Jayshree Ullal has made a hardware supplier look like a software company at precisely the moment artificial intelligence is making networks harder to build. Arista Networks reported first-quarter 2026 revenue of $2.71 billion, up 35 per cent, and GAAP operating income of $1.16 billion. Its operating margin was 42.7 per cent; on the company's adjusted measure, it reached 47.8 per cent. Cash from operations was $1.69 billion. Few businesses exposed to chips, optics and large customers combine that growth with those margins.
The same quarter captures the risk beneath the achievement. Arista's 2025 filing showed that two customers generated 26 per cent and 16 per cent of annual revenue. Together, they accounted for $3.78 billion of the company's $9.01 billion total. Arista does not name them in the filing. Their scale gives the company technical credibility and purchase volume, but it also means two procurement decisions can alter growth, pricing and inventory across the supplier.
Ullal's next leadership task is therefore not simply to sell more switches into the AI boom. It is to turn leadership with a few cloud buyers into a broader architecture spanning AI clusters, conventional data centres, campuses and wide-area networks. Arista calls that strategy its centres of data. The economics will be durable only if the operating system and network data become more valuable as hardware standards change.
Ethernet's second act
Arista's first disruption was cloud networking. It used merchant silicon, a consistent Extensible Operating System and a leaf-spine architecture to challenge incumbent networking vendors inside hyperscale data centres. Customers could automate large fleets and obtain detailed telemetry without accepting a vertically closed hardware stack. That model suited technology groups building infrastructure at unprecedented scale.
AI gives Ethernet a second, more demanding opportunity. Training and inference clusters move enormous volumes of data among accelerators. Congestion or packet loss can leave costly processors waiting, lowering utilisation and extending the time required to complete a job. The network is therefore part of the computing system's economics rather than a background connection.
Scale-out networks, which connect accelerators across racks, are moving from proprietary technologies such as InfiniBand towards enhanced Ethernet. Arista participates in the Ultra Ethernet Consortium and sells Etherlink systems designed for low latency, congestion control and visibility into AI workloads. Its 7800 universal AI spine uses large buffers and virtual output queues to manage bursts without blocking other traffic. A single EOS software base can cover the broader data-centre environment.
The more contested opportunity is scale-up, the extremely fast connection among accelerators within a rack or tightly coupled system. Proprietary links remain dominant because performance and integration matter more than openness at this layer. Arista supports an open Ethernet-for-scale-up effort, but adoption is not guaranteed. Accelerator designers may prefer interconnects tuned to their chips, while cloud companies can develop custom networking. Ullal must win a technical argument before scale-up becomes a material revenue pool.
Optics add another transition. In March, Arista helped establish a multi-source agreement for XPO, a liquid-cooled pluggable module delivering 12.8 terabits per second. The design is intended to reduce networking racks by as much as 75 per cent and save up to 44 per cent of floor space compared with conventional pluggable optics. It also seeks a middle path between familiar modules and co-packaged optics, bringing density and cooling gains without making the optical component inseparable from the switch.
These engineering choices determine serviceability, power consumption and vendor competition. If co-packaged optics mature quickly, the market may move past XPO. If operators value replaceable modules, the new standard can extend Arista's model. Supporting copper, radio-frequency and several optical reaches in one design increases flexibility but demands close coordination across suppliers.
Exceptional economics, concentrated power
Arista entered 2026 with strong momentum. Revenue grew 29 per cent in 2025 to $9.01 billion and GAAP net income reached $3.51 billion. The company had shipped 150 million cumulative ports. First-quarter product revenue was $2.31 billion and service revenue $398 million. Research and development increased to $344 million, yet operating expenses remained only about 19 per cent of revenue.
That efficiency reflects a focused portfolio, outsourced manufacturing and the leverage of software developed across many hardware generations. It also reflects large orders from sophisticated customers that require relatively little conventional marketing. The concentration makes sales efficient while increasing bargaining power on the buyer's side. A hyperscaler can qualify alternative suppliers, design its own equipment or delay a deployment with limited warning.
Arista describes its 2025 mix as 48 per cent Cloud and AI Titans, 32 per cent enterprise and 20 per cent AI and specialty providers. The two largest customers sit inside the first group, but the category is broader than two accounts. AI-specialty groups provide another growth route as model developers and cloud challengers build clusters. Many of those companies, however, depend on external funding or a few large contracts and may be less financially resilient than established cloud platforms.
Enterprise expansion is the diversification test. Campus switching, wireless access, security, observability and software-defined wide-area networking can bring EOS and CloudVision beyond the data centre. Arista acquired Broadcom's VeloCloud portfolio to connect branches and distributed sites. It appointed a president and chief operating officer and expanded senior leadership, giving Ullal more operating capacity for a broader go-to-market model.
The trade-off is competition. In cloud switching, Arista built share through technical focus. In campuses and branches, it meets vendors with long channel relationships, bundled security and extensive installed bases. Enterprise buyers expect local service, financing and integrated support. Revenue may become less concentrated, but selling costs can rise and the elegant margins of hyperscale hardware may be difficult to preserve.
Software must carry the moat
EOS is Arista's strongest defence against commoditisation. A common operating image across platforms reduces complexity, while the network data lake records state and telemetry that operators can analyse through CloudVision. The company can use that data for automation, troubleshooting and security. In 2025 it extended its Autonomous Virtual Assist with agentic capabilities for event correlation and continuous monitoring.
Artificial intelligence inside network operations is commercially different from networking for AI. The former can reduce labour and downtime for customers across ordinary infrastructure. It creates recurring software and service value even if the pace of accelerator deployment slows. It can also increase switching costs because operational processes and historical data accumulate around the platform.
Ullal must avoid turning that advantage into the lock-in Arista once criticised. Customers chose open, standards-based networking partly to preserve hardware and ecosystem choice. Agentic operations need transparent controls, auditable actions and support for third-party systems. A network assistant that works only with Arista devices may help retention but limit adoption in mixed estates.
Security is another convergence point. As campuses, clouds and AI clusters connect, network telemetry can identify abnormal behaviour and enforce segmentation. Arista can integrate security without becoming a full security-software conglomerate. The risk is strategic sprawl: each adjacent feature brings new competitors and support obligations.
India becomes an operating base
Asia is important to both supply and diversification. In 2025 Arista expanded its commitment to India through local production of high-performance data-centre and campus switches and enterprise access points. It also built an AI-for-networking centre of excellence focused on automated operations, analytics and self-healing systems. India is therefore not only a sales market; it contributes manufacturing, research and product engineering.
The arrangement can improve access to Indian enterprises and public projects that favour local production. It places engineers close to a large base of global capability centres operating complex networks for multinational companies. Domestic data-centre construction and sovereign artificial-intelligence initiatives create demand for high-performance Ethernet beyond the traditional US hyperscalers.
Localisation also introduces execution questions. High-end network equipment depends on semiconductors and optics sourced across Asia and the United States. Assembly in India does not eliminate those dependencies. Component qualification, quality control and export rules must be managed without fragmenting the product line. Arista's outsourced manufacturing model requires precise inventory decisions when lead times and customer orders change.
Japan, South Korea, Singapore and Australia offer mature enterprise and cloud markets, while Southeast Asia is adding data-centre capacity. Chinese demand is constrained by technology controls and local competition. Across the region, operators may prefer several suppliers to reduce geopolitical and procurement risk, which helps Arista enter accounts but makes concentration reduction a long process.
Growth must become less fragile
Arista forecast roughly $2.8 billion of second-quarter revenue and a non-GAAP operating margin between 46 and 47 per cent. That guidance implies continued growth with little immediate erosion in profitability. Its customer satisfaction score was also unusually high, supporting the claim that concentration reflects successful relationships rather than temporary discounting.
Yet the next architecture transition can change supplier positions quickly. Accelerators may integrate more networking. Scale-up links may remain proprietary. Optical technology may shift packaging. Customers may bring more design in-house. Cisco, Nvidia and specialist vendors have greater resources in parts of the stack. Arista must invest ahead without losing the cost discipline that distinguishes it.
Ullal has led Arista since 2008 and made consistent software across merchant hardware into one of enterprise technology's most profitable models. Her record justifies confidence in another transition. But a company valued for AI growth should be measured by more than shipments to the largest builders. Enterprise adoption, service revenue, AI-specialty customer quality and share across open scale-up networks will show whether the franchise is widening.
The decisive figure is not next quarter's margin. It is the 42 per cent of revenue attached to two customers. If Arista can keep those relationships while allowing the percentage to fall through faster growth elsewhere, Ullal will have turned hyperscale success into infrastructure breadth. If concentration remains fixed, the company's extraordinary economics will continue to depend on buyers powerful enough to redesign them.
