Arista, Credo, and Lumentum Gain as AI Networking Becomes the Next Bottleneck

Google News put three AI networking stocks in focus as investors shifted attention from processors toward the connections keeping thousands of accelerators productive.

Arista Networks, Credo Technology, and Lumentum occupy different parts of that increasingly important data path. Arista supplies Ethernet systems, Credo specializes in high-speed connectivity, and Lumentum provides optical components.

The conflict is no longer simply Nvidia against other chipmakers. It is Nvidia’s integrated networking stack against specialized suppliers betting that cloud operators want broader choice.

That distinction matters because an expensive accelerator generates little value while waiting for data. The network must move information quickly, predictably, and within strict power limits.

The three companies are not interchangeable investments. Each faces different competitors, customer risks, product cycles, and technical transitions.

Their recent growth nevertheless points toward the same conclusion. AI infrastructure spending is moving beyond GPUs and into the links connecting racks, clusters, and entire data centers.

What Google News Revealed About the AI Networking Trade

The latest Google News attention reflects a real change in infrastructure spending, not simply another collection of AI stock ideas.

Training and serving large models requires many accelerators to operate as one system. Those processors exchange model parameters, intermediate results, and memory traffic throughout each workload.

This pattern creates east-west traffic, meaning data moves heavily between servers rather than primarily entering or leaving the facility. Conventional enterprise networks were not designed for that intensity.

AI clusters also generate synchronized traffic bursts. Thousands of processors can transmit simultaneously, creating congestion that leaves costly computing capacity idle.

Networking has therefore become part of the computing system itself. Switches, cables, optical modules, network interface cards, and control software collectively determine how effectively the cluster operates.

Arista addresses the switching and software layer. Its systems build Ethernet fabrics, which are networks connecting accelerators across racks and data center halls.

Credo works closer to the physical connection. Its active electrical cables contain electronics that condition high-speed signals while using copper for short-distance links.

Lumentum operates in photonics, the technology that generates and manipulates light for data transmission. Its lasers and optical components support longer or more bandwidth-intensive connections.

These niches overlap at their boundaries, but they solve different engineering problems. Copper remains attractive across shorter distances because it can offer lower cost and power consumption.

Optics becomes necessary as distance, bandwidth, and signal integrity requirements rise. Switching platforms coordinate the resulting traffic and prevent one overloaded link from slowing the entire cluster.

The market’s interest follows measurable business growth. Arista reported first-quarter 2026 revenue of $2.709 billion, up 35.1% from the prior year.

Credo reported fiscal fourth-quarter revenue of $437 million, representing 157% year-over-year growth. Full-year revenue more than tripled to $1.335 billion.

Lumentum reported fiscal third-quarter revenue of $808.4 million. That figure increased 90% year over year and reached a company record.

Those results do not establish that every future AI cluster will use these suppliers. They show that connectivity spending is already reaching company income statements.

The trend also reaches beyond one product cycle. New accelerator systems require faster external networks, denser in-rack connections, and more optical capacity between computing buildings.

That expansion explains why the Google News headline resonates. The AI infrastructure story has developed several valuable bottlenecks, rather than one GPU bottleneck.

However, strong demand does not remove investment risk. Specialized companies can grow quickly when their niche expands, then suffer when customers redesign systems or suppliers add capacity.

The more useful question is not whether AI needs networking. It is which architecture captures the spending, and which vendors retain their position during each transition.

Arista AI Networking Turns Ethernet Into a Computing Fabric

Arista’s opportunity comes from making open Ethernet perform reliably under workloads once associated with specialized supercomputer networks.

Arista built its position by supplying switches and software to cloud operators. Its Extensible Operating System gives customers one consistent software foundation across multiple hardware platforms.

That consistency matters in very large environments. Operators need automation, visibility, and repeatable configurations across thousands of devices.

AI changes the performance requirement. Packet loss, congestion, or unpredictable latency can delay synchronized calculations across an entire cluster.

Arista combines high-capacity switches with traffic-management features designed for these conditions. Virtual output queuing separates traffic flows and reduces head-of-line blocking, where one congested destination delays unrelated data.

Large packet buffers absorb temporary traffic bursts. Network telemetry helps operators identify congestion, failing links, and uneven workload distribution before those problems waste substantial computing time.

Arista’s recent numbers support the demand case. Its first-quarter revenue grew 35.1%, while operating cash flow reached $1.69 billion.

The company also introduced its 7060XE7 portfolio in June 2026. These systems support 1.6-terabit connections and target both scale-out and scale-up AI fabrics.

Scale-out networking links multiple servers or racks into a larger cluster. Scale-up networking connects processors more tightly, often within a rack or rack-scale system.

Arista historically held a stronger position in scale-out Ethernet. Moving into scale-up would expand its addressable role inside future AI systems.

The company has also announced high-density liquid-cooled pluggable optics. Arista says the approach can reduce networking racks by up to 75% and floor space by up to 44%.

Those figures are company estimates, not independent performance guarantees. They still reveal the constraints customers are asking suppliers to solve.

Data center operators cannot expand networking indefinitely by adding conventional racks. Every additional component consumes floor space, electricity, cooling capacity, and maintenance time.

Arista’s central argument is that Ethernet can remain the common network standard while becoming suitable for demanding AI workloads. That promises customers more supplier choice and familiar operating practices.

The opposing strategy comes from Nvidia. Its Spectrum-X platform combines Ethernet switches, BlueField data processing units, network adapters, optics, and software.

Nvidia says Spectrum-X Ethernet can provide 1.6 times the networking performance of traditional Ethernet fabrics. That comparison reflects Nvidia’s own testing and configuration assumptions.

The important point is architectural. Nvidia sells an integrated platform optimized around its accelerators, while Arista sells a networking system designed for heterogeneous environments.

Cloud providers have reasons to consider both. An integrated stack can simplify deployment and performance tuning, particularly for customers standardizing on Nvidia hardware.

Open Ethernet can reduce dependency on one supplier. It can also accommodate accelerators, network adapters, and optical products from several vendors.

Arista must prove that openness does not require sacrificing predictable performance. Nvidia must prove that integration provides enough value to justify tighter platform dependence.

Customer concentration adds another pressure point. Large cloud operators can represent substantial portions of a networking supplier’s revenue and purchasing commitments.

A hyperscaler can change suppliers, develop custom equipment, or pause deployment after a major build cycle. Any of those decisions can alter growth quickly.

Arista’s expanding enterprise business provides some diversification. However, enterprise campus networking does not mirror the spending profile or growth rate of hyperscale AI fabrics.

The next test is whether Arista’s 1.6-terabit products move from technical announcements into significant deployments. Product availability alone does not establish lasting market leadership.

Credo Active Electrical Cables Own a Critical Short Link

Credo’s niche sits between passive copper and optics, where short AI connections need more reach without accepting unnecessary optical power consumption.

A passive copper cable carries electrical signals without signal-processing electronics. It works well at short distances, but higher data rates make signal loss increasingly difficult.

An optical cable converts electrical data into light. It supports longer distances and high bandwidth, but introduces additional components, cost, and power demands.

An active electrical cable, or AEC, occupies the middle. Electronics inside the cable restore signal quality, allowing copper to carry high-speed data farther than passive connections.

That middle ground has become valuable inside dense AI racks. Accelerators, switches, and network interface cards need many short, dependable connections.

Credo’s ZeroFlap AEC products target this use. The company says its design improves link stability and reduces cable-related interruptions.

“Link flap” describes a connection that repeatedly moves between working and failed states. Even a brief interruption can disrupt tightly synchronized AI jobs.

The business has expanded rapidly alongside this demand. Credo’s fiscal 2026 results showed quarterly revenue rising 157% year over year.

Full-year revenue reached $1.335 billion, compared with $436.8 million in fiscal 2025. Credo also ended the year with $1.4 billion in cash and short-term investments.

Management forecast first-quarter fiscal 2027 revenue between $465 million and $475 million. That forecast represented continued sequential growth from the fourth quarter.

Credo does more than sell complete cables. Its portfolio includes retimers, digital signal processors, optical components, and connectivity products reaching speeds up to 1.6 terabits.

A retimer receives a degraded electrical signal and transmits a cleaner version. A digital signal processor performs the calculations needed to recover data from imperfect high-speed signals.

These products give Credo exposure to several connection designs. They also place the company against large semiconductor suppliers with greater resources and broader customer relationships.

Broadcom, Marvell, Nvidia, and other networking vendors can address portions of the same data path. Customers can also redesign racks around different cable lengths or optical technologies.

Credo’s acquisition of DustPhotonics shows how it is preparing for that transition. The deal added silicon photonics capabilities, which integrate optical functions with semiconductor manufacturing techniques.

This move reduces the risk of relying entirely on copper. It also introduces integration challenges and places Credo deeper into a competitive optical market.

The copper-versus-optics debate is often presented as a single winner replacing the other. Actual data centers use both, based on distance, bandwidth, reliability, and power requirements.

Credo benefits when engineers optimize each link instead of applying optics everywhere. Its AEC products can preserve copper within connections where passive cables no longer perform adequately.

The risk is that the optimal boundary moves. New packaging, co-packaged optics, or changes in rack architecture can shrink the distances where AECs offer the best tradeoff.

Customer concentration creates another uncertainty. Rapidly growing component suppliers frequently depend on a small number of hyperscale customers during early adoption cycles.

That dependence can amplify growth when one customer expands. It can produce the opposite effect when that customer pauses orders or qualifies another supplier.

Credo’s growth rate therefore deserves context. Revenue expansion confirms strong current demand, but it does not guarantee stable market share across future architectures.

A practical deployment with Rebellions offers one concrete example. The companies integrated ZeroFlap AECs into the RebelPOD enterprise inference platform.

Credo says that design targets faster cluster stabilization and better accelerator utilization. Buyers should treat those statements as supplier claims until independent operating data becomes available.

The larger implication remains important. AI networking performance can depend on a cable that receives little attention beside the processors it connects.

That creates opportunity for specialists. It also means a seemingly small design change can move revenue between copper, optical, and integrated platform suppliers.

Lumentum Optical Networking Rides the Shift From Copper to Light

Lumentum benefits when electrical links reach physical limits and AI data centers require more lasers, optical engines, and switching components.

Copper signals weaken as distance and data rates increase. Higher speeds also make interference and power consumption harder to control.

Optical connections carry data as light, making them suitable for longer distances and dense bandwidth. They connect racks, data halls, campuses, and increasingly shorter links inside AI systems.

Lumentum supplies lasers, optical components, and subsystems used in these networks. It does not sell the complete data center fabric that Arista or Nvidia offers.

That specialization creates direct exposure to rising optical content. It also leaves Lumentum dependent on system makers and module customers that control the final product.

The company’s fiscal third-quarter results illustrate the current demand cycle. Revenue reached $808.4 million, rising 90% from $425.2 million one year earlier.

GAAP gross margin reached 44.2%, while GAAP operating margin reached 21.6%. Both improved significantly from the preceding quarter.

Management attributed part of the improvement to laser-chip strength and a favorable product mix. It also highlighted pump lasers and narrow-linewidth laser assemblies.

These components support scale-across connections, which link multiple computing sites or data center buildings. That category becomes more important when one facility cannot supply enough power or space.

Lumentum is also targeting optical circuit switches. These systems redirect light paths without repeatedly converting signals between optical and electrical formats.

Optical circuit switching can help operators reconfigure large clusters while reducing conversion overhead. It does not replace packet switching across every workload.

The company expects co-packaged optics to become another growth driver. Co-packaged optics places optical components close to the switch silicon instead of relying entirely on removable modules.

Moving optics closer can reduce the electrical distance inside a system. That approach targets power and signal-integrity problems at very high bandwidths.

However, co-packaged optics introduces difficult questions about manufacturing, cooling, repair, and serviceability. Replacing a failed pluggable module is easier than repairing optics integrated near an expensive switch.

That tension gives Lumentum several possible revenue paths. Conventional pluggable modules can keep growing even if co-packaged optics adoption takes longer than suppliers expect.

A faster transition could create demand for new laser and optical-engine designs. It could also shift value toward companies controlling packaging and complete switch systems.

Nvidia’s Rubin roadmap places photonics directly within the platform strategy. Its Rubin networking design includes Spectrum-6 Ethernet and co-packaged optical switch systems.

Nvidia claims these systems provide five times better power efficiency than traditional methods. That remains a vendor comparison tied to its selected design and baseline.

The announcement still confirms the direction of travel. Networking power, reliability, and bandwidth have become first-order concerns in new accelerator platforms.

Lumentum has gained another connection to this roadmap through its work with Nvidia on optics. Partnerships can accelerate demand, but they do not eliminate negotiating or concentration risks.

A customer with significant purchasing power can press suppliers on pricing. It can also qualify alternative components to protect its own supply chain.

Manufacturing capacity poses a separate challenge. Laser and photonic products require specialized fabrication, packaging, testing, and quality control.

Lumentum announced a new U.S. manufacturing facility for advanced lasers in March 2026. Expanding production can support demand, but capacity investments depend on accurate long-range forecasts.

Overbuilding would pressure utilization and margins. Underbuilding could leave orders unfilled while competitors secure customer qualifications.

The next reported milestone is close. Lumentum scheduled its fiscal fourth-quarter results for August 11, 2026.

That report should reveal whether the fiscal third-quarter acceleration continued. Guidance will also show how management views optical demand beyond current customer orders.

Lumentum’s niche therefore has a clear tailwind but no guaranteed outcome. AI systems need more optical connectivity, while product form and supplier share remain unsettled.

Nvidia Pressures Every Independent Networking Supplier

The primary conflict is specialized choice against Nvidia’s integrated system, not Ethernet against Ethernet or copper against optics.

Nvidia no longer approaches networking as an accessory to its GPU business. It sells switches, network adapters, data processing units, cables, optics, software, and rack-scale systems.

Its Mellanox acquisition supplied much of the foundation. Nvidia then expanded InfiniBand while developing Spectrum-X for customers that prefer Ethernet.

This dual approach closes an obvious opening. Independent Ethernet suppliers cannot assume Nvidia will remain confined to a proprietary networking standard.

Nvidia can optimize communication across the accelerator, network interface, switch, and software stack. That coordination can improve deployment speed and predictable workload performance.

It can also simplify accountability. A customer buying one integrated platform has fewer vendors to contact when a cluster fails to meet expectations.

Arista, Credo, and Lumentum offer a different proposition. Each gives customers access to specialized technology without requiring the entire Nvidia networking stack.

The model appeals to cloud operators that design their own infrastructure. These buyers often combine products from several suppliers and maintain internal networking software.

Supplier diversity can improve negotiating leverage and reduce dependence. It can also protect operators from a shortage or delay affecting one vendor.

However, a multi-vendor system requires integration and validation. Every combination of switch, adapter, cable, optical module, and software can introduce another failure point.

The resulting decision is not ideological. Operators will compare total cluster output, power use, reliability, deployment time, and long-term flexibility.

This creates three distinct pressure tests.

Arista must show that its Ethernet fabrics deliver consistent AI workload performance across a broad equipment ecosystem. Strong switching hardware alone is insufficient.

Credo must keep its AEC products valuable as bandwidth rises and rack designs change. It also needs optical products ready when copper stops being optimal.

Lumentum must convert industry optical growth into durable revenue and margins. It cannot assume that every optical architecture assigns the same value to its components.

Nvidia faces its own risks. Customers can resist vendor concentration, particularly when the same supplier controls scarce accelerators and their surrounding network.

Cloud operators also develop custom accelerators. Google, Amazon, Microsoft, and Meta have incentives to use networking systems that support several computing architectures.

The Ultra Ethernet Consortium reflects this demand for broader standards. Its members have worked on Ethernet specifications aimed at AI and high-performance computing workloads.

Standards can expand the overall market for interoperable equipment. They can also make it easier for customers to replace one component supplier with another.

The distinction between open and proprietary systems is not absolute. Nvidia describes Spectrum-X as standards-based Ethernet, while still differentiating through tightly coordinated hardware and software.

Arista also relies on merchant silicon and industry standards while adding proprietary software, systems engineering, and traffic-management capabilities.

Credo and Lumentum similarly participate in standards while defending product-specific intellectual property. Every supplier combines interoperability with differentiation.

Investors should therefore avoid a simple winner-takes-all conclusion. Several architectures can coexist because customers prioritize different workloads and operating models.

A frontier training cluster can justify an integrated system optimized for maximum performance. A cloud platform serving many customers might emphasize flexibility and equipment choice.

An enterprise inference deployment can prioritize easier installation and predictable operating costs. A distributed cluster can require advanced optics between separate facilities.

These scenarios support multiple suppliers, but they do not guarantee equal economics. The strongest position belongs to whoever controls the hardest component to replace.

Today, Nvidia controls the accelerator platform. Arista controls important network operating relationships, Credo has momentum in AECs, and Lumentum supplies difficult photonic components.

Those positions remain contestable. New interconnect standards, custom silicon, silicon photonics, or changes in accelerator packaging can redraw the value chain.

The Google News framing captures the opportunity but can obscure this dependency. A niche looks dominant only while customers continue designing around it.

Three Signals to Watch After the Google News Spotlight

The next three signals will show whether these companies hold durable positions or simply benefit from a strong spending cycle.

The first signal is Arista’s conversion of AI product announcements into recognized revenue. Management previously targeted substantial growth from AI networking during 2026.

Watch customer adoption of the 7060XE7 systems and Arista’s progress in scale-up networking. Success would move the company closer to the accelerator rack.

That outcome would strengthen the case that open Ethernet can compete with Nvidia across more of the cluster. Delays would leave Arista concentrated in familiar scale-out deployments.

Gross margin also deserves attention. Large hyperscale customers can generate significant revenue while producing a less favorable product mix.

Arista’s first-quarter results established a strong growth baseline. Later quarters must show that demand converts without undermining economics.

The second signal is Credo’s customer and product diversification. Its recent revenue growth creates a demanding comparison for future quarters.

Watch whether optical products, retimers, and additional AEC deployments reduce reliance on a narrow customer or product group. Broader adoption would make growth more defensible.

The DustPhotonics integration offers an early indicator. Credo must turn acquired silicon photonics capability into qualified products without distracting from its copper franchise.

A successful transition would let Credo serve customers as link requirements move from electrical to optical. Weak execution would expose the limits of its current niche.

Reliability data also matters. Credo markets ZeroFlap around connection stability, but customers ultimately care about cluster uptime and accelerator utilization.

Additional production deployments would offer stronger evidence than laboratory specifications. Independent benchmarks would make supplier performance claims easier to compare.

The third signal is Lumentum’s August 11 earnings report and subsequent outlook. Current results show rapid optical demand and significant margin expansion.

Watch growth in cloud and AI products, manufacturing capacity commentary, and progress in co-packaged optics or optical circuit switches.

Continued revenue growth with stable margins would support the argument that Lumentum holds valuable photonics capacity. Weaker guidance would suggest ordering or product-mix volatility.

The timing of co-packaged optics remains especially important. A gradual transition preserves demand for pluggable products while giving suppliers time to improve manufacturing.

A rapid transition can expand component value but increase execution risk. Delays can postpone expected revenue while leaving recently added capacity underused.

Readers should also track Nvidia’s second-half Rubin deployments. Those systems will test whether its integrated networking approach gains broader adoption among major cloud providers.

If customers standardize heavily around Spectrum-X and Nvidia photonics, independent suppliers face greater pressure. Mixed deployments would support a more diverse networking market.

None of these signals turns the three companies into automatic stock recommendations. Revenue growth can coexist with high expectations, customer concentration, and volatile capital spending.

The more durable takeaway concerns infrastructure architecture. AI performance increasingly depends on communication between processors, not only the processors themselves.

That shift gives switches, active cables, lasers, and optical systems a larger role in determining cluster economics. It also makes networking failures more expensive.

Google News has helped move that layer into public view. The next earnings reports and deployments will determine whether Arista, Credo, and Lumentum keep their specialized advantages.

For developers and enterprise buyers, the practical question is straightforward. Does the chosen network keep accelerators working while preserving flexibility, reliability, and manageable power use?

For investors, the standard should be equally demanding. Follow verified revenue, customer diversification, manufacturing execution, and production deployments rather than treating every AI networking claim as equivalent.