Arista Networks company profile: How EOS and Ethernet are powering its AI data center expansion

Arista Networks, Inc. (NYSE: ANET) is a Santa Clara, California-headquartered networking technology company that develops switching, routing, wireless, software and network-management products for artificial intelligence infrastructure, hyperscale cloud data centers and enterprise networks. Founded by Andy Bechtolsheim, Ken Duda and David Cheriton, the business evolved from Arastra, established in 2004, before launching as Arista Networks in 2008 under the leadership of Chief Executive Officer Jayshree Ullal.

Arista Networks completed its New York Stock Exchange listing in 2014 and subsequently joined the S&P 500. Over the following decade, the company developed from a challenger in high-speed cloud switching into one of the most important suppliers of networking infrastructure used by hyperscale cloud operators, large enterprises and increasingly artificial intelligence infrastructure providers.

The foundation of that expansion has been the Extensible Operating System, commonly known as EOS. Instead of building separate operating environments for each generation of switching and routing hardware, Arista Networks developed a common software architecture intended to run consistently across its portfolio. That approach gave cloud operators a programmable and highly automated alternative to more tightly coupled legacy networking environments.

Artificial intelligence is now pushing that architecture into a new phase. Thousands or even hundreds of thousands of accelerators inside AI clusters need to exchange enormous quantities of data with very low latency, making networking one of the most important determinants of utilisation, training speed and overall infrastructure economics.

Arista Networks is positioning Ethernet, EOS and its broader Etherlink portfolio as an open networking layer capable of supporting that transition.

The financial scale already reflects significant momentum. Arista Networks generated approximately $9.01 billion of revenue in 2025, up 28.6% from $7.00 billion in 2024, while second-quarter 2026 revenue reached a record $3.04 billion, increasing 37.7% year over year. GAAP operating margin reached 45.4% during the quarter and net income was approximately $1.21 billion, demonstrating that rapid AI infrastructure expansion is occurring alongside unusually strong profitability.

The central question for the Arista Networks company profile is therefore no longer whether Ethernet can support large-scale cloud environments. That has already been demonstrated. The more important question is whether Arista Networks can extend its software-led architecture into increasingly specialised AI infrastructure while simultaneously building a much broader enterprise networking franchise.

How did Arista Networks use EOS to challenge traditional networking architecture?

The core strategic asset behind Arista Networks is not simply its portfolio of Ethernet switches. EOS was designed as a common, programmable network operating system running across multiple generations of hardware, allowing network state and software processes to be managed in a consistent architecture.

That design allowed Arista Networks to bring automation, programmability and telemetry into large-scale networking environments at a time when many traditional networks still depended heavily on manual device-by-device configuration.

The architecture aligned naturally with hyperscale cloud operators, which increasingly wanted networks to behave more like programmable infrastructure than collections of individual appliances.

EOS subsequently became the foundation for a much wider software stack. Arista Networks’ Network Data Lake aggregates device state, telemetry, packet information, flow data, sensor information and third-party inputs, while CloudVision provides network automation and management across data centers, campus infrastructure, routing and multi-cloud environments.

Arista AVA extends that model into artificial intelligence-assisted network operations by analysing infrastructure conditions and helping operators identify potential failures, security issues and configuration problems.

This software consistency is commercially important because customers can deploy different generations of Arista Networks hardware while maintaining a relatively uniform operational model. It also allows the company to use merchant silicon supplied by specialist semiconductor manufacturers rather than designing every switching chip internally, concentrating more of its differentiation on system architecture, software, reliability and network operations.

The model also creates dependencies. Arista Networks relies heavily on third-party suppliers for advanced silicon, memory, optics and manufacturing, meaning rapid AI infrastructure growth can simultaneously increase customer demand and intensify supply-chain risk.

Why has Ethernet become an important battleground in artificial intelligence infrastructure?

Traditional data center networks were largely designed to connect servers, storage systems, users and external applications. Artificial intelligence clusters create a very different traffic pattern because large numbers of accelerators may need to exchange data continuously while training or serving models.

If those accelerators cannot communicate efficiently, expensive computing hardware can remain underutilised. Networking therefore directly affects the economic return on AI infrastructure.

This has created an industry-wide contest around how accelerators should be interconnected. Specialised and proprietary technologies continue to play an important role, but Arista Networks is betting that high-speed Ethernet can increasingly support scale-out, scale-up and scale-across architectures while preserving an open ecosystem of switches, accelerators, network interface cards and optical components.

Ethernet already benefits from an enormous installed ecosystem. Enterprises, cloud operators and network engineers understand how to operate it, while semiconductor and optics companies continue increasing bandwidth and improving congestion management specifically for AI workloads.

The challenge is that AI networks impose much stricter requirements on latency, packet loss and congestion than conventional enterprise traffic. Arista Networks therefore needs to demonstrate that Ethernet can provide not merely enormous headline bandwidth but predictable performance when very large accelerator clusters are operating simultaneously.

How is the 7060XE7 portfolio changing Arista Networks’ role inside AI data centers?

Arista Networks introduced the 7060XE7 family as a new generation of AI networking systems capable of supporting connectivity of up to 1.6 terabits per second per port.

The platforms deliver up to approximately 102.4 terabits per second of system bandwidth, giving Arista Networks substantially greater capacity for next-generation AI fabrics. The first air-cooled 64-port systems are scheduled for availability in the fourth quarter of 2026, with liquid-cooled and higher-density configurations following in the first quarter of 2027.

The significance extends beyond faster switching.

Arista Networks increasingly describes its role in terms of rack-scale AI networking, where switches, optics, cooling and accelerator connectivity must be designed together rather than treated as completely separate infrastructure layers.

This change reflects the physical reality of AI data centers. As accelerator racks consume extraordinary amounts of electrical power and require sophisticated liquid cooling, networking equipment can no longer be considered an isolated layer installed after compute decisions have already been made.

Cable lengths, optical power, switch placement, thermal design and network topology can affect the architecture of the entire rack.

Arista Networks is working with semiconductor and cloud partners around this transition. The 7060XE7 generation uses next-generation switching silicon including Broadcom Tomahawk 6, while the company has highlighted collaboration with Oracle Cloud Infrastructure and Advanced Micro Devices around large-scale Ethernet AI networking.

The commercial opportunity is substantial, but the most important test will be how quickly qualification programmes become repeatable production deployments.

Why is Arista Networks developing liquid-cooled XPO optics for AI clusters?

Bandwidth is only one part of the AI networking challenge because optical connectivity itself consumes increasing amounts of power and physical space.

As network links move toward 800-gigabit and 1.6-terabit speeds, traditional pluggable optics can become increasingly demanding from a thermal and power perspective.

Arista Networks helped establish the XPO multi-source agreement around a 12.8-terabit liquid-cooled optical module designed specifically for high-density AI networks. The architecture is intended to support as much as 204.8 terabits per second of front-panel density per Open Compute Project rack unit.

The company has also modelled potential infrastructure benefits from the technology.

In an Arista Networks reference design for a hypothetical 400-megawatt AI data center containing 128,000 graphics processing units, the company estimated that XPO could reduce required switch racks by approximately 75% and save around 44% of floor space compared with the modelled conventional optical architecture.

Those figures should not be interpreted as universal savings. They are scenario-specific engineering estimates based on assumptions about one large theoretical deployment.

The broader strategic point remains important. Lower-power optics can allow more of a data center’s limited electrical capacity to be allocated to accelerators, while greater bandwidth density can reduce physical infrastructure requirements and potentially simplify the overall network.

Arista Networks is therefore moving deeper into the technologies surrounding the switch itself, even where it does not manufacture every optical component.

How does CloudVision expand Arista Networks beyond hardware?

Hardware still generates most of Arista Networks’ reported revenue, so describing the company primarily as a software business would be misleading.

In 2025, product revenue accounted for approximately $7.58 billion of the company’s $9.01 billion total revenue, while service revenue contributed around $1.43 billion.

Software nevertheless plays an outsized strategic role because EOS, CloudVision and related applications help determine why customers choose Arista Networks hardware and how deeply the company becomes integrated into network operations.

CloudVision provides a common management layer across data center switching, campus networking, wireless infrastructure, routing and multi-cloud environments. Rather than managing every networking domain through an entirely separate operating model, Arista Networks is attempting to give customers a common view based on telemetry collected through EOS and the Network Data Lake.

Services create another recurring component. Arista A-Care provides technical support, software updates and hardware replacement options, with service revenue rising 31.3% year over year during Q2 2026 as the installed base expanded and support contracts renewed.

This creates an important economic dynamic. A switch sale produces substantial upfront product revenue, but customer relationships can continue through support, software, management and future hardware refreshes.

As Arista Networks expands further into campus, branch and wide-area networking, the number of infrastructure domains governed through the same software architecture can increase.

Why did Arista Networks acquire VeloCloud and expand into enterprise WAN networking?

Arista Networks historically built its strongest franchise in cloud and data center networking, while larger incumbent networking vendors maintained much broader portfolios spanning campus, branch and wide-area infrastructure.

The acquisition of the VeloCloud SD-WAN business from Broadcom addressed one of those gaps.

Arista Networks completed the acquisition on June 30, 2025, for approximately $300 million in cash, with the transaction publicly announced the following day.

VeloCloud provides software-defined wide-area networking technology that allows enterprises to connect branches, corporate sites and cloud applications while intelligently managing traffic across different connectivity providers.

The strategic fit extends beyond simply adding another product.

Arista Networks already offered data center switching, campus infrastructure, Wi-Fi and routing. VeloCloud expands that footprint into the branch and WAN edge, allowing the company to offer a broader networking architecture to enterprise customers.

This supports Arista Networks’ wider “Centers of Data” strategy, under which artificial intelligence centers, conventional data centers, campuses and WAN environments increasingly operate as interconnected domains rather than separate networking silos.

Earlier acquisitions including Untangle and Pluribus Networks also broadened Arista Networks into security-oriented edge networking and distributed cloud fabrics.

VeloCloud is particularly significant because SD-WAN is already deeply embedded in enterprise infrastructure, giving Arista Networks access to customers and use cases that extend beyond the hyperscale environments where its original reputation was established.

How dependent is Arista Networks on a small number of hyperscale customers?

Customer concentration remains one of the most important structural risks in the Arista Networks business model.

Two end customers each accounted for more than 10% of annual revenue during 2025, representing approximately 26% and 16% respectively. Together, those relationships therefore contributed around 42% of company revenue.

Arista Networks did not identify those two customers by name in its latest annual concentration disclosure.

Large hyperscale customers create enormous opportunities because a single network deployment can require substantial quantities of switches, optics and related services. The same concentration also creates volatility because capital expenditure plans at one or two cloud operators can materially affect company-wide demand.

Pricing creates another tension. Large-volume customers typically receive discounts, meaning an increasing concentration of hyperscaler revenue can create gross-margin pressure even while total revenue rises.

Arista Networks is attempting to diversify this exposure by expanding across enterprise customers and specialist AI infrastructure providers.

Its 2025 strategic customer mix was approximately 48% Cloud and AI Titans, 32% Enterprise and 20% AI and Specialty Providers, showing that the business has become broader than its largest hyperscale relationships even though customer concentration remains significant.

Expansion into campus, WAN and security therefore has a financial purpose alongside its technology rationale. Greater enterprise penetration could make revenue less dependent on the deployment cycles of a small number of enormous cloud customers.

What do Arista Networks’ latest financial results reveal about operating leverage?

Arista Networks’ recent results demonstrate unusually strong operating economics for a company still growing at a substantial rate.

Second-quarter 2026 revenue reached approximately $3.04 billion, up 37.7% from $2.21 billion a year earlier. Product revenue increased to around $2.61 billion while service revenue reached approximately $430.5 million.

Gross profit was roughly $1.91 billion, producing a consolidated gross margin of approximately 62.9%.

GAAP operating margin reached 45.4%, compared with 44.7% a year earlier, while non-GAAP operating margin increased to 49.9%.

Net income rose to approximately $1.21 billion from $888.8 million, while diluted GAAP earnings per share increased to $0.95 from $0.70.

The combination matters because rapid artificial intelligence infrastructure growth does not automatically produce attractive supplier economics. Hardware companies can face heavy inventory requirements, high component costs and aggressive pricing pressure, yet Arista Networks is currently converting a substantial portion of incremental revenue into operating profit.

The company also generated approximately $2.78 billion of operating cash flow during the first six months of 2026, compared with $1.84 billion during the corresponding period a year earlier.

Management’s third-quarter 2026 outlook called for approximately $3.3 billion of revenue and a non-GAAP operating margin of between 48% and 49%. Those figures remain management expectations rather than guaranteed outcomes, but they indicate that Arista Networks expected demand momentum to continue after crossing $3 billion of quarterly revenue.

Why does Arista Networks maintain such a large liquidity position?

Arista Networks had approximately $13.3 billion of cash, cash equivalents and marketable securities at June 30, 2026, including around $2.29 billion of cash and approximately $11.05 billion of marketable securities.

That financial position gives management considerable flexibility to fund research and development, increase component commitments, make acquisitions, repurchase shares or absorb volatility in hyperscaler deployment schedules without relying heavily on external financing.

Liquidity is particularly relevant during the current AI infrastructure cycle because Arista Networks is carrying substantial inventories and making larger supply commitments to reduce lead times.

Inventories stood at approximately $2.54 billion at June 30, including around $616 million of evaluation inventory placed with customers for qualification and deployment programmes.

That inventory should not automatically be interpreted as a sign of weak demand because new networking platforms often require extended customer testing. It nevertheless increases risk if expected AI deployments are delayed or products under evaluation do not progress into commercial orders.

Arista Networks’ cash resources should also be considered alongside its exceptionally large supply commitments.

At June 30, the company had approximately $9.7 billion of non-cancellable purchase obligations, including around $9.4 billion expected within the following 12 months. Management increased those commitments as rapid AI-network deployments coincided with tighter conditions in memory and silicon markets.

Those commitments can help secure scarce components and support customer demand, but they also expose Arista Networks to greater inventory risk if deployment schedules change or expected demand fails to materialise.

The company also reported approximately $8.4 billion of future revenue associated with deferred revenue, contract liabilities and other performance obligations, with around 91% expected to be recognised over the following two years.

That amount provides useful revenue visibility but should not be interpreted as an unconditional sales backlog because contract terms, shipment timing and support obligations affect eventual recognition.

What could challenge Arista Networks as AI networking becomes more competitive?

Competition is likely to intensify because AI networking represents one of the most valuable infrastructure markets created by the generative artificial intelligence boom.

Traditional networking companies, semiconductor vendors, accelerator suppliers and hyperscale cloud operators developing their own systems all have incentives to capture a larger portion of the networking stack.

Arista Networks’ open Ethernet strategy can be an advantage because customers may prefer architectures capable of supporting accelerators and other components from multiple suppliers.

That same openness means customers can also evaluate alternative switches, silicon and software rather than becoming completely dependent on one vendor.

Supply-chain risk remains substantial. Arista Networks relies heavily on third-party contract manufacturers and semiconductor suppliers, while tighter memory and silicon availability has encouraged management to make larger inventory and purchase commitments.

If suppliers cannot provide components quickly enough, Arista Networks may be unable to convert customer demand into recognised revenue. If demand weakens after commitments have already been made, the company could instead face excess inventory.

Customer concentration magnifies those risks because changes in architecture or capital spending at one or two large cloud companies can materially affect overall demand.

Artificial intelligence infrastructure also evolves rapidly, forcing Arista Networks to anticipate changes in accelerator connectivity, optics, cooling and networking protocols well before large deployments reach production.

Enterprise expansion introduces a different competitive challenge. Campus networking, Wi-Fi, SD-WAN and branch infrastructure are markets where Arista Networks does not have the same historical dominance it developed in hyperscale switching.

VeloCloud and the wider enterprise portfolio therefore need to win customers against deeply established competitors.

What will determine whether Arista Networks remains a critical networking platform in the AI era?

Arista Networks has already demonstrated that its EOS-led architecture can scale from an unconventional networking startup into a company serving more than 10,000 customers and supporting more than 150 million cloud networking ports.

The challenge ahead is not reproducing that first transformation. It is extending the architecture into an artificial intelligence infrastructure market where technical requirements are changing faster than conventional data center networks did.

The first measurable test will be AI fabric adoption.

The 7060XE7 portfolio, Etherlink architecture and emerging 1.6-terabit systems need to move from customer qualification into repeatable production deployments across hyperscalers, AI companies and specialist cloud providers.

Sustained growth will depend on those systems demonstrating that Ethernet can provide the latency, reliability and congestion management required when tens of thousands of accelerators operate together as one computing system.

The second test will be whether Arista Networks can preserve its software differentiation as networking hardware becomes increasingly specialised.

EOS, CloudVision, the Network Data Lake and AVA provide a common operational layer spanning multiple generations of equipment. If customers increasingly value consistent automation, observability and policy management across AI, data center, campus and WAN environments, software may become more important rather than less important as physical infrastructure becomes more complex.

Enterprise diversification represents another important threshold.

VeloCloud gives Arista Networks a credible wide-area networking product, while campus switching, wireless networking and security broaden customer relationships beyond the data center. Greater adoption across these areas could gradually reduce dependence on a relatively small number of hyperscale customers.

The final test will be economic discipline.

Artificial intelligence networking can support exceptional growth, but large purchase commitments, higher inventories, expensive optics and rapid product transitions create substantial execution risk when demand forecasts are wrong.

Arista Networks currently has strong margins, substantial cash generation and significant liquidity, giving the company considerable room to manage that transition. Those economics will nevertheless need to remain resilient as AI infrastructure becomes increasingly dense and hardware-intensive.

Arista Networks has already moved far beyond being simply a manufacturer of high-speed switches. EOS provided the software foundation, hyperscale cloud established the data center franchise, enterprise expansion widened the addressable market and artificial intelligence is now pulling networking deeper into the architecture of the computer itself.

Whether Arista Networks remains one of the key beneficiaries of that shift will depend on how effectively it converts Ethernet’s openness, EOS software consistency and its expanding AI networking portfolio into reliable production infrastructure at massive scale.