6T segment scaling from under $4.0B in 2026 to more than $40B by 2030 at an +80% growth rate, while the older 800G segment persists as a roughly $16B market. The scale-up networking market alone, the domain Arista has not yet captured, is expected to grow from about $17B in 2026 to nearly $80B by 2030.
Found this content valuable? Share it with your network! Help others discover these insights by sharing the newsletter. Your support makes all the difference!
2. Company History:
Arista was founded in October 2004, originally as Arastra, by three people whose names carry weight in Silicon Valley: Andy Bechtolsheim, David Cheriton, and Ken Duda.
Bechtolsheim was employee number one and chief hardware designer at Sun Microsystems, and together with Cheriton, a Stanford computer science professor, he had earlier built Granite Systems, sold to Cisco in 1996, and famously written one of the first checks to Google in 1998.
The two funded Arista largely out of their own pockets, nearly $100M, bypassing venture capital entirely. Duda, the 3rd founder and the first employee of Granite, supplied the systems software that became the heart of the company.
The founders’ thesis was highly contrarian at the time:
Incumbent networking equipment ran on monolithic, fragile operating systems, where a fault in one part of the code could bring down the entire box, and where hardware and software were tightly bundled together, allowing vendors to lock customers in.
Arista's bet was to decouple the two:
It would build switches on standard merchant silicon, the same chips available to anyone, and
Differentiate entirely through software: a Linux-based, modular, programmable operating system called EOS, the Extensible Operating System, where every networking function ran as an independent process that could fail, restart, and be patched without bringing down the rest of the system.
The first products shipped in 2008, the same year the company rebranded to Arista Networks and brought in Jayshree Ullal as CEO. Ullal had run the data centers and switching business at Cisco, and her arrival converted a brilliant engineering shop into a commercial machine.
Here is an episode of Nicolai Tangen’s In Good Company podcast with Jayshree Ullal that is worth listening to after finishing the article:
The early customers were the people who cared most about microseconds and uptime: high-frequency trading firms, supercomputing sites, and the first wave of web-scale operators.
Arista won them by delivering consistent low latency and a software experience that made networks programmable rather than hand-configured.
The real inflection came with the cloud…
As the hyperscalers built out their data centers in the 2010s, they needed exactly what Arista had built, switching that could scale horizontally, be automated like software, and stream real-time telemetry.
Arista rode the 100G transition to a leadership position in high-speed switching and, by 2020, held the #1 share in 100-, 200-, and 400G ports. The IPO came in June 2014 at $226M raised, with the stock jumping nearly +30% on its first day. Arista entered the S&P 500 in 2018.
Management today frames the company's evolution in two acts:
Arista 1.0 was the data center switching insurgent that displaced incumbents at the cloud titans.
Arista 2.0 is the broader, AI-era platform spanning what the company calls its Centers of Data: AI Centers, Data Centers, Campus Centers, and WAN Centers, all unified by one operating system and one management plane.
Already a Pro subscriber? Feel free to skip this section and jump straight to the full report below.
This article is exclusively available to paid subscribers of Jimmy’s Journal.
Inside the full report, we break down:
Why Arista’s software-driven model earns SaaS-like economics inside a hardware revenue base - and why the screen P/E multiple overstates how expensive the stock actually is
The product portfolio behind the moat and the switching costs that keep Microsoft and Meta locked in
How our DCF, target price, and competitive positioning come together in the debate over whether Arista can defend its leadership vs. peers
This is the type of institutional-quality research usually reserved for professional investors.
↳ Upgrade to Jimmys Journal Pro using the link below to unlock the full investment thesis, just $297/year (about $24/mo):
3. Business Model:
Arista sells… boxes?
On the surface, Arista sells boxes. Switches and routers.
Those products are manufactured by partners such as Jabil, Sanmina, and Foxconn in places like Malaysia, Vietnam, Mexico, and other countries, then shipped through fulfillment centers to customers.
But that isn’t the best way to understand the company…