Broadcom, Arista, and Vertiv Target Different AI Infrastructure Bottlenecks Beyond NVIDIA

Google News surfaced three companies tied to the AI buildout beyond NVIDIA, but the list is not really about finding another GPU winner. Broadcom, Arista Networks, and Vertiv occupy three separate pressure points around AI processors. One supplies custom accelerators and networking silicon, another builds Ethernet systems, and the third handles power and cooling.

The distinction matters because NVIDIA remains central to AI computing. Its latest quarter produced $89 billion in data center revenue, up 117% from one year earlier. The company is expanding from GPUs into networking, CPUs, software, and complete rack-scale systems.

That expansion creates the central tension behind the three-stock thesis. Broadcom, Arista, and Vertiv benefit when customers install more accelerated computing, yet NVIDIA increasingly sells more of the surrounding system. These companies are complementary suppliers today, but parts of their opportunity sit close to NVIDIA’s expanding product boundary.

What the Google News Story Actually Changed

The new signal is not that three companies suddenly entered AI infrastructure. It is that investors are separating the buildout into distinct bottlenecks.

The three-stock thesis appeared on MarketBeat on August 30, 2026. Google News then distributed the story through its AI chips and data centers feed. The companies were Broadcom, Arista Networks, and Vertiv.

The selection divides an AI data center into three supporting layers. Broadcom addresses custom computing and chip-level connectivity. Arista supplies network systems that move data across clusters. Vertiv provides electrical and thermal systems that keep those clusters operating.

None is a direct replacement for NVIDIA. That point prevents the headline from turning into a misleading comparison between unrelated businesses.

NVIDIA supplies the processors, networking products, software, and system designs used across many AI installations. Broadcom also designs custom accelerators, commonly called ASICs. An ASIC is a processor built for a specific customer or workload instead of broad general use.

Arista concentrates on Ethernet fabrics. A fabric is the network of switches and links that lets thousands of processors exchange data as one computing system.

Vertiv works farther down the physical stack. Its equipment distributes power, manages heat, and protects data center operations when workloads become unusually dense.

The common factor is hyperscaler capital expenditure. Hyperscalers are major cloud operators that build computing capacity at enormous scale. Their infrastructure budgets pay for much more than GPUs.

Every accelerator requires memory, network bandwidth, power conversion, backup systems, and cooling. Additional processors can increase demand across those categories, although spending does not reach every supplier equally.

That is why the original story is more useful as a map than as a stock list. It identifies three constraints that determine whether purchased computing capacity becomes usable capacity.

Google News readers should also distinguish discovery from verification. An aggregator can surface a timely argument, but it does not validate the financial claims inside that argument.

For this analysis, recent company disclosures provide the relevant operating evidence. They show that all three businesses are growing, while also exposing different uncertainties around customers, competition, timing, and expectations.

The timing is important. NVIDIA reported its latest results on August 26, only four days before the MarketBeat story appeared. Its data center results reinforced the scale of spending behind the broader infrastructure chain.

NVIDIA recorded $96.2 billion in total quarterly revenue. Data center revenue reached $89 billion, increasing 18% sequentially and 117% year over year.

Those figures do not prove that every adjacent supplier will grow at the same rate. They do show why investors are examining the equipment surrounding each new accelerator deployment.

The changed question is therefore narrower than the headline suggests. Instead of asking which company replaces NVIDIA, readers are asking which supporting constraint receives the next infrastructure dollar.

Broadcom Extends the AI Trade Into Custom Silicon

Broadcom offers the clearest semiconductor extension beyond NVIDIA, but its opportunity depends heavily on a concentrated group of large customers.

Broadcom participates in AI infrastructure through custom accelerators and networking components. These two businesses place it close to the computing layer while avoiding a simple GPU comparison.

Custom accelerators give large cloud operators more control over workloads, power use, system design, and supply. A customer can optimize a processor for selected training or inference tasks rather than buying only general-purpose accelerators.

Inference is the process of running a trained model to generate an answer or prediction. Its economics become increasingly important as AI services move from experiments into daily production.

Broadcom also supplies networking silicon used inside switches and connections between computing devices. Those products help data move through a cluster without leaving expensive processors idle.

The company’s latest reported numbers support the infrastructure thesis. Broadcom generated $22.187 billion in fiscal second-quarter revenue, an increase of 48% from the prior-year period.

Its semiconductor revenue attributed to AI reached $10.8 billion. Management said that figure grew 143% year over year, driven by custom accelerators and AI networking.

Broadcom also projected $16 billion in AI semiconductor revenue for its third quarter. That projection represented expected growth above 200% from one year earlier, according to its quarterly results.

These are company-reported figures and forecasts. They show significant momentum, but the forecast still requires confirmation when Broadcom reports its next results.

The scheduled report on September 2 gives this story an unusually near-term test. Investors will be able to compare the company’s earlier guidance with reported demand, shipments, and its updated outlook.

Broadcom’s role also creates a more complicated relationship with NVIDIA. Custom accelerators can displace some general-purpose GPU demand within selected workloads. Yet the same AI clusters can still use NVIDIA systems elsewhere.

Broadcom is therefore complementary at the industry level and competitive within particular deployments. The boundary changes according to each customer’s workload, technical resources, and preferred architecture.

Networking creates another overlap. NVIDIA sells NVLink, InfiniBand, Ethernet switches, data processing units, and related software. Broadcom provides components that equipment makers and cloud companies use in alternative or complementary network designs.

This makes the real contest broader than Broadcom versus NVIDIA. It concerns how much infrastructure value remains available to merchant component suppliers, custom-chip programs, and open networking systems.

Broadcom’s diversification also cuts both ways. Semiconductor demand does not stand alone inside the company. Its infrastructure software operations, including VMware, affect revenue, cash flow, debt management, and how investors interpret the business.

Customer concentration is the more direct AI risk. Designing custom processors requires substantial engineering work, and relatively few organizations can order them at hyperscaler volumes.

A delayed program or a change in one major customer’s roadmap can therefore affect growth. Broadcom can add customers over time, but each successful program requires a lengthy design and deployment cycle.

There is also an execution risk hidden behind rapid growth. Forecast demand must become completed designs, available manufacturing capacity, finished systems, and accepted customer deployments.

The next earnings report should clarify whether the projected acceleration remained intact. It should also show whether demand broadened beyond a small set of programs.

For readers arriving through Google News, Broadcom is the closest of the three companies to the chip narrative. However, it is not simply a less expensive version of NVIDIA.

It represents a different claim. The company is betting that hyperscalers will keep funding specialized silicon and high-speed connectivity alongside general-purpose accelerated computing.

Arista Networks Tests Ethernet at AI Scale

Arista’s opportunity rests on Ethernet becoming a dependable fabric for enormous AI clusters, not merely a familiar data center standard.

An AI training cluster produces traffic patterns that differ from ordinary enterprise computing. Thousands of processors exchange synchronized data, while a delayed connection can slow a larger job.

Network performance therefore affects how efficiently customers use their processors. Adding compute without sufficient bandwidth can leave expensive equipment waiting for data.

Arista sells complete switching systems and its EOS network operating software. This differs from Broadcom, which supplies silicon used across many vendors’ products.

Arista’s argument centers on Ethernet. The standard has a broad supplier base and extensive operational history, but AI workloads impose demanding requirements for congestion control, latency, and reliability.

The company has been developing systems for scale-out, scale-up, and scale-across networks. Scale-out connects more computing nodes. Scale-up links processors within a tightly integrated system. Scale-across joins separate clusters or facilities.

Arista reported $3.036 billion in second-quarter revenue, its first quarter above $3 billion. Revenue increased 37.7% from one year earlier and 12.1% from the preceding quarter.

Its GAAP operating margin reached 45.4%, while its non-GAAP operating margin was 49.9%. Those margins show that growth has not depended solely on sacrificing current profitability.

The company also introduced 1.6 terabit-per-second platforms, including liquid-cooled configurations for dense AI systems. Some products are scheduled for late 2026, while other configurations are planned for early 2027.

That timeline matters because announced products do not produce immediate proof of broad adoption. Customers still need to test the systems, qualify them for new clusters, and deploy them at scale.

Arista’s second-quarter disclosure described broad-based growth. However, total company revenue does not provide a perfect measurement of AI networking demand.

Arista also serves cloud, campus, routing, and enterprise environments. Those businesses reduce dependence on a single market, but they complicate efforts to isolate AI-related revenue.

The competitive pressure comes from several directions. NVIDIA can bundle processors with its own networking technology. Broadcom supplies switching silicon to multiple equipment makers. Cisco and other established vendors also want a larger role in AI data centers.

Customers may favor complete NVIDIA architectures when integration speed matters most. Others may choose Ethernet to preserve vendor choice or integrate equipment from several suppliers.

The technical outcome will not be decided by headline bandwidth alone. Operators will examine job completion time, network utilization, failure recovery, software tooling, energy consumption, and operational complexity.

Arista’s EOS software can strengthen its position because network operations become harder as clusters grow. A consistent operating environment helps teams automate configuration and diagnose failures.

Still, strong growth creates an expectations risk. When investors already assume continued expansion, a modest slowdown can change the market narrative even if the underlying business remains healthy.

Customer concentration deserves attention here as well. Large cloud operators can account for significant purchasing volumes, and their deployment schedules do not always move smoothly between quarters.

A customer can pause orders while shifting from one system generation to another. That pause may affect reported revenue without ending the longer infrastructure program.

This makes Arista an execution story rather than a simple bandwidth story. It must deliver new systems on schedule, win deployments, and maintain software advantages as competitors improve.

The next evidence should come from product availability and customer adoption. Shipments of the announced 1.6-terabit systems would show whether Arista can convert its roadmap into deployed capacity.

Readers using Google News to track AI chips should watch networking as a separate category. Faster processors increase the cost of network delays, making interconnect performance more economically important.

Yet that relationship does not guarantee a single winner. Ethernet’s growth can support Arista while also attracting more competition from NVIDIA, Cisco, and other system vendors.

Vertiv Turns Power and Cooling Into the Hard Limit

Vertiv sits farthest from the processor, yet it addresses the constraint that software cannot remove: dense computing requires electricity and heat management.

AI systems place unusual demands on data center facilities. Modern accelerators concentrate significant electrical load within each rack, and nearly all consumed energy eventually becomes heat.

Older facilities were often designed for lower rack densities. Installing advanced processors can therefore require upgraded power distribution, cooling equipment, monitoring, and backup capacity.

Vertiv sells those supporting systems. Its portfolio includes uninterruptible power supplies, power distribution equipment, thermal management, racks, and related services.

This position makes Vertiv less dependent on which processor architecture wins. A data center using NVIDIA GPUs, custom Broadcom accelerators, or another chip platform still requires reliable electrical and cooling systems.

However, independence from a chip vendor does not mean independence from the infrastructure cycle. Vertiv relies on customers completing data center projects and accepting equipment according to construction schedules.

The company reported $3.274 billion in second-quarter sales, up 24% from one year earlier. Organic sales growth, which excludes acquisitions and currency effects, reached 18%.

Operating profit increased 44%, while adjusted operating profit rose 51%. Vertiv also reported $1.1 billion in operating cash flow for the quarter.

Management raised its full-year 2026 guidance. At the midpoint, it expected $14 billion in sales and organic growth of 31% compared with 2025.

The updated guidance supports the case that power and thermal demand is reaching recognized revenue. It also establishes demanding expectations for the remaining year.

Vertiv’s physical role can look safer than semiconductor competition because electricity and heat are unavoidable. The commercial reality is less automatic.

Data center construction involves permitting, utility connections, financing, contractors, and long equipment lead times. A delay at any stage can shift revenue between reporting periods.

Orders can also arrive in large batches. This creates lumpiness, meaning quarterly results may change because of project timing rather than a lasting demand shift.

Customers are increasingly considering liquid cooling because air cooling becomes harder at higher rack densities. Liquid systems move heat through fluid placed closer to the computing equipment.

The transition expands Vertiv’s opportunity but adds design and execution challenges. Cooling architectures vary across facilities, processors, rack designs, and customer preferences.

Vertiv must coordinate with chip companies, server manufacturers, construction firms, and data center operators. The company’s position depends on integrating its equipment into a wider system rather than selling a standalone product.

Competition also remains substantial. Electrical equipment and thermal management attract established industrial suppliers, specialized cooling firms, and internal engineering efforts from large operators.

The durability question is whether Vertiv can maintain differentiation as demand expands. Capacity shortages can support suppliers during rapid construction, but customers may gain negotiating leverage when supply catches up.

Power availability creates another risk. Equipment demand cannot fully convert into deployments when utilities cannot provide additional capacity or transmission connections.

Local opposition can slow projects because data centers affect electricity systems, land use, water planning, and community infrastructure. Those issues differ across regions and can alter construction schedules.

Vertiv consequently offers the purest exposure to the physical bottleneck, but also the greatest exposure to project execution. Its results can remain strong while individual quarters fluctuate.

Google News coverage often groups power companies, data center operators, and equipment suppliers under one AI infrastructure label. Those categories have different revenue models and risk profiles.

Vertiv sells equipment and services rather than electricity or computing capacity. Readers should track its orders, backlog conversion, organic growth, and cash flow instead of relying on general data center spending headlines.

The Real Contest Is Bottlenecks Versus Expectations

Broadcom, Arista, and Vertiv can all benefit from the same buildout, yet none escapes customer concentration, competitive pressure, or demanding expectations.

The three businesses should not be treated as interchangeable AI infrastructure stocks. Each translates infrastructure spending into revenue through a different mechanism.

Broadcom depends on design wins for custom processors and networking silicon. Those wins can generate large volumes, but they concentrate risk among a limited number of sophisticated customers.

Arista depends on customers adopting Ethernet systems for demanding AI fabrics. Its hardware and software must compete against bundled architectures and other networking vendors.

Vertiv depends on physical construction and facility upgrades. Its opportunity can persist across chip architectures, but project timing affects when orders become revenue.

The companies also operate on different planning cycles. A custom accelerator can require years of design work before deployment. Networking systems move through qualification and cluster construction. Power equipment follows facility engineering and installation schedules.

These differences mean one quarter of weakness does not carry the same message across all three companies. Broadcom could face a delayed chip program. Arista could encounter a network transition. Vertiv could experience postponed site completion.

NVIDIA remains the essential reference because it is expanding into more of the system. The company no longer sells only processors to customers assembling everything else independently.

Its platforms include GPUs, CPUs, networking hardware, interconnects, software, and reference system designs. That wider scope can accelerate deployments while pressuring suppliers near its product boundaries.

Broadcom has the most direct competitive overlap through accelerators and networking components. Arista encounters NVIDIA in networking architecture. Vertiv remains more complementary because NVIDIA does not supply complete facility power and cooling portfolios.

Even Vertiv must follow NVIDIA’s roadmap closely. A change in rack density, cooling method, or electrical design can reshape what customers require from facility suppliers.

The resulting primary contest is not NVIDIA against three replacements. It is the growth of independent infrastructure bottlenecks against increasingly integrated computing platforms.

Open standards can preserve space for specialized suppliers. Ethernet allows customers to combine products from multiple vendors, while standard facility interfaces support a broad equipment market.

Integration can pull in the opposite direction. Customers under deployment pressure may accept a more complete architecture from one supplier to reduce qualification work and operational uncertainty.

Economics will decide many of these choices. Customers compare system utilization, power use, deployment speed, reliability, and the total cost of operating each AI service.

The rapid growth reported by all four companies shows that current demand is not merely hypothetical. It does not settle how profits will be distributed when supply expands or customer priorities change.

Valuation risk also exists even without quoting market prices or earnings multiples. Strong recent results can lead investors to assume that elevated growth will continue with little interruption.

Infrastructure cycles rarely move in a straight line. Customers can digest earlier purchases, redesign facilities, switch product generations, or delay construction because of power constraints.

Accounting classifications create another limitation. Companies define AI-related revenue differently, and some do not provide a complete segment-level breakdown.

Broadcom reports semiconductor revenue attributed to AI. Arista reports companywide results while discussing AI networking. Vertiv ties broader infrastructure demand to AI without isolating every AI-related sale.

These figures should not be placed in a direct ranking. They measure different products, reporting boundaries, fiscal calendars, and stages of deployment.

There is also no guarantee that faster hyperscaler spending benefits suppliers proportionally. Large customers can develop internal technology, negotiate aggressively, or concentrate orders among fewer vendors.

Google News can make the theme appear unified because many stories use the same AI infrastructure label. The operating evidence reveals a chain of businesses with distinct dependencies.

That is the useful reversal in this story. Looking beyond NVIDIA does not reduce exposure to NVIDIA’s demand cycle. It changes which part of that cycle an investor or industry observer is measuring.

Three Signals to Watch After the Headlines

The thesis now faces three concrete tests: Broadcom’s forecast, Arista’s product adoption, and Vertiv’s conversion of physical demand into completed revenue.

The first signal arrives on September 2, when Broadcom is scheduled to report fiscal third-quarter results. Its earlier guidance called for approximately $16 billion in AI semiconductor revenue.

A result near that forecast would support management’s claim that custom accelerators and AI networking are accelerating. A material shortfall would raise questions about program timing, customer concentration, or shipment execution.

The updated outlook will matter just as much as the reported quarter. Custom silicon programs involve long planning cycles, so forward guidance can reveal whether current growth extends into later deployments.

The second signal is customer adoption of Arista’s newest AI fabric platforms. Announced specifications are useful, but deployments provide stronger evidence about competitive position.

Readers should watch whether Arista ships its planned 1.6-terabit systems on schedule. Customer commentary about scale-up Ethernet would also clarify how the technology performs inside tightly connected systems.

Broader adoption would strengthen the case for Ethernet as an alternative to proprietary interconnects. Delays or limited deployments would leave more room for NVIDIA’s integrated networking approach.

The third signal is Vertiv’s ability to deliver its raised full-year outlook. Its midpoint calls for $14 billion in sales and 31% organic growth.

Meeting that guidance would show that orders are converting into installations despite construction and utility constraints. Weak conversion would suggest that physical project delays are catching up with reported demand.

Cash flow should accompany revenue growth. Strong cash generation indicates that recognized sales are turning into collected cash rather than expanding working-capital needs.

These three tests also give readers a better framework for interpreting future Google News coverage. A new chip announcement matters differently from a networking deployment or a data center construction delay.

Teams evaluating AI services should care about the same chain, even if they never buy individual infrastructure components. Supply constraints influence cloud capacity, deployment schedules, reliability, and operating costs.

Knowledge workers experience the consequences indirectly. Faster infrastructure can improve model availability and response times, while constrained capacity can produce usage limits or delayed enterprise rollouts.

The right question is therefore not which company becomes the next NVIDIA. It is whether specialized suppliers keep capturing value as AI systems become denser, more connected, and more integrated.

Watch the evidence in order: Broadcom’s reported AI revenue, Arista’s deployed network systems, and Vertiv’s completed facility projects. If all three advance, the second infrastructure layer is broadening. If one stalls, the reason will identify the bottleneck that demand alone could not solve.