Arista Networks makes the specialized switches and software that move data around inside large data centers and corporate offices. When a cloud company or AI provider needs thousands of computers to talk to each other at enormous speed, they buy Arista's Ethernet switches and the EOS operating system that runs on them. Arista sells the hardware, then earns recurring revenue from support contracts called PCS (post-contract support) that customers renew year after year. In 2025, product sales made up 84% of total revenue and services made up the remaining 16%. The three main customer groups are large cloud and AI companies (48% of revenue), enterprise businesses like banks and hospitals (32%), and smaller AI-focused providers (20%). The diagram below traces where the money goes.

Five years of financial data tell a consistent story: this business has been growing fast and generating a lot of real cash. Revenue climbed from $2.9 billion in 2021 to $9.0 billion in 2025. That is more than three times larger in just four years. The growth has not come at the cost of profitability.

Gross margin, which measures how much money is left after making and delivering the products, has stayed remarkably steady. It was 63.8% in 2021 and sat at 64.1% in 2025. Holding margins flat while tripling revenue is unusual. It means the company has not had to slash prices to win customers. Free cash flow, the actual cash the business produces after all its operating costs, tells the same story. It was $1.0 billion in 2021 and reached $4.4 billion in 2025. The company also carries no net debt. In fact, it holds $10.7 billion in cash and marketable securities on its balance sheet.

The reason margins have held up is that Arista uses a single operating system, EOS, across its entire product range. One software platform serving every product means engineering improvements apply everywhere at once. It also means customers who build their networks around EOS face real switching costs if they ever want to change vendors. That stickiness helps protect pricing.

Despite the strong numbers, there are specific risks documented in Arista's own filings that are worth understanding clearly.

Two customers together accounted for 42% of total revenue in 2025. One single customer represented 26% of revenue and another represented 16%. If either of those customers slows down its spending, delays orders, or shifts to a competitor, the revenue impact would be immediate and large. Arista's own filings describe the timing of these large orders as unpredictable. Orders can shift between quarters with little warning.

Arista relies primarily on one company, Broadcom, for the switching chips that go inside its products. There is no written agreement guaranteeing supply. If Broadcom changes its pricing, delays shipments, or stops supplying Arista, the company cannot easily find a replacement. New tariffs on goods made in Malaysia, Vietnam, and Mexico, where Arista's contract manufacturers operate, add another layer of cost risk. Arista's filing states that if these higher costs cannot be passed on to customers, gross margins would be squeezed.

There is also competitive pressure that is structural, not just cyclical. Nvidia sells its GPU chips bundled with its own proprietary NVLink interconnect networking. When a customer buys Nvidia GPUs, they get Nvidia's networking solution as part of the package. That is a direct challenge to Arista's position in AI data centers. InfiniBand, another proprietary networking technology, has historically dominated supercomputer clusters for the same reason: it comes bundled with hardware that AI researchers already want. Arista is betting that open Ethernet standards will win out over these proprietary alternatives, but that outcome is not yet settled.

The financial trajectory is strong across every measurable dimension. But the key question is whether the conditions that produced this trajectory will continue. The AI networking boom has pulled forward enormous amounts of capital spending by cloud and AI companies. The company's own filing acknowledges that customers may overestimate their needs and cancel orders, and that demand estimates for new AI products are difficult to forecast.