On June 9, 2026, Arista Networks announced a new family of 1.6-terabit switches for artificial-intelligence fabrics. The product language was dense, as networking language tends to be: 64 ports, rack-scale infrastructure, air and liquid cooling, scale-up and scale-out. Buried inside the specifications was a more consequential statement. Arista was no longer describing itself simply as a supplier of high-performance switches. It was preparing to deliver rack-scale systems.
That shift captures the point at which Jayshree Ullal’s company has arrived. Arista spent most of its modern history proving that an open, software-led Ethernet architecture could displace proprietary networking inside the world’s largest cloud data centres. It now has to carry the same argument into the most tightly coupled part of the AI machine, where accelerators communicate at extraordinary speed, power and thermal limits shape every design decision, and incumbent suppliers can combine chips, interconnects and software into one controlled system.
The commercial backdrop is formidable. Arista produced $9.006 billion of revenue in 2025, 28.6 per cent more than the previous year, while maintaining a 64.1 per cent gross margin and a 42.8 per cent operating margin under generally accepted accounting principles. Net income reached $3.511 billion. In the first quarter of 2026, revenue rose another 35.1 per cent to $2.709 billion, operating margin held at 42.7 per cent and operating cash flow reached $1.69 billion.
Those figures make Arista one of the rare infrastructure companies that can grow at a pace associated with software while producing the margins of a highly disciplined platform business. They also raise the standard by which Ullal must now be judged. A company valued for near-flawless execution does not have much room for an ordinary product transition.
The next phase is harder than the one Arista has already won. In cloud networking, the company could challenge a slower incumbent with merchant silicon, one operating system and a close relationship with technically demanding customers. In AI, it faces competitors that own more of the compute stack, customers powerful enough to design their own systems and a market in which the network has become inseparable from the economics of the data centre.
Two unnamed customers accounted for 26 per cent and 16 per cent of Arista’s 2025 revenue. The company relies principally on Broadcom for merchant switching silicon. Its fastest growth is tied to a capital-spending cycle concentrated among a small group of hyperscalers and AI builders. The very customers that validated Arista’s model have the engineering resources and purchasing leverage to reshape it.
Ullal’s answer is to make Ethernet more capable, Arista’s software more pervasive and the company’s addressable market broader. The Etherlink portfolio is moving from connecting AI clusters across racks towards connecting accelerators within rack-scale systems. CloudVision, the network data lake and AI-assisted operations are intended to turn hardware into a continuously managed service. Campus networking and the VeloCloud acquisition extend the same architecture from the data centre to branch offices and industrial sites.
This is an expansion of product scope, but it is also a test of corporate identity. Arista became exceptional by refusing the complexity and fragmentation that accumulated inside larger networking companies. It now wants to serve AI centres, data centres, campuses and wide-area networks without becoming the kind of sprawling supplier it once disrupted.
Ullal made Ethernet the operating language of the cloud. Her defining task is to make it the fabric of AI without allowing Arista’s discipline to disappear inside the opportunity.
A former Cisco executive chose the architecture Cisco could not easily copy
When Ullal joined Arista in 2008, the company was a small Silicon Valley start-up with celebrated founders, an unconventional operating system and almost no commercial scale. She arrived from Cisco, where she had spent more than 15 years and led a data-centre, switching and services operation responsible for roughly $10 billion of business.
The move was not a retreat from corporate life. It was a wager that the structure of networking was about to change. Public cloud companies were building computing estates at a scale traditional enterprises had never attempted. They needed networks that could be automated, observed and modified through software. They also wanted greater control over hardware costs and product roadmaps than established vendors were accustomed to giving them.
Arista’s founders had designed the technical foundation for that market. Andy Bechtolsheim brought systems and silicon judgement. David Cheriton contributed distributed-systems insight. Kenneth Duda created the Extensible Operating System, or EOS, around a state-oriented architecture running on standard Linux. Ullal supplied the commercial discipline, customer access and organisational force required to turn the engineering into a company.
The strategic choice was deceptively simple. Arista would not design every important switching chip itself. It would use high-performance merchant silicon, initially giving customers the economic benefits of a broader semiconductor ecosystem. The company would concentrate its proprietary value in software, system design, quality, support and the speed with which it could translate customer requirements into deployable products.
Cisco could source merchant silicon too, and it had vastly greater sales reach. What it could not easily reproduce was Arista’s organisational coherence. Established product families carried separate software histories, commercial commitments and internal constituencies. Arista had one operating system from the beginning. A feature written for one part of the portfolio could be extended across the rest without maintaining a collection of incompatible code bases.
That consistency mattered to cloud customers more than traditional branding. At hyperscale, operational variation becomes cost. A different command structure, software release or failure mode across product lines forces engineers to maintain exceptions. Arista offered a network that could be treated more like a distributed computing platform than a collection of appliances.
Ullal also understood the psychology of the buyers she was pursuing. Cloud engineers did not want a vendor to conceal complexity behind proprietary packaging. They wanted direct access to product teams, programmable systems and evidence that the supplier would adapt around their architecture. Arista’s intimacy with a small number of demanding customers became an advantage before it became a concentration risk.
The company went public in 2014. It entered the S&P 500 in 2018. By the end of 2025, it had shipped 150 million ports and built a business whose profitability would have been difficult to imagine when Ullal left Cisco. The achievement was not merely taking share from an incumbent. It was recognising that cloud networking would be won by a different operating model.
EOS turned a hardware catalogue into a compounding system
The centre of Arista’s model remains EOS. The software stores system state in a central database and distributes changes through a publish-and-subscribe mechanism. Processes can operate independently, recover without bringing down the whole switch and expose information to automation tools in a consistent form.
That design sounds technical because it is. Its business consequence is easier to grasp. Arista can add hardware platforms without recreating the operating environment around each one. Customers can manage fixed switches, modular chassis, campus products and routing systems through a common set of behaviours. Product expansion therefore has a lower organisational cost than it would under a fragmented architecture.
The operating system also produces data. Telemetry from the network can be stored in Arista’s NetDL data lake and managed through CloudVision, giving customers a continuously updated view of configuration, performance, security and events. A network that once required administrators to inspect individual boxes becomes a system that can be queried, automated and audited.
This is why Arista’s economics should not be confused with those of a commodity hardware assembler. Merchant silicon reduces the need to fund every chip development programme, but the value does not disappear. It moves into software architecture, board design, signal integrity, power efficiency, testing and support. Customers are buying reliable behaviour across thousands of devices, not simply ports in a metal enclosure.
The model has compounded over time. Each new deployment creates more operational data. Each additional platform makes the common software environment more useful. Features developed with the largest customers can be adapted for enterprises. Support experience feeds product quality. The installed base produces service revenue and lowers the perceived risk of choosing Arista for another part of the network.
Ullal has protected that compounding effect by resisting the temptation to manufacture growth through unrelated acquisitions. Arista has bought capabilities, including wireless networking, network detection and response, observability and wide-area networking, but the strategic filter has generally remained clear: the acquired technology must strengthen the common operating system and data architecture.
The discipline becomes more important as the portfolio widens. A single operating system is an advantage only if customers experience the company as one system. If campus, security, AI and wide-area products acquire separate commercial and technical layers, Arista’s principal differentiation begins to erode.
Ullal’s record suggests that she understands this danger. Arista does not usually enter a category merely because the revenue pool is large. It enters when the company believes EOS, telemetry and cloud operating principles can change the economics of that category. AI is the largest test yet of whether that method still scales.
The financial machine leaves little room for strategic confusion
Arista’s 2025 income statement reveals an unusually efficient company. Revenue exceeded $9 billion with approximately 5,115 full-time employees. Research and development spending was $1.237 billion, up 24.1 per cent, while total operating expenses represented only 21.3 per cent of revenue. Cash generated from operations reached $4.372 billion.
The company ended the year with $10.7 billion of cash, cash equivalents and marketable securities. That balance gives Ullal the ability to secure components, fund new product introductions, repurchase shares and make acquisitions without relying on debt markets. It also creates a buffer against the abrupt changes in demand and supply that have become normal in AI infrastructure.
High margins are not incidental to the strategy. They indicate that Arista has retained pricing power even while using widely available silicon. The company’s value rests in the complete system and its operating reliability. Customers may negotiate aggressively over hardware, but the cost of network instability inside a cloud or AI cluster is far larger than the saving achieved by selecting a cheaper switch.
The first quarter of 2026 extended the pattern. Product revenue reached $2.311 billion and service revenue was almost $398 million. Non-GAAP operating margin was 47.8 per cent. Arista guided to approximately $2.8 billion of second-quarter revenue while keeping expected non-GAAP operating margin between 46 and 47 per cent.
Few companies can expand this quickly without allowing expenses to grow ahead of sales. Ullal’s organisation has done the opposite. Research and development rose materially in 2025, including higher personnel and prototype costs, while operating leverage improved. The company spent more on the future without weakening the present.
That combination shapes investor expectations. Arista is not valued as a cyclical equipment vendor that can excuse a margin decline during a major transition. It is expected to deliver innovation, growth and profitability at the same time. Ullal herself has repeatedly framed those three qualities as inseparable.
The constraint is strategic selectivity. A large cash balance can make adjacent markets appear more attractive than they are. Strong margins can conceal the early inefficiency of new businesses. Rapid AI demand can encourage suppliers to build capacity before standards and customer architectures have settled.
Arista’s advantage has always been judgement about where not to compete. The financial machine will remain exceptional only if Ullal preserves that restraint while the AI market offers more plausible opportunities than any networking company can pursue coherently.
The hyperscalers built Arista’s scale and remain its largest vulnerability
Arista’s relationship with giant cloud customers is the source of its technical authority. These companies operate networks large enough to expose weaknesses that ordinary enterprise deployments may never encounter. They force suppliers to improve congestion management, telemetry, failure recovery, power consumption and software automation. Winning their business is a form of product validation.
It is also a source of financial concentration that cannot be diversified away quickly. Arista’s two largest end customers represented 42 per cent of 2025 revenue, up from 35 per cent in 2024. One accounted for 26 per cent and the other for 16 per cent. The company’s top two resellers represented more than half of year-end accounts receivable.
The customers are unnamed in Arista’s annual filing, which is less important than the structure of the dependence. A delay in one hyperscaler’s project, a shift towards internally designed equipment or a decision to allocate more spending to a vertically integrated supplier can move billions of dollars of demand. Quarterly growth can remain excellent while underlying bargaining power becomes more concentrated.
AI makes the relationship even more complex. The largest buyers are no longer simply purchasing network capacity. They are designing complete computing systems around accelerators, memory, optics, power and software. Some can specify custom silicon, fund new interconnect standards or work directly with contract manufacturers. They may use Arista in one part of the architecture and an internally designed or rival system in another.
Ullal has managed concentration by becoming difficult to replace rather than pretending the exposure does not exist. Arista works closely with customers on roadmaps, qualifies products to demanding standards and embeds EOS into operating processes. The more network domains that use the same software and telemetry, the greater the cost of fragmentation.
Service quality reinforces that position. Arista reported a net promoter score of 89 in 2026, with 94 per cent of surveyed customers strongly positive. In enterprise technology, such figures are partly marketing instruments, but they also reflect an operating culture built around direct technical accountability. Cloud customers tolerate fewer layers between a problem and the engineers responsible for solving it.
Diversification must nevertheless become more than a long-term aspiration. Enterprise data centres, financial institutions, campuses, industrial networks and wide-area connectivity can reduce dependence on a handful of cloud titans. The challenge is that these markets require a broader sales channel, more varied product configurations and different support economics.
Arista became efficient by serving customers that bought at enormous scale and understood the technology deeply. A broader customer base is strategically safer but operationally more expensive. Ullal must diversify revenue without importing the complexity that Arista’s hyperscale model helped it avoid.
AI changes the network from a layer into part of the machine
Traditional data-centre applications can tolerate a degree of variation in network performance. AI training cannot. Thousands of accelerators work on one job, repeatedly exchanging large amounts of data. If part of the cluster waits for congested links or delayed packets, expensive compute sits idle. The relevant measure is not how fast an individual switch can move traffic under ideal conditions. It is how quickly and predictably the whole system can complete the job.
This turns networking into an economic control point. Accelerators are scarce, power-hungry and costly. A fabric that improves their utilisation can justify a premium because it increases the output of the entire data centre. A fabric that behaves unpredictably can destroy the economics of a multibillion-dollar deployment.
Arista is well positioned in scale-out networking, which connects racks and clusters across the AI data centre. Its leaf-and-spine architectures, large buffers, virtual output queuing and telemetry were developed through years of cloud deployment. Etherlink combines those systems with features designed for AI traffic, including load balancing and job-level observability.
The more difficult frontier is scale-up networking. This is the tightly coupled domain inside a rack or rack-scale system, where accelerators need very high bandwidth and low latency to behave as one computing unit. Proprietary interconnects have held an advantage because the same supplier can optimise the accelerator, link, switch and software together.
Arista’s strategic proposition is that Ethernet can move inward. An open standard can gain performance capabilities associated with proprietary systems while preserving customer choice across chips, optics and equipment vendors. The argument resembles the one Arista made in cloud networking: openness becomes commercially powerful when the engineering is good enough that customers no longer have to trade performance for flexibility.
The outcome is not predetermined. Nvidia can integrate networking with accelerators, systems and software at a depth Arista cannot match alone. Cisco, HPE after its acquisition of Juniper, Broadcom and specialist suppliers all have reasons to shape the standard. Hyperscalers may support openness while maintaining custom architectures that preserve their own differentiation.
Ullal does not need Ethernet to eliminate every proprietary link. Arista needs it to become a credible scale-up option for customers that want a multi-vendor ecosystem, and then to make EOS the best way to operate that environment. Even a partial opening of rack-scale AI could create a market comparable to the cloud networking transition that built the company.
The risk is timing. Standards can take longer to mature than capital markets expect, while proprietary systems improve continuously. Arista must invest before demand is fully visible without mistaking participation in an industry workstream for a product win.
Ethernet for scale-up is now a product roadmap, not a slogan
Arista’s work on Ethernet for Scale-Up Networks, known as ESUN, moved into public view in October 2025 through an Open Compute Project workstream. The objective is to create standards-based scale-up systems open to multiple suppliers. By 2026, the company was pairing that standards effort with products across switching, optics and rack design.
The 7060XE7 series announced in June is the clearest expression of the strategy. The platforms support 1.6-terabit interfaces and are designed for rack-scale AI infrastructure. Arista described them as the foundation for both scale-out and scale-up fabrics, with air-cooled and liquid-cooled configurations intended to operate at the power densities of next-generation systems.
The availability schedule matters. The first 64-port air-cooled version is expected in the fourth quarter of 2026. Additional liquid-cooled and 128-port configurations are planned for the first quarter of 2027. These are forward product commitments, not installed revenue. Ullal must convert a persuasive architecture into qualified deployments while customers are making some of the largest infrastructure decisions in corporate history.
Arista’s universal AI spine, built around the 7800 family, addresses a different part of the fabric. Virtual output queuing is designed to prevent head-of-line blocking, while deep buffers absorb the sudden traffic bursts produced by AI workloads. Cluster Load Balancing distributes flows more intelligently, and CloudVision Universal Network Observability links network behaviour to individual jobs.
That combination illustrates Arista’s preferred method. Hardware handles the physical demands. EOS controls the system. Telemetry explains what happened. The software layer then uses the data to improve operations. The company is not trying to win through a single benchmark; it is trying to make the entire fabric predictable enough that customers can operate it at scale.
Predictability may be more valuable than peak speed. AI infrastructure teams need to know why a training run slowed, which path created congestion and whether a component is about to fail. An open network that can be observed and automated may produce better fleet economics than a faster black box, particularly for customers running mixed accelerators and several generations of equipment.
The product roadmap also exposes Arista to a new kind of integration responsibility. Once the company moves from switches towards rack-scale systems, customers will judge thermal design, cabling, optics, mechanical compatibility and deployment support as part of the offer. The boundary between network vendor and systems supplier begins to dissolve.
Ullal built Arista by concentrating on the layer where it had unusual authority. AI is pulling the company across more layers because the customer’s problem no longer respects traditional product boundaries.
Optics, cooling and power are becoming executive decisions
The physical constraints of AI data centres are changing the economics of networking. Faster links consume more power and generate more heat. Dense racks shorten cable distances but make cooling harder. Pluggable optics provide flexibility and serviceability, yet the number and power of modules can consume valuable rack space and electrical capacity.
Arista’s XPO multi-source agreement, announced in March 2026, is an attempt to change that equation. The proposed liquid-cooled optical module delivers 12.8 terabits per second and supports a front-panel density of 204.8 terabits per open-compute rack unit. Arista says the design can reduce networking racks by as much as 75 per cent and save up to 44 per cent of floor space compared with traditional pluggable optics.
The numbers are engineering targets, but their strategic meaning is clear. Networking suppliers are now competing for power, cooling capacity and physical space, not merely bandwidth. A more efficient optical architecture can allow customers to install more compute within the same facility envelope. That moves the network into conversations once dominated by server, accelerator and data-centre design teams.
It also increases supply-chain exposure. Arista depends on a limited number of component providers and principally on Broadcom for switching chips. The company uses third-party manufacturers and does not control every capacity decision upstream. AI demand has already led it to increase purchase commitments and inventory in order to reduce lead times.
Merchant silicon was an advantage when it allowed Arista to ride a broad semiconductor roadmap without carrying the full cost of chip development. It can become a constraint when a predominant supplier serves competitors, sets delivery priorities or chooses a technical direction that does not align perfectly with Arista’s schedule.
Ullal’s response has been to deepen collaboration with suppliers while keeping the software and system architecture differentiated. The approach preserves capital efficiency but requires exceptional forecasting. Order too little and customers move to available alternatives. Order too much and a rapid standards change leaves the company with expensive inventory.
Tariffs and export restrictions add another layer. Components may cross several borders before a finished system reaches the customer. Trade policy can alter cost, availability and approved destinations with little warning. For a company whose revenue is concentrated in large projects, a delayed component can postpone recognition of an entire deployment.
AI networking is therefore becoming a capital-allocation business as much as a product business. Ullal must decide which optical technologies to support, how much inventory to secure, when to commit to liquid cooling and where open standards will mature quickly enough to justify capacity. The quality of those decisions will be visible in both revenue growth and gross margin.
Campus and wide-area networking are the diversification test
Arista’s expansion into campus networking is strategically logical and culturally difficult. Enterprises increasingly want the same automation, telemetry and operational consistency across data centres, offices, wireless networks and branches. EOS and CloudVision can provide a common control environment, reducing the number of systems an IT team must maintain.
The company has built wired and wireless products, network-access control, security features and industrial switches designed for demanding physical environments. Its VESPA technology supports large wireless mobility domains. Agentic capabilities in Arista’s autonomous virtual assistant correlate events, monitor systems and help troubleshoot across domains.
The $300 million acquisition of VeloCloud from Broadcom in June 2025 extended the architecture into software-defined wide-area networking. VeloCloud connects branches, campuses and data centres across public and private links. Integrated with Arista’s campus and data-centre portfolio, it can turn a collection of sites into one managed network.
This gives Arista a more complete enterprise proposition and a route to reduce hyperscaler concentration. It also moves the company into Cisco’s historic centre of strength: broad product coverage, channel distribution and long-standing enterprise relationships. HPE’s ownership of Juniper creates another integrated competitor across data centre, campus and wide-area networking.
The operating model is different from cloud sales. A hyperscaler may buy enormous volumes through a concentrated engineering relationship. Enterprise customers purchase through resellers and systems integrators, require local support and often modernise in stages. Sales and marketing costs can rise before revenue reaches comparable scale.
Arista must resist imitating the incumbents it is trying to displace. The objective is not to produce an equivalent box for every category. It is to make the enterprise network behave according to cloud principles: common software, real-time state, automation and fewer operational silos.
VeloCloud will be an early measure of integration discipline. If the product becomes another console and another sales motion, the acquisition adds breadth without simplicity. If it is absorbed into CloudVision and the EOS data model, Arista can offer a genuine client-to-cloud architecture.
Ullal’s phrase for the wider strategy is the connection of centres of data: AI centres, data centres, campus centres and wide-area centres. The language is broader than the company that entered the public market in 2014. The economic question is whether one architecture can serve all four without forcing Arista to build four separate organisations.
Ullal is making the leadership bench visible before it becomes urgent
Arista’s strategy has long been associated with a small group of unusually influential leaders. Ullal directs the business and customer agenda. Bechtolsheim shapes advanced systems, silicon and optics. Duda is the architect of EOS and the technical culture around it. The concentration has produced clarity, but it also creates institutional risk as the company becomes larger and more complex.
In 2025, Ullal began making the next management structure more explicit. Todd Nightingale joined as president and chief operating officer with responsibility focused on enterprise data centres, campus networks and manufacturing. He had previously led major networking businesses at Cisco and served as chief executive of Fastly. His appointment gives Arista an operator familiar with both the incumbent it challenges and the enterprise markets it wants to penetrate.
Duda was promoted to president and chief technology officer, formalising his authority over the technology system. Tyson Lamoreaux became senior vice-president for cloud and AI networking. Chantelle Breithaupt, who joined as chief financial officer in 2024, now oversees a business generating more than $4 billion of annual operating cash flow and carrying a far larger inventory and supply-chain commitment than Arista did before the AI cycle.
The division of responsibility is sensible. Nightingale can industrialise enterprise execution and manufacturing. Duda can protect the software architecture. Lamoreaux can focus on the largest cloud and AI customers. Ullal can remain the integrator across customers, capital allocation, product direction and culture.
It also creates a succession question that Arista can no longer leave abstract. Ullal has led the company since its commercial launch in 2008 and combines the roles of chief executive and chairperson. The board says it reviews CEO succession through its nominating and governance committee, but no public plan can capture the extent to which her judgement and customer credibility are embedded in the business.
The purpose of a strong bench is not merely to identify a future chief executive. It is to prove that Arista can make high-quality decisions without routing every issue through one leader. AI systems, campus expansion, supply-chain commitments and standards work are too broad for a founder-like command structure to remain the only source of coherence.
Nightingale’s appointment is especially revealing because Arista has historically been cautious about adding senior executives from outside. The company is accepting some integration risk in exchange for operating scale. Ullal must give new leaders enough authority to matter while preventing the organisation from acquiring the layers and politics that accompany many large technology companies.
The transition from exceptional team to durable institution is often where successful founder-era companies lose speed. Arista is approaching that transition while its market is accelerating, which makes deliberate succession planning more difficult and more necessary.
Governance has to match the scale of Ullal’s authority
Ullal’s influence is reinforced by ownership. As of April 2026, she beneficially controlled approximately 29.3 million Arista shares, or 2.3 per cent of the company, much of it through family trusts. Her economic exposure is therefore tied overwhelmingly to long-term equity value rather than cash salary.
Her annual base salary remains $300,000, with a target bonus of the same amount. Total compensation reported for 2025 was approximately $2.9 million, and the ratio to the median Arista employee’s compensation was about 15 to one. Those figures are restrained by the standards of a large US technology company, particularly one with Arista’s market performance.
Low cash pay does not eliminate governance questions. Combining the chair and chief executive roles places substantial authority in one office. Arista has appointed Daniel Scheinman as lead independent director, with responsibility for convening independent directors, communicating with the chief executive and raising issues on behalf of outside board members.
The board faced a warning in 2025 when only 62 per cent of votes cast supported the advisory resolution on executive compensation, down from more than 90 per cent in each of the previous two years. Investor concerns centred less on Ullal’s own pay than on a large off-cycle award to Duda, disclosure around short-term incentives and the absence of multi-year performance measures in some equity awards.
Arista responded with extensive shareholder outreach, including discussions with investors representing a substantial portion of outstanding shares, and revised elements of the compensation structure. The episode is useful because it shows the difference between a company with excellent operating results and one exempt from institutional scrutiny. Performance can explain unusual pay decisions; it should not prevent the board from justifying them.
Duda’s technical importance is difficult to overstate. EOS is the foundation of Arista’s value, and retaining its architect matters. Yet dependence on irreplaceable individuals is itself a risk that compensation cannot permanently solve. The stronger governance response is to build teams capable of extending the architecture beyond its original creators.
Ullal has earned unusually broad authority through 18 years of execution. The next stage of stewardship is to make that authority transferable. Investors should be able to believe in Arista’s standards, capital discipline and customer culture even when the people who created them are no longer making every major decision.
Her Asian connection is embedded in the operating system of the company
Ullal was born in London, raised in India and educated in the United States. Her career belongs to Silicon Valley, but the cross-border formation matters to the way Arista has built engineering capacity. She has never treated Asia as an ornamental market or a biographical footnote. It is part of the company’s talent base, supply chain and customer landscape.
Arista opened a research and development centre in Bangalore in 2010, only two years after its commercial launch. The decision placed Indian engineers close to the development of EOS and cloud networking rather than confining the operation to support or back-office work. As the company expanded, India became one of several locations contributing to a common software architecture.
That structure mirrors Ullal’s own career. She moved through engineering, product management and commercial leadership before becoming a chief executive. The distinction between technical and business authority has always been less rigid inside Arista than in companies where engineering is treated as a function to be managed rather than a source of strategy.
Asia also sits inside the physical economics of the business. Semiconductor production, component testing, optics, contract manufacturing and data-centre construction are distributed across the region. Trade restrictions, tariffs and tension around Taiwan can affect supply and delivery even when the customer and final assembly are elsewhere.
For Asian cloud providers, telecommunications groups, financial institutions and governments, Arista offers an architecture that does not require a US hyperscaler to own the entire stack. Open Ethernet, multi-vendor silicon and common software can support greater infrastructure choice, though the company itself remains subject to US export controls and the realities of a concentrated technology supply chain.
Ullal’s significance to Asian business is therefore larger than representation. She has shown how an executive formed across India and the United States can lead a globally consequential infrastructure company without making identity the centre of the corporate narrative. The evidence is operational: where engineering is located, how customers are served and which standards the company chooses to support.
That perspective becomes more valuable as AI infrastructure turns into an arena of national policy. Countries want access to compute, control over data and resilience against geopolitical interruption. Arista cannot resolve those tensions, but an open networking model can give customers more architectural room than a vertically closed system.
The hardest version of Arista begins now
Ullal’s first great achievement was to turn a start-up built by eminent engineers into a commercially disciplined challenger. Her second was to preserve that discipline as Arista became a public company, entered the S&P 500 and grew into one of the most important suppliers to the cloud economy.
The third phase requires a different form of leadership. Arista must invest ahead of a rapidly changing AI architecture, participate in standards before their commercial value is certain and support customers building systems more complex than any previous data centre. It must broaden into enterprise networking while maintaining hyperscale efficiency. It must reduce dependence on a small number of buyers without weakening the relationships that drive its product advantage.
Competition will be more integrated than it was during Arista’s rise. Nvidia can connect accelerators, networking and software. Cisco can use its installed base and growing AI portfolio. HPE and Juniper bring data-centre and enterprise assets together. Broadcom is both a critical supplier and a powerful strategic actor. Customers themselves can design hardware and influence standards.
Arista’s defence is not size. It is architectural trust. Customers believe EOS will behave consistently, telemetry will expose what the network is doing and the company will work directly on difficult problems. Those qualities are slow to build and easy to damage through organisational expansion.
The financial pressure will be equally demanding. A 64 per cent gross margin and operating margin above 40 per cent leave little tolerance for an undisciplined systems push. Rack-scale products, liquid cooling and optics can increase engineering and inventory requirements. Enterprise sales can raise operating costs. Large customer projects can make quarterly growth volatile.
Ullal has enough cash, technical credibility and market access to take the risk. What she cannot buy is timing. Ethernet must become capable inside the rack while customers are still willing to consider an open alternative. The 1.6-terabit roadmap must arrive when accelerator platforms and cooling systems are ready. CloudVision must turn network data into operational advantage before rivals make observability a standard feature.
The company does not need to own the whole AI stack. Indeed, Arista’s history argues against trying. Its opportunity is to become the neutral, programmable fabric across heterogeneous compute, the layer customers trust precisely because no accelerator or cloud platform controls it completely.
That is a larger ambition than selling switches and a narrower one than becoming another vertically integrated systems giant. The distinction is where Ullal has always operated best. She wins by choosing the part of the architecture that becomes more valuable as the rest of the industry grows more complex.
Arista’s first era proved that merchant silicon and one operating system could overturn the hierarchy of cloud networking. The AI era will determine whether the same principles can survive at rack scale, under higher power, tighter coupling and fiercer control of the stack.
If Ullal succeeds, Ethernet will not merely connect the AI factory. It will become one of the mechanisms that keeps the factory open.
