AI infrastructure is no longer only a contest over accelerators. Faster clusters need switching, optical links and data-center interconnects that can keep expensive compute working instead of waiting on the network. The commercial evidence is arriving, but it is not reported on a comparable basis.
Research current as of September 4, 2026.
Featured image: Photo by Jordan Harrison on Unsplash.
AI networking demand is showing up in four different places: Broadcom’s combined accelerator-and-networking pool, Arista’s Ethernet systems, Ciena’s optical transport and Marvell’s data-center silicon. The growth is measurable. A clean league table is not. Each company draws the boundary around “AI exposure” differently, and customer concentration is high enough to matter.
The useful investor question is not whether AI needs more bandwidth. It does. The harder question is where that need becomes reported revenue, and how much of the reported number belongs to networking rather than adjacent products.
Broadcom’s latest quarter put the issue in sharp relief. The company reported $16.7 billion of AI semiconductor revenue, up 221% from a year earlier. That figure includes custom AI accelerators and networking. It is large, current and commercially meaningful. It is also too broad to treat as pure networking revenue.
Ciena sits at the opposite end of the disclosure spectrum. Optical networking produced $1.19 billion in its fiscal third quarter, 71.3% of total revenue. Management attributes the current investment wave to AI, but the financial table does not label each optical dollar as AI-related. The category is clean. The demand attribution is not.
A network bottleneck is not one product category
Modern AI clusters create several connectivity problems at once. Scale-up links connect accelerators inside tightly coupled systems. Scale-out fabrics connect servers and racks across a data center. Optical transport carries traffic between buildings, campuses and regions. Custom silicon sits inside switches, transceivers and compute systems.
Those layers can benefit from the same capital-spending cycle without sharing the same economics. A merchant switch-silicon vendor, a systems company and an optical transport supplier face different product cycles, margins and customer dependencies. Adding their “AI” numbers together would produce a confident-looking total with little analytical value.
The central measurement problem: the companies with the largest AI claims often have the broadest definitions. The companies with the cleanest product categories often cannot isolate how much demand is specifically caused by AI.
Four companies, four different kinds of evidence
Broadcom: the largest disclosed pool is also the most bundled
Broadcom reported fiscal third-quarter revenue of $29.59 billion. Semiconductor solutions contributed $20.84 billion, or 70% of the total. Within that business, management said AI semiconductor revenue reached $16.7 billion and forecast $21.7 billion for the fourth quarter.
The number deserves weight. It is recognized revenue, not a backlog target or a total-addressable-market slide. Yet the company explicitly describes the driver as custom AI accelerators and networking. That makes Broadcom the largest exposure in this comparison, but not the cleanest way to isolate network spending.
What remains mixed: the split between custom accelerators, switch silicon, networking components and other AI semiconductor products.
Concentration adds another layer. Broadcom’s fiscal 2025 annual filing said its top five end customers represented about 40% of revenue. The AI ramp can be enormous and still depend on a small set of buyers, deployment schedules and supply commitments.
Arista Networks: strong economics, incomplete AI attribution
Arista reported $3.04 billion of second-quarter revenue, up 37.7% from a year earlier. Product revenue was $2.61 billion, and the company guided to about $3.3 billion for the third quarter. Its non-GAAP operating margin reached 49.9% in the reported quarter.
The operating evidence is unusually strong for a networking systems company. Arista has also introduced 1.6-terabit AI fabric platforms for scale-up, scale-out and scale-across networks. Those launches show product readiness. They do not tell investors how much of current revenue came from those new systems.
The 2025 annual filing helps frame the customer base. Cloud and AI Titans represented about 48% of revenue, AI and Specialty Providers 20%, and Enterprise 32%. Core products, which include AI, cloud and data-center networking, represented 65%. Two customers accounted for 26% and 16% of annual revenue.
What remains unreported: a standalone AI Ethernet revenue figure.
Arista has the best margin profile in this group, but “AI customer” and “AI networking revenue” are not interchangeable. That distinction should survive even when the operating trend is excellent.
Ciena: the cleanest optical category, with a concentration warning
Ciena’s fiscal third-quarter revenue rose 37% to $1.67 billion. Optical networking contributed $1.19 billion, up from $815.5 million a year earlier. The company raised full-year revenue guidance to $6.42 billion at the midpoint and guided to a 45% adjusted gross margin for the fourth quarter.
This is the most direct networking exposure in the comparison. Optical transport is the business, not an optional feature attached to a larger software or compute portfolio. Still, Ciena does not divide optical sales into AI data-center interconnect, telecom modernization and other capacity demand.
The customer base is concentrated. Two customers each exceeded 10% of revenue and together represented 41.7% of the quarter. A few large deployment programs can therefore make reported growth look smoother, or more durable, than the underlying order cadence.
What remains inferential: the exact share attributable to AI workloads rather than broader network upgrades.
Ciena is the cleanest instrument for testing whether optical demand is arriving. It is not a pure read-through on AI alone.
Marvell: the best end-market split, not a pure connectivity number
Marvell reported $2.74 billion of fiscal second-quarter revenue, up 37% year over year. Data-center revenue was $2.17 billion, or 79% of the total, and grew 46%. Management cited strong connectivity demand and expects its custom business to accelerate in the second half of fiscal 2027.
The end-market split is useful because it shows where current revenue sits. The limitation is equally important. Marvell defines the data-center market broadly. It includes AI systems, Ethernet switching, storage, servers and data-center interconnect. Connectivity is a major driver, but $2.17 billion is not a pure AI-networking figure.
Marvell’s fiscal 2026 filing also showed that its ten largest customers represented 82% of annual revenue. Design wins can create long-lived economics, but they can also make growth dependent on a small number of programs.
What remains projected: the expected second-half acceleration in the custom business.
Exposure comparison
| Company | Primary layer | Strongest current evidence | Key measurement gap | Concentration signal |
|---|---|---|---|---|
| Broadcom AVGO |
Custom accelerators and networking silicon | $16.7B Q3 AI semiconductor revenue | Networking is not separated from accelerators | Top five end customers were about 40% of FY2025 revenue |
| Arista ANET |
Ethernet switching systems and software | $3.04B Q2 revenue, up 37.7% | No standalone AI Ethernet revenue | Two FY2025 customers were 26% and 16% of revenue |
| Ciena CIEN |
Optical networking and data-center interconnect | $1.19B Q3 optical revenue | AI-driven demand is not isolated | Two Q3 customers were 41.7% of revenue |
| Marvell MRVL |
Connectivity and custom data-center silicon | $2.17B Q2 data-center revenue | End market includes compute, storage and servers | Top ten customers were 82% of FY2026 revenue |
What the latest numbers prove
- Bandwidth spending is already commercial. These are recognized revenue figures from current quarters, not only product announcements.
- The expansion reaches several layers. Switching systems, optical transport and data-center silicon are all growing.
- Customer concentration is structural. Hyperscale AI infrastructure is bought by a small number of companies, so supplier growth can remain lumpy.
- Definitions matter more as the numbers get larger. Broad AI labels can include compute, networking, storage and software in the same bucket.
What the numbers do not prove
- They do not identify a single winner. The companies participate in different layers and can grow at the same time.
- They do not make every dollar incremental. Some spending replaces conventional cloud or telecom investment rather than adding to it.
- They do not guarantee stable margins. Product mix, component supply, customer discounts and acceptance timing can change profitability.
- They do not remove forecasting risk. Management guidance depends on deployment schedules, design wins and a small set of large customers.
Three checks for the next earnings cycle
First, watch whether disclosure becomes more precise. A separate AI networking or AI Ethernet revenue line would improve comparability. Until then, category definitions should remain part of the analysis.
Second, compare revenue growth with concentration. Growth that comes with a rising share from one or two customers may still be attractive commercially, but it is less diversified.
Third, track margins and acceptance timing. Faster networking products can carry strong economics, but early deployments, supply constraints and large-customer pricing can change that result.
MarketInsiderLab verdict
AI networking has moved beyond the presentation stage. Revenue is accelerating across switch systems, optical transport and connectivity silicon. That is the durable part of the story.
The rankings depend on the question. Broadcom has the largest disclosed AI revenue pool, but it bundles networking with custom accelerators. Arista combines strong growth with exceptional operating margins, yet does not isolate AI Ethernet revenue. Ciena provides the cleanest optical-networking line item. Marvell gives the clearest data-center end-market split while mixing connectivity with compute, storage and server exposure.
The sharp conclusion is simple: the network buildout is real, but the reported numbers are not interchangeable. The next step is better disclosure, not a bigger basket.
Continue with the Market Radar archive, the related analysis of AI data-center power and grid exposure, the Technology & Semiconductor sector page and the Broadcom stock page.
Primary sources
- Broadcom fiscal Q3 2026 results, September 2, 2026.
- Broadcom fiscal 2025 Form 10-K.
- Arista Networks Q2 2026 results, August 4, 2026.
- Arista Networks 2025 Form 10-K.
- Ciena fiscal Q3 2026 results, September 3, 2026.
- Marvell fiscal Q2 2027 results, August 27, 2026.
- Marvell fiscal 2026 Form 10-K.
Editorial disclosure: This article separates reported revenue, product announcements, management guidance and MarketInsiderLab analysis. References to AI demand reflect company disclosures unless stated otherwise.
Investment disclosure: This material is for general informational and educational purposes. It is not personalized investment advice, a recommendation to buy or sell any security, or a promise of future performance. Public-market investments can lose value. Verify current filings and consider your own objectives and risk tolerance.