LWAYS OPEN
The artificial-intelligence infrastructure boom began with an extraordinary concentration of attention on GPUs. The next stage is increasingly about everything required to connect those processors efficiently. Arista Networks (ANET) occupies a strategically important position in that transition as AI clusters expand from thousands toward tens or potentially hundreds of thousands of accelerators. Its opportunity is therefore larger than conventional data-center switching, but so are the expectations. This FRL research study examines whether Ethernet-based AI networking can become another major layer of the infrastructure cycle, whether Arista can capture a durable share of that economics, and which technological, competitive and customer-concentration risks could challenge the thesis.
Research Thesis
The central finding of our research is that the AI infrastructure cycle is changing the economic importance of networking.
The first phase of generative AI infrastructure was dominated by accelerator scarcity. NVIDIA (NVDA) became the clearest expression of that constraint because access to high-performance GPUs determined how quickly companies could build AI computing capacity.
But adding accelerators does not create useful computing capacity in isolation.
Large AI models divide workloads across enormous numbers of processors. Those processors must continuously exchange data. As cluster size, model complexity and accelerator performance increase, the network connecting the system becomes progressively more important to overall utilisation and performance.
That creates a second-order infrastructure opportunity.
More AI compute → more data movement → higher bandwidth requirements → more sophisticated network fabrics → greater economic importance of AI networking.
Arista Networks is attempting to capture this transition through high-speed Ethernet switching, its EOS software architecture and the expanding Etherlink portfolio.
The structural opportunity is substantial. The principal uncertainty is whether Arista can preserve its technological and economic position as NVIDIA, Cisco and alternative networking architectures compete for the same rapidly expanding infrastructure budgets.
There is also a second distinction that matters for investors and market observers: a strong structural market does not automatically imply that every level of market valuation is equally attractive.
The stronger AI networking becomes, the higher the expectations Arista may ultimately have to satisfy.
What Does Arista Networks Do in AI Infrastructure?
Arista Networks provides high-speed networking systems and software that connect computing infrastructure inside cloud and AI data centers. In large AI clusters, accelerators must exchange enormous quantities of information with very low latency. Arista’s Etherlink platforms use high-speed Ethernet, including 400G, 800G and emerging 1.6T architectures, to create the network fabrics connecting these increasingly powerful computing systems.
This is becoming important because an AI data center is not simply a building filled with GPUs.
It is an integrated computing system.
If thousands of expensive accelerators cannot communicate efficiently, utilisation deteriorates and the economic return on the entire cluster can decline. Network performance therefore influences how efficiently infrastructure worth billions of dollars can actually operate.
Arista’s position is particularly interesting because the company has spent years developing high-speed Ethernet networks for some of the world’s largest cloud operators.
Its strategy is now expanding that expertise into what the company calls AI Centers.
Arista’s broader “Centers of Data” strategy combines AI Centers, conventional Data Centers, Campus Centers and WAN Centers around its Extensible Operating System, or EOS, and related network-management technology.
The significance is strategic: Arista is attempting to evolve from a supplier of high-performance cloud switches into a provider of networking architecture spanning the increasingly interconnected infrastructure of the AI economy.
Why AI Compute Creates a Networking Opportunity
The economics of AI networking begin with a relatively simple problem.
Accelerators are becoming faster, but AI workloads are simultaneously becoming larger.
Training and increasingly sophisticated inference workloads cannot always be processed by one accelerator or even one server. Work must be distributed across large clusters.
That creates enormous machine-to-machine traffic.
The challenge is particularly acute during collective operations, when accelerators repeatedly exchange and aggregate information across the cluster. A network bottleneck can leave expensive computing resources waiting for data.
The consequence is economically important.
If an organisation has invested billions of dollars in accelerators, servers, buildings, power and cooling, improving network efficiency can increase utilisation of the entire infrastructure base.
Networking expenditure therefore represents only part of the total AI data-center cost while influencing the productivity of substantially more expensive computing assets.
Arista’s Etherlink architecture is designed around this problem.
Its AI networking portfolio already spans 400G and 800G Ethernet, and in June 2026 the company introduced its next-generation 7060XE7 Series supporting 1.6-terabit networking.
The new architecture is designed for both scale-out and scale-up AI fabrics and reflects an important transition in the industry.
As AI systems move toward rack-scale architectures, the network increasingly becomes part of the computing system itself rather than an independent communications layer.
Arista describes its newest 7060XE7 systems as providing up to 102.4 terabits per second of total bandwidth with 800G and 1.6T connectivity options.
The direction is clear:
AI processors are getting faster, and the network connecting them must accelerate with them.
Arista’s Position in the AI Data-Center Value Chain
The AI infrastructure value chain can be simplified as:
AI Demand
→ Hyperscaler Capital Expenditure
→ Accelerators and Custom Silicon
→ Servers
→ Networking
→ Optical Connectivity
→ Power and Cooling
→ Data Centers
→ Cloud Capacity
→ AI Applications
Arista sits primarily in the networking layer.
That position is less visible than the accelerator market, but potentially attractive because the network connects nearly every major element of an AI cluster.
The company’s economic exposure also extends beyond hardware.
EOS provides a common software foundation across Arista’s networking platforms. This matters because very large data-center networks are difficult to operate manually. Automation, telemetry, configuration consistency and fault detection become increasingly valuable as infrastructure scales.
Arista therefore competes not simply on the number of ports contained in a switch but on the architecture surrounding those switches.
The company’s 2025 Form 10-K illustrates the changing business mix.
Approximately 48% of revenue came from Cloud and AI Titans, 32% from enterprise customers and 20% from AI and Specialty Providers.
By product category, approximately 65% of revenue came from the company’s Core segment, including AI, cloud and data-center networking.
Those figures demonstrate both the strength and one of the principal risks of the Arista model.
The company has unusually direct exposure to some of the world’s largest infrastructure spenders.
That provides access to extraordinary capital investment.
It also creates concentration.
Corporate Evidence: What Arista’s Latest Results Show
The most recent evidence remains constructive.
Arista released its second-quarter 2026 results on 4 August and provided a third-quarter revenue outlook above prevailing Wall Street expectations.
The market significance was not simply that Arista delivered another strong quarter.
Management continued to see sufficient networking demand to support further expansion while the company simultaneously broadened its business beyond its traditional cloud customer base.
This diversification matters.
Arista built its reputation serving enormous cloud environments where reliability, network automation and high-speed switching are critical. The company is now expanding further into enterprise campus and branch networking while AI creates another potential growth vector at the high-performance end of the portfolio.
The combination creates a more diversified long-term model than a pure AI infrastructure supplier.
At the same time, the newest product cycle shows that Arista is allocating substantial engineering resources directly toward AI fabrics.
The 1.6T portfolio announced in June represents more than an incremental increase in port speed. It is designed for rack-scale AI architectures where networking increasingly functions as a backplane for tightly integrated computing systems.
This transition could enlarge Arista’s addressable market if Ethernet becomes a major standard for large AI clusters.
But that outcome cannot yet be assumed.
Hyperscaler Capex and the Demand Engine Behind Arista
The most powerful external driver behind Arista’s opportunity is the unprecedented capital-spending cycle underway among global hyperscalers.
Alphabet has indicated that its 2026 capital expenditure could reach approximately $175–185 billion.
Meta expects approximately $130–145 billion of 2026 capital expenditure.
Microsoft continues to invest tens of billions of dollars each quarter in cloud and AI capacity. In its fiscal second quarter, Microsoft reported capital expenditure of $37.5 billion, with approximately two-thirds directed toward shorter-lived assets, primarily GPUs and CPUs. Management also indicated that customer demand continued to exceed available supply.
These figures should not be confused with Arista’s addressable revenue.
Hyperscaler capital expenditure includes GPUs, CPUs, buildings, land, electrical infrastructure, cooling, storage and many other investments.
But the spending does establish something strategically important.
The computing base that requires networking is expanding at an extraordinary rate.
As hyperscalers deploy more accelerators, network capacity must expand alongside them.
The relationship is not one-for-one, but it is structural.
The stronger and more persistent the AI compute cycle becomes, the larger the potential networking requirement.
For Arista, this makes hyperscaler capital expenditure one of the most important external variables to monitor.
Ethernet vs. InfiniBand: The Battle for AI Networking
This is where the Arista thesis becomes more complicated.
Ethernet is not the only architecture available for AI networking.
NVIDIA has built a powerful networking position around InfiniBand and increasingly around Spectrum-X Ethernet. NVIDIA’s ability to combine accelerators, networking and software creates an integrated infrastructure proposition that represents a significant competitive force.
Arista’s counter-position is based on open, standards-oriented Ethernet.
Ethernet has several potential advantages.
It is deeply established throughout global data-center infrastructure, has a broad supplier ecosystem and allows customers to avoid dependence on a single vertically integrated vendor.
For hyperscalers operating enormous heterogeneous infrastructure environments, those characteristics can be economically valuable.
Arista’s Etherlink strategy is therefore based on the proposition that Ethernet can provide the performance characteristics required for AI clusters while retaining openness and architectural flexibility.
The company’s systems incorporate congestion management, load balancing and other capabilities required to address the unusual traffic patterns generated by AI workloads.
The industry is also moving toward increasingly standardised AI Ethernet architectures through developments such as the Ultra Ethernet ecosystem.
But declaring Ethernet the inevitable winner would be premature.
NVIDIA controls an extraordinarily influential part of the AI computing stack. Its ability to optimise networking around its own accelerators creates technological advantages that independent networking suppliers must continuously match.
The most plausible outcome may not be a winner-takes-all architecture.
Different AI clusters may adopt different networking solutions depending on workload, scale, accelerator architecture, cost and customer preference.
For Arista, it may therefore be sufficient for high-performance Ethernet to capture a substantial share of a rapidly expanding market.
Competitive Position and Economic Moat
Arista’s strongest competitive asset may not be any individual switch.
It is the combination of hardware, software and operational experience accumulated through years of hyperscale deployment.
EOS provides a consistent operating environment across Arista platforms.
This can reduce operational complexity as networks expand and allows customers to automate and monitor infrastructure through a common architecture.
Reliability also matters disproportionately in AI infrastructure.
When a networking failure interrupts a large AI cluster, the economic cost is potentially much greater than the price of the failed networking component itself.
This creates an environment in which established performance and operational history can become meaningful barriers to entry.
Arista nevertheless faces formidable competition.
NVIDIA (NVDA) can integrate networking with the world’s dominant AI accelerator ecosystem.
Cisco Systems (CSCO) possesses enormous enterprise relationships and is itself seeing rapidly rising AI infrastructure demand. Cisco reported approximately $9.3 billion of hyperscaler AI infrastructure orders during fiscal 2026 and expects roughly $7.5 billion of AI infrastructure revenue in fiscal 2027.
Broadcom (AVGO) occupies another strategically important position because its switching silicon underpins significant portions of the networking ecosystem.
White-box networking represents another potential pressure point, particularly among hyperscalers capable of developing more infrastructure internally.
Arista therefore operates in a valuable market precisely because that market attracts powerful competitors.
Its moat will ultimately depend on whether EOS, engineering execution, Ethernet expertise and customer relationships continue producing enough operational value to prevent networking hardware from becoming increasingly commoditised.
Valuation Versus Fundamental Growth
This is the point at which company quality and investment analysis must be separated.
Arista can simultaneously possess:
a strong balance sheet,
high-quality margins,
excellent exposure to cloud infrastructure,
a credible AI networking opportunity,
and a demanding market valuation.
There is no contradiction between those statements.
Financial markets price expectations rather than simply current operating quality.
As Arista becomes more widely identified as an AI infrastructure beneficiary, future AI networking growth becomes progressively embedded in those expectations.
This raises the execution threshold.
Strong revenue growth may no longer be sufficient if markets expect exceptional growth.
Successful AI deployments may not be sufficient if markets already assume rapid adoption.
This phenomenon is becoming visible across AI infrastructure.
Cisco’s most recent results provide a useful illustration. The company reported strong AI infrastructure orders and guided annual revenue above expectations, yet its shares initially declined after the report because market expectations had already become demanding.
The lesson is broader than Cisco.
As the AI infrastructure cycle matures, expectation risk can become almost as important as fundamental risk.
For Arista, valuation discipline therefore remains essential when interpreting otherwise strong corporate evidence.
The Counterargument
The strongest counterargument to the Arista AI infrastructure thesis is not that networking demand disappears.
It is that the economics may be captured differently than markets currently expect.
Hyperscalers have enormous engineering resources and strong incentives to reduce infrastructure costs. They can develop proprietary systems, use white-box hardware or diversify suppliers.
NVIDIA can integrate networking more tightly with its accelerator platform.
Cisco is increasing its exposure to hyperscaler AI infrastructure.
New architectures could alter how AI clusters are constructed.
And perhaps most importantly, the hyperscaler capital-spending cycle itself could eventually slow.
The current AI infrastructure buildout assumes that enormous investments in computing capacity will produce sufficient economic returns through cloud services, advertising, software, AI agents and other applications.
If monetisation fails to keep pace with infrastructure spending, capital discipline could eventually become more important.
That would affect the entire value chain.
Arista would not need to lose technological leadership for growth expectations to fall. A slower expansion of the total infrastructure pool could be sufficient.
Risks to the Research Thesis
Customer concentration remains one of the most important structural risks.
Serving the world’s largest cloud companies provides enormous opportunity, but purchasing decisions by a small number of customers can materially influence revenue.
Technology is the second major risk.
AI networking is developing rapidly. The transition from 400G to 800G and now toward 1.6T demonstrates how quickly architectures evolve. Leadership must therefore be continuously re-earned.
Competition is the third.
NVIDIA, Cisco, Broadcom-based platforms and internally developed hyperscaler solutions ensure that attractive economics will not remain uncontested.
A fourth risk is AI infrastructure overinvestment.
If current capital expenditure eventually creates excess computing capacity, orders could become cyclical even if the long-term AI trend remains intact.
Finally, valuation itself can amplify operating disappointment.
When expectations become elevated, even fundamentally strong results may generate weak market reactions if they fail to exceed what has already been discounted.
The clearest thesis breaker would be sustained evidence that Arista is losing strategic relevance inside next-generation AI clusters despite continued growth in total AI infrastructure spending.
If AI compute continued expanding rapidly while Arista consistently failed to participate in the corresponding networking opportunity, the structural thesis would require fundamental reassessment.
FRL Research Outlook
Arista Networks occupies one of the more interesting second-order positions in the AI infrastructure value chain.
The company does not manufacture the accelerators driving the AI revolution.
Instead, it provides part of the infrastructure that allows increasingly large collections of those accelerators to operate as coordinated computing systems.
That distinction could become more important as the industry evolves.
The first AI infrastructure bottleneck was compute.
The next bottlenecks increasingly include networking, memory, power, cooling and physical data-center capacity.
Arista’s opportunity is therefore connected to a broader transition: AI infrastructure is evolving from a collection of individual components into tightly integrated systems in which overall performance depends on every layer.
The current evidence supports the view that networking is becoming strategically more important.
Hyperscaler capital expenditure remains extraordinary. AI cluster sizes continue to expand. Ethernet performance is advancing from 400G through 800G toward 1.6T. Arista is expanding its product architecture specifically to address this market.
Yet the research thesis depends on more than AI growth alone.
Arista must maintain technological relevance, preserve major customer relationships, defend attractive economics against powerful competitors and demonstrate that Ethernet can capture a substantial share of next-generation AI fabrics.
Markets must also distinguish between the quality of the company and the price assigned to that quality.
For FRL, this is the central analytical conclusion:
Arista Networks represents a strategically important way to understand the networking layer of the AI infrastructure cycle. The structural opportunity remains substantial, but the next phase will increasingly be determined by market-share evidence, Ethernet adoption, hyperscaler capital allocation and whether fundamental growth can continue to meet increasingly demanding expectations.
Frequently Asked Questions
What does Arista Networks do in AI infrastructure?
Arista Networks provides high-speed Ethernet networking hardware and software used to connect servers, accelerators and other computing infrastructure inside large data centers. As AI clusters grow, accelerators must exchange increasingly large volumes of data. Arista’s Etherlink and EOS technologies are designed to provide the high-bandwidth, low-latency network fabrics required to operate these systems efficiently.
How does Arista Networks benefit from AI data centers?
AI data centers require considerably more than GPUs. Large accelerator clusters need high-speed switches, network fabrics and sophisticated software to move data efficiently. Growth in AI computing capacity can therefore increase demand for 400G, 800G and eventually 1.6T networking. Arista’s opportunity depends on capturing a meaningful share of this expanding Ethernet infrastructure market.
Does Arista Networks compete with NVIDIA?
Yes, in parts of AI networking, but the relationship is more nuanced than a simple company-versus-company comparison. NVIDIA offers InfiniBand and Spectrum-X networking alongside its accelerators, while Arista focuses heavily on standards-based Ethernet architectures. Both can participate in the expansion of AI infrastructure, and different customers may choose different networking architectures depending on their requirements.
Why is Ethernet important for AI data centers?
Ethernet already forms the networking foundation of much of global cloud infrastructure. Extending Ethernet into AI clusters can provide customers with a broad supplier ecosystem, interoperability and reduced dependence on proprietary architectures. The challenge is achieving the latency, congestion management and reliability required by extremely large accelerator clusters.
What are the main risks facing Arista Networks?
Important risks include concentration among large cloud customers, competition from NVIDIA and Cisco, increasing hyperscaler internal development, rapid technological transitions, potential AI infrastructure overcapacity and high market expectations. The central long-term question is whether Arista can retain a strategically important position as AI networking architectures evolve.
Research Notice
This publication constitutes general financial market research prepared by Final Resurrection Limited for informational and analytical purposes. It is based on publicly available information and independent research assessment. It does not constitute investment advice, financial advice, portfolio management, or a recommendation to buy or sell any financial instrument. Investment decisions remain solely the responsibility of the individual investor.