Arista Networks
| Arista Networks, Inc. | |
|---|---|
| Type | Public company |
| Traded as | NYSE: ANET |
| Industry | Computer networking |
| Founded | 2004 as Arastra; incorporated as Arista Networks in 2008 |
| Founder | Andy Bechtolsheim, David Cheriton, Kenneth Duda |
| Headquarters | Santa Clara, California, United States |
| Key people | Jayshree Ullal (Chair and Chief Executive Officer) |
| Products | Ethernet switches, routers, network operating system, network management software |
Arista Networks, Inc. is an American provider of client-to-cloud networking for large data centre, artificial intelligence, campus and routing environments. It is headquartered in Santa Clara, California, and is among the principal suppliers of high-speed Ethernet switching to the operators of the world's largest data centres.
The company was founded in 2004 as Arastra and incorporated under its present name in 2008. It listed on the New York Stock Exchange in June 2014 under the symbol ANET, entered the S&P 100 in 2026, and was recognised on the 2026 Fortune 500 list of the largest American corporations by revenue.
Arista's significance in the industry is disproportionate to its size relative to Cisco Systems, the incumbent it has consistently taken share from. It established that a network switch could be built from standard merchant silicon and differentiated by software, an argument that ran against the vertically integrated model the industry had followed for two decades, and it did so at exactly the moment when the growth of hyperscale data centres created customers who wanted precisely that.
History
Founding
Arista was founded by Andy Bechtolsheim, David Cheriton and Kenneth Duda. Bechtolsheim and Cheriton had previously founded Granite Systems, an Ethernet switching company acquired by Cisco Systems in 1996, and both had been among the first investors in Google. Duda had worked with them at Granite and subsequently at Cisco.
The company's founding thesis was that the economics of networking hardware had changed. Merchant silicon, meaning switching chips designed and sold by specialist semiconductor companies rather than developed in-house by the switch vendor, had become good enough that the hardware was no longer the difficult part. What remained difficult was the software, and the incumbent operating systems had been written in the 1990s for a world of much smaller networks and much less demanding availability requirements.
The Extensible Operating System
The product of that thesis is EOS, the Extensible Operating System, which is what the company is genuinely differentiated by.
EOS is built on an unmodified Linux kernel. Its design principle is a multi-process state-sharing architecture that separates state information entirely from the processes that operate on it, in contrast with older network operating systems built on monolithic kernels in which a fault anywhere can bring down the whole device.
Protocol processing, security functions and hardware drivers all run in user space as independent processes called agents. At the centre sits an in-memory database, SysDB, which holds system state and distributes changes through a publish and subscribe model. Agents never communicate with one another directly; they publish changes to the database, which notifies whichever agents have subscribed to that category of state.
Three consequences follow, and they are what customers buy.
Fault containment: because agents are isolated in user space, a defect or crash in one does not cause a system failure.
Stateful restart: an agent that restarts re-subscribes to the database and resumes from where it stopped rather than rebuilding its state, allowing recovery within tens of milliseconds and rapid network convergence.
Hitless upgrade: patches can be applied to an individual process in real time, and full operating system images can be replaced while packet forwarding continues, which removes the planned downtime that network upgrades have traditionally required.
A single software image runs across the entire product range, from campus switches to the largest data centre platforms, which reduces the operational burden on customers who would otherwise manage several different operating systems.
The architecture has since been extended beyond the individual device. NetDL, the Network Data Lake, centralises the state formerly held in each switch's SysDB across the whole network, allowing analysis, troubleshooting and automation to operate on a complete picture rather than device by device.
Litigation with Cisco
In December 2014, six months after Arista's listing, Cisco Systems filed two lawsuits alleging infringement of patents and of copyright in the command line interface used to configure its equipment.
The copyright claim was the more consequential in principle, since a network operating system's command syntax is what engineers learn, and a finding that it could be owned would have made switching vendors far harder for customers. The patent claims proceeded before the United States International Trade Commission, which issued limited exclusion and cease-and-desist orders concerning two features and upheld an import ban on infringing products, obliging Arista to redesign them.
The dispute was settled in August 2018, with Arista paying Cisco US$400 million and the parties agreeing a five-year standstill. Commentary at the time treated the litigation as having been at least partly competitive in purpose, since it was brought against the company taking share from Cisco most rapidly.
Acquisitions
Arista has expanded its range principally through acquisition of specialist companies.
It bought Mojo Networks in August 2018, adding cloud-managed wireless networking, and Metamako in September 2018, whose ultra-low-latency products, aimed at financial trading, became the 7130 series.
Big Switch Networks was acquired in February 2020, adding network monitoring and software-defined fabric capability, and Awake Security in October 2020, bringing network detection and response, which formed the basis of the company's security portfolio.
Pluribus Networks, a unified cloud networking company, was acquired in August 2022 for an undisclosed sum.
In July 2025 Arista expanded its artificial intelligence-driven campus and branch offering and acquired the VeloCloud software-defined wide area networking portfolio from Broadcom. The combination was presented as extending the company's automation and monitoring approach from the data centre to the branch office, with zero-touch provisioning, proactive monitoring and automated troubleshooting.
Markets
Cloud and artificial intelligence
Arista's largest customers are the operators of hyperscale data centres, and a substantial proportion of its revenue has come from a small number of them, principally Microsoft and Meta Platforms. This concentration is the most frequently cited risk in analysis of the company, since the loss or reduction of a single customer would be material.
The growth of artificial intelligence training has been the dominant recent driver. Training a large model requires thousands of processors working on the same computation simultaneously, exchanging enormous quantities of data between each step, and the network connecting them determines how much of the expensive processing capacity is actually used. A cluster whose network cannot keep up leaves processors idle, which is the most expensive form of waste in the industry.
This has made the back-end network, the one connecting processors to one another within a cluster, a subject of intense competition. The alternative technology, InfiniBand, supplied principally by Nvidia, has historically dominated high-performance computing. Arista's position is that Ethernet, with the enhancements developed under the Ultra Ethernet consortium, will prevail because it is an open standard supported by many suppliers, is familiar to the engineers who operate data centres, and scales to the sizes required. Which technology predominates in large training clusters is among the more consequential open questions in data centre infrastructure.
Campus and enterprise
Arista has extended from the data centre into campus networking, meaning the wired and wireless networks of offices, hospitals, universities and comparable sites, a market considerably larger by unit volume and dominated by Cisco. The proposition is the same operating system, the same management software and the same operational model as the data centre, which appeals to organisations that already run Arista in one and would rather not learn two systems.
Routing
The company also supplies routing platforms to service providers and to large enterprises, applying the same merchant silicon and common software approach.
Products
Switching platforms
Arista's hardware is organised principally into the 7000 series of Ethernet switches, spanning fixed-configuration units for the top of a server rack through to large modular chassis carrying hundreds of high-speed ports, alongside the 7130 series inherited from Metamako for applications where latency measured in nanoseconds matters, principally electronic trading.
The design philosophy throughout is the use of merchant silicon from suppliers including Broadcom and others rather than switching chips developed in-house. The argument is that a specialist semiconductor company serving the whole industry will out-invest any single equipment vendor's internal chip team, and that the vendor's effort is better spent on software and on system design. The approach was contrarian when the company was founded, since the incumbent's competitive position rested substantially on proprietary silicon, and it has since become the industry norm.
CloudVision
CloudVision is the management platform, and it is the commercial expression of the state-sharing architecture. Because every device publishes its state to a database rather than holding it privately, that state can be aggregated network-wide, which allows configuration to be applied as a single intended state rather than device by device, changes to be validated before deployment, and the network's condition at any past moment to be reconstructed for troubleshooting.
This is the basis of the company's move towards what the industry calls intent-based networking: the operator specifies what the network should do and the system determines and verifies the configuration required, rather than the operator configuring each device and inferring the result.
Security
The security portfolio, built on the Awake Security acquisition, applies the same principle. Because the network devices already observe all traffic and already publish state centrally, detection of anomalous behaviour can be performed on data the network is generating anyway, rather than by deploying separate monitoring infrastructure. The products are marketed under the heading of zero trust networking.
Context: the networking industry
Arista's position is best understood against how enterprise networking developed.
Ethernet, invented at Xerox PARC in the 1970s and standardised thereafter, won the local network standards contest of the 1980s and 1990s against Token Ring and other alternatives, principally because it was simple, cheap and open. Cisco Systems, founded in 1984, built the dominant position in the equipment that connects Ethernet networks together, and by the 2000s held a share of enterprise switching and routing that few companies have held in any market.
That position rested on three things: proprietary silicon, a proprietary operating system with a command syntax that a generation of network engineers had trained on and been certified in, and a sales and support organisation reaching every large organisation in the world. Each was a genuine barrier, and the graveyard of companies that attempted to compete is well populated.
What changed was the customer. The rise of very large data centre operators from the late 2000s created buyers with their own engineering capability, with networks far larger than any enterprise, with no attachment to the incumbent's certification programme, and with an explicit preference for open standards and multiple suppliers. They were prepared to buy from a company with no track record if the product was better, which no traditional enterprise buyer would have done.
Arista built for those customers specifically, and the fact that the same architecture subsequently proved attractive to conventional enterprises is a later development rather than the original plan.
Finances
Arista's financial profile is unusual for a hardware company. Gross margins have run in the region of sixty per cent or above, well beyond what hardware assembly alone would support, which reflects that a substantial part of the value is in software delivered with the hardware. Operating margins have exceeded forty per cent in recent years, and the company has carried no meaningful debt and accumulated substantial cash.
Revenue growth has been rapid but uneven, reflecting the timing of large customers' capital programmes: periods of very high growth have alternated with flat quarters when a hyperscale customer paused construction. The company's guidance has accordingly been treated cautiously by investors, and its share price has been correspondingly volatile despite consistently strong results.
Leadership
Jayshree Ullal has been chief executive since 2008 and chair of the board since 2022. She spent fifteen years at Cisco Systems, where she rose to senior vice president responsible for a business generating billions of dollars in revenue, before joining Arista.
She was named to Barron's list of the world's best chief executives in 2018 and 2019, and appeared on the Fortune Most Powerful Women list in 2026. She is among the small number of women leading large American technology companies and among a smaller number still to have taken one from founding to the Fortune 500.
Andy Bechtolsheim, a co-founder of Sun Microsystems and one of the most consequential hardware engineers of the period, has remained closely involved in Arista's product architecture, and Kenneth Duda has served as chief technology officer with responsibility for the software.
Assessment
Arista is an unusual case of a company succeeding against an entrenched incumbent in a market with strong switching costs, and the reasons are instructive.
It arrived at the moment when a new class of customer appeared. The operators of hyperscale data centres were, unlike traditional enterprises, technically sophisticated, willing to evaluate an unproven supplier, and interested in buying capable hardware cheaply and differentiating through their own software. They were also building networks of a size that exposed the weaknesses in operating systems designed twenty years earlier. Arista's architecture, with its state separation, fault isolation and upgrade without downtime, answered problems those customers actually had.
The strategic risk is the mirror image of the strength. Revenue concentrated among a handful of hyperscale buyers is revenue subject to those buyers' capital expenditure decisions and to their periodic interest in designing their own switches, which several have pursued. The expansion into campus, routing and security is a response to that, and it takes the company into markets where Cisco's incumbency is far stronger and where the technical argument that won the data centre carries less weight.
The artificial intelligence build-out has, for the present, made the concentration profitable rather than dangerous, since the customers concerned are spending on networking at rates without precedent. Whether that persists, and whether Ethernet or InfiniBand prevails in the clusters, will determine a good deal of the company's next decade.