Cisco: Can Networking Become the Hidden Winner of the AI Infrastructure Boom?
Company Deep Dive · The Executive Intelligence Series | September 19, 2026
Cisco has been the default answer to enterprise networking for three decades. The AI era is the first time that position has been both an asset and a liability. This deep dive examines whether networking can become the hidden winner of the AI infrastructure boom.
AI computing is not just about GPUs. Every GPU in a training cluster has to communicate with every other GPU, often thousands of them, at speeds that would have been unthinkable a decade ago. The network is not a supporting character in AI infrastructure. It is the circulatory system.
AI back-end switch sales surpassed front-end switch sales for the first time in the second quarter of 2026 — a major inflection point in how networking spending is allocated. Cisco's AI infrastructure orders grew 4.5x in a single fiscal year. The question is whether a company built for enterprise networking can capture value in a market increasingly defined by hyperscale AI clusters.
The Startup: From Plumbing to Platform
Cisco Systems has been the default answer to enterprise networking for three decades. Routers, switches, wireless controllers, firewalls — if data moved inside a company, it probably moved through Cisco gear. That dominance made Cisco one of the most valuable technology companies of the dot-com era, and it made the company complacent for nearly a decade afterward.
The problem was structural. Cisco's growth had stalled in the mid-2010s as cloud computing shifted networking spend away from enterprise campuses and into hyperscale data centers. Those data centers — operated by Amazon, Google, Microsoft, and Meta — bought networking differently. They wanted open standards, disaggregated hardware and software, and suppliers who would compete on price and speed rather than bundling. Cisco's proprietary model, which had served it well in the enterprise, was a liability in the cloud.
Starting around 2018, Cisco began a slow transformation. It separated hardware from software, embraced subscription revenue, and repositioned itself as a platform company rather than a box company. The acquisition of Splunk for $28 billion in March 2024 was the clearest signal that Cisco intended to compete on software, data, and security rather than just networking hardware. The acquisition of Galileo Technologies in April 2026, folded into the Splunk Observability portfolio, extended that strategy into AI agent observability.
But the transformation that matters most for this analysis is not software. It is the AI networking opportunity — and whether Cisco can convert a legacy position in enterprise networking into a durable position in the infrastructure layer of the AI stack.
The CODEW Lens: Cisco spent two decades selling the safest networking decision. In the AI era, the safest networking decision may no longer be Cisco. The company's transformation is a bet that it can change that.
The AI Networking Problem
There are two distinct networking challenges in AI data centers:
Scale-out networking connects GPUs within a data center, forming the back-end fabric that moves gradients, activations, and model weights between accelerators during training and inference. This is the fastest-growing segment of the networking market. In the first quarter of 2026, Ethernet switch sales in AI back-end networks more than doubled and accounted for about two-thirds of data center switch sales in AI clusters.
Scale-across networking connects data centers to each other, linking geographically distributed AI clusters over long distances. As power and space constraints force AI infrastructure to spread across multiple sites, scale-across networking is becoming a distinct opportunity.
The AI networking market has become the most dynamic segment of the data center switch market. AI back-end switch sales surpassed front-end switch sales for the first time in the second quarter of 2026, a major inflection point in how networking spending is allocated. Ethernet has overtaken InfiniBand in AI back-end networking and now accounts for roughly two-thirds of data center switch sales in AI clusters.
The CODEW Lens: The network is the bottleneck nobody talks about until it becomes the bottleneck everybody talks about. AI clusters are only as fast as the fabric that connects them.
The Technology: Silicon One and Nexus One
Cisco's AI networking strategy rests on two pillars: custom silicon and an open systems platform.
Silicon One — Cisco introduced the Silicon One family of networking chips in 2019. At the time, the company was moving away from proprietary ASICs and toward a unified silicon architecture that could serve multiple markets — service provider routing, data center switching, and enterprise — from a single design lineage. That bet is now paying off in the AI era.
The G300 — announced in February 2026, the G300 delivers 102.4 terabits per second of switching capacity. It powers Cisco's N9000 and 8000 series switches, which support clusters of more than one million GPUs. The G300 includes microsecond-level telemetry, intelligent load balancing, and the industry's largest packet buffer for predictable AI performance.
Nexus One — Cisco's unified AI networking platform brings silicon, systems, optics, software, and a single operating model into one integrated solution. The platform supports multiple silicon options — including Cisco Silicon One and NVIDIA Spectrum-X Ethernet switch silicon — across N9000 switches and optics.
This is a deliberate strategic choice. Cisco is not trying to force hyperscalers to buy Cisco silicon exclusively. It is offering a platform that can integrate with whatever silicon the customer prefers, while still capturing value through systems, optics, software, and management.
Acacia Optics — Cisco acquired Acacia Communications in 2021 for its coherent optical technology. That acquisition is now producing significant returns. Cisco's CEO said Acacia was set to grow 200% in fiscal 2026. By the third quarter of FY26, Acacia had shipped over 750,000 400G and 40,000 800G coherent pluggable optics.
The CODEW Lens: The G300 is not a faster version of a previous chip. It is a bet that AI workloads require a different kind of switch — one with telemetry, load balancing, and power efficiency built in from the start.
The AI Infrastructure Opportunity
Cisco's AI order book has grown from essentially zero two years ago to a scale that now materially affects the company's financial outlook.
AI Infrastructure Snapshot · FY 2026
Hyperscaler AI orders: $9.3B (up ~4.5x from FY25)
Q4 alone: $4B
FY26 AI revenue delivered: ~$4B
FY27 AI revenue guide: $7.5B
Neocloud/sovereign/enterprise AI orders: $400M+ in Q4
The mix of orders tells an important story. Approximately 60% of hyperscaler AI orders were Silicon One-based systems, and 40% were optics, showing that Cisco is selling complete networking systems rather than just components.
Beyond hyperscalers, Cisco is seeing AI demand broaden. In the fourth quarter of fiscal 2026, Cisco recorded more than $400 million in AI infrastructure orders from neoclouds, sovereign cloud operators, and enterprise customers, bringing that annual total above $1 billion. Enterprise Nexus switch orders tagged for AI deployments rose more than 85% sequentially.
Cisco's campus and branch networking business, which historically grew at roughly 3% to 4%, has expanded about 20% for several quarters. The company attributes that growth to infrastructure refresh cycles, demand for lower-latency networking, and heightened security concerns around aging equipment. Cisco has identified more than $100 billion in upgrade and refresh opportunity involving its own installed base over the next several years.
The CODEW Lens: The AI order book is not just a hyperscaler story. The enterprise opportunity is still early, and it may end up being the more durable half of Cisco's AI business.
Ethernet vs. InfiniBand
The AI networking market is defined by an architectural battle that has significant implications for Cisco.
InfiniBand, developed by Mellanox (now part of NVIDIA), was traditionally the interconnect of choice for high-performance computing and AI training. It offered lower latency and higher bandwidth than Ethernet, and it was purpose-built for the demanding communication patterns of distributed training. Two years ago, InfiniBand accounted for nearly 80% of data center switch sales in AI back-end networks.
That has changed. Ethernet has overtaken InfiniBand in AI back-end networking and now accounts for roughly two-thirds of data center switch sales in AI clusters.
The reason is economic. InfiniBand imposes significant infrastructure and operational costs. It often requires organizations to maintain a separate networking infrastructure distinct from their standard Ethernet-based systems. As AI clusters scale and as inference workloads grow relative to training, the cost and flexibility advantages of Ethernet become more important.
This shift benefits Cisco. Cisco is an Ethernet networking company. Its entire portfolio — Silicon One, Nexus, Acacia optics — is built around Ethernet. The company is a key steering member of the Ultra Ethernet Consortium, which is developing the next generation of Ethernet standards for AI workloads.
The CODEW Lens: InfiniBand was designed for supercomputers. AI data centers are not supercomputers. They are factories, and factories run on Ethernet.
The Business Model: Networking, Security, Observability
Cisco is not just selling switches into AI data centers. It is selling an integrated platform that spans networking, security, and observability.
| Segment | Q4 FY26 Revenue | YoY Growth | Signal |
|---|---|---|---|
| Networking | $9.79B | +28% | 8th consecutive quarter of double-digit order growth |
| Security | $2.23B | +14% | Firewall revenue +30% for two consecutive quarters |
| Observability | $275M | +6% | Splunk + Galileo extend into AI agent observability |
The strategic logic is that observability is becoming inseparable from networking and security. As AI agents operate across distributed infrastructure, enterprises need visibility into network performance, security posture, and AI behavior from a single platform.
The CODEW Lens: Cisco is not selling switches. It is selling a platform that manages, secures, and observes the infrastructure that runs AI. The switch is the entry point. The platform is the business.
Funding and Capital: The Splunk Bet
Cisco's $28 billion acquisition of Splunk, completed in March 2024, was the largest in the company's history. It was also a bet that software and data would matter more than hardware in the next phase of enterprise technology.
That bet is now being tested in the AI era. Splunk gives Cisco a platform for collecting, analyzing, and acting on machine data — the telemetry generated by networks, security systems, and AI workloads. The Galileo acquisition extends that platform into AI agent observability, where the ability to monitor and evaluate AI behavior is becoming a requirement rather than a nice-to-have.
Cisco's financial position supports continued investment. In fiscal 2026, the company reported record revenue of $63.3 billion, up 12% year over year, and a record non-GAAP operating margin of 34.8%. Cisco achieved its highest productivity metrics in 30 years, measured by revenue, non-GAAP operating margin, and earnings per employee.
For fiscal 2027, Cisco guided revenue to $72.2 billion to $73.4 billion, representing approximately 15% growth at the midpoint — a second straight year of accelerating growth for a company that grew 5% in fiscal 2025.
The CODEW Lens: Splunk was expensive, and the integration is not complete. But it gives Cisco something Arista does not have: a software and data platform that can capture value beyond the switch.
Competitive Landscape
Cisco competes in AI networking against a diverse set of players:
| Competitor | Position | Cisco's Response |
|---|---|---|
| Arista Networks | Largest data center Ethernet switch vendor; strong in hyperscaler AI fabrics | Nexus One open platform; Silicon One silicon; incumbency in enterprise |
| NVIDIA | Dominant in AI compute; owns InfiniBand; Spectrum-X Ethernet silicon | Supports NVIDIA silicon in Nexus switches; partners on reference architectures |
| Celestica | Contract manufacturer turned AI switch leader | Competes on systems integration, software, and optics |
| HPE (Juniper) | Expanded portfolio through Juniper acquisition | Cisco's enterprise incumbency and security integration |
| Broadcom | Merchant silicon supplier to many switch vendors | Cisco designs its own Silicon One silicon; reduced dependency |
In the broader data center switching market, Arista leads with approximately 21.1% share, followed by Cisco at 17.8% and Celestica at 13.2%. In AI back-end Ethernet switching, Celestica leads, followed by NVIDIA, with Arista third and Cisco fourth — but Cisco recorded the largest share gain in the first quarter of 2026.
The CODEW Lens: Cisco both competes with and partners with NVIDIA. That co-opetition reflects the reality that no single vendor owns the AI networking stack — and Cisco is betting it can win by being the vendor that connects everyone else.
Risks
Gross margin pressure. Cisco's fourth-quarter gross margin fell 210 basis points on a heavier hardware mix and rising memory costs. Management expects a slight gross-margin headwind through fiscal 2027. AI infrastructure revenue carries lower margins than Cisco's traditional software and services business.
Competitive intensity. Arista remains the leader in data center Ethernet switching, and its EOS operating system has a loyal following. Celestica and NVIDIA have gained share in AI back-end networks. Cisco's share gains are encouraging but do not yet reverse Arista's leadership position.
Customer concentration. Cisco's AI order growth is heavily dependent on a small number of hyperscaler customers. If hyperscaler capital spending slows or shifts to competitors, Cisco's AI revenue trajectory could change quickly.
Security vulnerabilities. In September 2026, Cisco disclosed a critical vulnerability (CVE-2026-20212, CVSS 9.8) affecting Nexus 9000 switches equipped with Silicon One ASICs. The vulnerability could allow an unauthenticated remote attacker to execute code with root privileges. Cisco's security strategy depends on its ability to defend its own infrastructure as well as its customers'.
Integration risk. The Splunk and Galileo acquisitions represent a significant expansion into software and observability. Integrating these capabilities with Cisco's networking portfolio — and making them feel native rather than bolted on — is a multi-year execution challenge.
The CODEW Lens: The biggest risk is not that Cisco loses to Arista. It is that the AI networking opportunity turns out to be smaller and more concentrated than the current order book implies.
What's Next
Cisco's AI networking trajectory depends on several variables:
Scale-across networking. As AI clusters grow beyond the capacity of a single data center, the opportunity to connect distributed clusters becomes significant. Acacia's coherent optics position Cisco well for this transition.
Enterprise AI adoption. Cisco's enterprise AI orders are growing from a small base. If enterprises deploy private AI infrastructure at scale — rather than relying exclusively on public cloud — Cisco's incumbency in enterprise networking becomes a significant advantage.
Agentic AI bandwidth. Cisco executives cite OpenRouter data showing that AI agents consume about 60% of total inference capacity and their token consumption has increased 14-fold since February. An agent uses about 450% more bandwidth than a human performing the same task. If agentic AI adoption accelerates, network bandwidth demand will grow faster than most current models assume.
Silicon One proliferation. Cisco plans to roll out Silicon One comprehensively across its high-performance networking systems by fiscal 2029. Broader Silicon One deployment would reduce Cisco's dependency on merchant silicon and improve margins over time.
The CODEW Lens: The enterprise AI networking opportunity is still early. If Cisco can hold its position with hyperscalers while building the enterprise business, the next phase of growth could be broader and more durable than the current order book suggests.
The CODEW Take: Can Cisco Turn the AI Networking Boom into a New Growth Engine?
The short answer is yes, but the growth engine will look different from anything Cisco has built before.
The evidence is compelling. Cisco's AI infrastructure orders grew 4.5 times in a single fiscal year. The company is guiding to $7.5 billion in hyperscaler AI revenue in fiscal 2027, up from $4 billion. Networking orders have grown at double-digit rates for eight consecutive quarters. And the architectural shift toward Ethernet — which Cisco helped pioneer and still leads in enterprise deployments — is working in Cisco's favor.
But the more interesting story is not the headline numbers. It is the strategic repositioning.
Cisco is no longer trying to sell proprietary networking to hyperscalers who do not want it. It is selling an open platform — Nexus One — that integrates with whatever silicon, operating system, and accelerator the customer chooses. It is selling optics through Acacia that are essential for connecting distributed AI clusters. And it is selling observability and security through Splunk and Galileo that address the operational complexity of AI infrastructure.
The Splunk acquisition was expensive,ensive and the integration is not complete. But it gives Cisco something that Arista does not have: a software and data platform that can capture value beyond the switch. And it gives Cisco a way to compete not just on bandwidth and latency, but on the operational layer where AI workloads are managed, secured, and observed.
The risks are real. Gross margins are under pressure. Arista remains a formidable competitor. Hyperscaler concentration is a vulnerability. And the competitive landscape in AI networking is evolving faster than any single vendor can control.
But Cisco has something that most AI infrastructure companies lack: incumbency. The company's installed base spans thousands of enterprises, service providers, and government agencies. It has identified more than $100 billion in upgrade and refresh opportunity within that base. If enterprises deploy private AI infrastructure — which many will, for data sovereignty, security, and cost reasons — Cisco's position in enterprise networking gives it a distribution advantage that no startup can match.
The final consideration is timing. Cisco's AI orders today are dominated by hyperscalers. The enterprise AI networking opportunity is still early. If Cisco can hold its position with hyperscalers while building the enterprise AI business, the next phase of growth could be broader and more durable than the current order book suggests.
The verdict: Cisco will not be the largest AI infrastructure company. NVIDIA owns compute. Amazon, Google, and Microsoft own the cloud. But Cisco may be the company that connects them all — the networking layer that moves data between GPUs, between data centers, and between AI agents and the systems they operate. That is not the same as owning the stack. It is something more valuable: owning the layer that every other layer depends on.
The CODEW Lens: NVIDIA = compute. Amazon/AWS = cloud infrastructure. Cisco = networking and connectivity. The AI stack is being built by companies that own their layer — and Cisco is positioning itself as the layer every other layer depends on.
The AI Networking Glossary
Scale-Out Networking — Connects GPUs within a single data center, forming the back-end fabric for distributed training and inference.
Scale-Across Networking — Connects data centers to each other over long distances, linking geographically distributed AI clusters.
InfiniBand — A high-performance interconnect developed by Mellanox (now NVIDIA), traditionally used for supercomputing and AI training.
Ethernet — The dominant networking standard for data centers. Ethernet has overtaken InfiniBand in AI back-end networking.
Silicon One — Cisco's family of networking chips, including the G300, designed for service provider, data center, and enterprise markets.
Nexus One — Cisco's unified AI networking platform, supporting multiple silicon options and operating systems.
Coherent Pluggable Optics — Optical transceivers used to connect network switches over distances of tens or hundreds of kilometers.
AI Back-End Network — The network that connects GPUs within an AI cluster. Distinct from the front-end network that connects the cluster to external systems.
Acquihire — An acquisition structured primarily to secure engineering talent, often with a separate technology license.
Observability — The ability to monitor and understand the state of a system based on its outputs. In AI, observability extends to monitoring AI agent behavior.
FAQ
Q: Why is Ethernet overtaking InfiniBand in AI networking?
InfiniBand imposes significant infrastructure and operational costs. It often requires organizations to maintain a separate networking infrastructure distinct from their standard Ethernet-based systems. As AI clusters scale and inference workloads grow relative to training, the cost and flexibility advantages of Ethernet become more important. Ethernet now accounts for roughly two-thirds of data center switch sales in AI clusters.
Q: How does Cisco compete with NVIDIA in AI networking?
Cisco both competes with and partners with NVIDIA. Cisco's N9000 switches support NVIDIA Spectrum-X Ethernet silicon, and Cisco offers NVIDIA Cloud Partner reference architecture-compliant designs. Cisco competes on systems integration, software, optics, and its open Nexus One platform — not on trying to force customers to buy Cisco silicon exclusively.
Q: What is the Silicon One G300?
The Silicon One G300 is Cisco's flagship networking chip, announced in February 2026. It delivers 102.4 terabits per second of switching capacity and powers Cisco's N9000 and 8000 series switches, which support clusters of more than one million GPUs. The G300 includes microsecond-level telemetry, intelligent load balancing, and the industry's largest packet buffer for predictable AI performance.
Q: Why did Cisco acquire Splunk?
Cisco acquired Splunk for $28 billion in March 2024 — its largest acquisition ever. The deal was a bet that software and data would matter more than hardware in the next phase of enterprise technology. Splunk gives Cisco a platform for collecting, analyzing, and acting on machine data, which is increasingly critical as AI workloads generate enormous telemetry across distributed infrastructure.
Q: What does the Nexus 9000 vulnerability mean for Cisco's AI strategy?
In September 2026, Cisco disclosed a critical vulnerability (CVE-2026-20212, CVSS 9.8) affecting Nexus 9000 switches with Silicon One ASICs. The vulnerability could allow an unauthenticated remote attacker to execute code with root privileges. Cisco released fixes and workarounds. The episode underscores that security is inseparable from networking strategy — and that Cisco's ability to defend its own infrastructure is part of its value proposition to customers.
Connected Resources
This Company Deep Dive connects to The CODEW Company Deep Dive Series, The Great AI Acquisition Wave, Why Nvidia Buys Rather Than Builds, Startup Spotlight, AI Infrastructure, Networking Watch, The CODEW Intelligence, and The Term Sheet.
The CODEW Stat
$9.3B orders · 4.5x growth · 102.4 Tbps Cisco took $9.3 billion in hyperscaler AI infrastructure orders in fiscal 2026 — a 4.5x increase over the prior year — and is guiding to $7.5 billion in AI revenue in fiscal 2027. Its Silicon One G300 chip delivers 102.4 Tbps of switching capacity. The company that built the enterprise network is now building the network that connects the AI economy. Cisco will not own the AI stack. It may own the layer every other layer depends on.
Reviewed by Erwin Castro
on
Saturday, September 19, 2026
Rating:
