6 AI Infrastructure Stocks Beyond Nvidia: CEG to AVGO
6 AI Infrastructure Stocks Beyond Nvidia
CEG · VST · VRT · MU · ANET · AVGO
Follow the electricity, cooling, memory and connections that turn expensive computing capacity into useful work.
Research date: October 8, 2026 · Company disclosures and educational analysis · No live price targets
A popular investing pitch asks readers to forget missing Amazon or Nvidia and focus on the next group of potential winners. The more useful insight is less dramatic: artificial intelligence needs an entire physical system, and a shortage anywhere in that system can limit the return on everything else.
That makes six AI infrastructure stocks worth studying together: Constellation Energy, Vistra, Vertiv, Micron, Arista Networks and Broadcom. Their businesses span power generation, electrical equipment, cooling, memory, network systems and custom accelerators. Each addresses a different constraint on deploying and operating AI capacity.
This is a research framework, not a prediction that these companies will reproduce Nvidia’s returns. Essential products do not automatically create exceptional shareholder returns. Competition, capital requirements, customer bargaining power and the valuation paid determine how much of industry growth ultimately reaches investors.
If AI demand expands, which bottleneck becomes more valuable, which company captures that value, and how much success does its share price already assume?
The AI infrastructure value chain is a system
An accelerator cannot deliver useful output without memory, communications, electricity and heat removal. However, these dependencies are not a literal sequential production line. Cooling and electrical systems support the whole facility, while memory and networking operate alongside compute. Several companies also compete or collaborate across overlapping layers.
| Stock | Main exposure in this framework | How demand becomes revenue | Central question |
|---|---|---|---|
| CEG Constellation | Reliable electricity generation | Energy sales, contracts and capacity | Do contract economics justify investment? |
| VST Vistra | Nuclear and flexible generation | Generation and retail energy economics | How much scarcity value is retained? |
| VRT Vertiv | Power and thermal management | Equipment, systems and services | Do orders convert into profitable deliveries? |
| MU Micron | HBM and broader memory | Memory shipments and product mix | Can margins survive the supply response? |
| ANET Arista | Ethernet systems and software | Network platforms and support | Does adoption translate into durable share? |
| AVGO Broadcom | Custom accelerators and networking chips | Design programs and semiconductor shipments | How durable are major customer programs? |
The table is a map of investment exposures, not a supplier list for a particular data center. A facility can use competing products, different power arrangements or a mixture of architectures. No claim is made that all six companies participate in every AI installation.
1. CEG: electricity availability becomes strategic
Constellation’s nuclear fleet gives it exposure to demand for dependable, low-carbon electricity. Data centers need substantial power across the day, making reliability valuable. Yet the existence of generation does not guarantee that a particular site can receive it: transmission, interconnection and local distribution remain separate constraints.
One important update changes the comparison with Vistra. Constellation completed its Calpine acquisition on January 7, 2026, combining nuclear operations with natural-gas and geothermal generation. Investors should therefore analyze CEG as a broader power business, rather than treating it as a pure nuclear company.[1]
On October 6, Constellation and Google announced a 20-year power purchase agreement supporting 890 megawatts of additional nuclear capacity through improvements at existing units. The capacity is incremental output from upgrades, not a newly built 890-megawatt reactor. The announcement illustrates how long-term customer commitments can support investment in existing assets.[2]
The investment test is contractual economics. A long agreement can improve revenue visibility while still requiring substantial capital expenditure and exposing the producer to operating costs, outages and delivery obligations. Contract duration alone does not reveal the return on capital or establish that shareholders receive an attractive price.
Track project schedules, investment requirements, plant availability and the share of generation sold under different arrangements. Scarcity creates an opportunity to negotiate, but future earnings depend on the terms negotiated. Nuclear reliability also does not remove maintenance outages, regulatory oversight or the cost of extending asset lives.
2. VST: the value of dependable and flexible capacity
Vistra provides another route into electricity demand, with generation and retail operations rather than a narrowly defined AI product. Nuclear assets can support sustained output, while gas generation can offer flexibility. How much flexibility is available depends on the plant, fuel access, operating conditions and market rules.
The strongest evidence for this thesis is a commercial commitment, not a general forecast about electricity consumption. In January 2026, Vistra announced 20-year agreements with Meta covering 2,176 megawatts of operating nuclear generation and 433 megawatts of planned additional output. Existing capacity and future uprates are economically different components and should be tracked separately.[3]
Existing assets may gain value when new supply takes years to permit and build. But that value is regional. A plant near a constrained market can face different economics from one elsewhere, and a national AI-demand forecast cannot substitute for understanding the relevant power market.
Vistra’s results also depend on hedging, retail obligations, fuel costs and availability. Higher wholesale prices do not flow directly into profit for every megawatt produced. Previously contracted output may not capture a current price spike, while retail commitments can change the net exposure.
For investors, the practical comparison with CEG is broader than nuclear versus gas. Examine the geographic mix, balance sheet, contract book, planned spending and sensitivity to market prices. A valuable fleet can still deliver disappointing equity returns if its acquisition or stock-market valuation already assumes unusually favorable conditions.
3. VRT: keeping dense computing systems operating
Vertiv addresses the infrastructure inside and around the facility: power delivery, backup systems and thermal management. More demanding racks can require new approaches to distributing electricity and removing heat. The relevant opportunity includes installation and service, not simply selling a standalone cooling component.
Its second-quarter 2026 release reported sales of $3.274 billion, up 24% year over year. Organic growth was 18%, with acquisitions and currency contributing to the remainder. Adjusted operating margin reached 22.6%. These distinctions matter because total growth and adjusted profitability answer different questions from underlying expansion and GAAP earnings.[4]
Liquid cooling can support higher-density systems, but adoption is not uniform across every data center. Existing buildings, rack configurations, water systems and customer preferences influence what is installed. Investors should avoid assuming that every announced computing project requires the same equipment mix or spending per rack.
Backlog can show demand visibility, yet it is not cash in the bank. Orders must survive scheduling changes and become deliveries, accepted systems and collected payments. Installation complexity can postpone recognition even when a customer remains committed to the project.
Watch organic sales, margins, working capital and cash conversion together. If revenue expands while inventory and receivables rise much faster, investigate the reason. Strong growth is more valuable when execution preserves profitability and the company does not need disproportionately more capital to support each additional dollar of sales.
4. MU: memory bandwidth enables compute utilization
Micron supplies memory products, including high-bandwidth memory used with advanced AI accelerators. Memory capacity and bandwidth are different: capacity concerns how much data can be held, while bandwidth concerns how quickly data moves. Both can constrain system performance, depending on the workload.
Micron’s product materials identify HBM as serving demanding AI and high-performance computing environments. That makes the company relevant to the infrastructure thesis, but not exclusively an AI business. Broader DRAM and NAND markets remain important to its overall operating picture.[5]
Its fiscal fourth-quarter 2026 release reported an 86.8% GAAP gross margin and an 87.0% non-GAAP gross margin. Those are the company’s reported quarterly figures, not assumptions that should be extended indefinitely. Extraordinary profitability makes an assessment of supply discipline and sustainable earnings especially important.[6]
Memory demand can expand while memory investment returns weaken. If supply grows faster, pricing may deteriorate even as shipments increase. HBM qualification, packaging and technical requirements can affect supply responsiveness, but they do not abolish competition or the possibility of overinvestment.
Monitor product mix, pricing, manufacturing yields, capital spending and customer commitments. Long-term arrangements can improve visibility, although their economic value depends on pricing provisions and purchase obligations. A low P/E based on exceptional earnings is not conclusive evidence of cheapness; normalized profitability can produce a very different valuation.
5. ANET: connecting accelerators efficiently
Arista builds Ethernet networking systems and software for large computing environments. When workloads span multiple accelerators, communication can limit useful output. Network performance involves congestion management, reliability, latency and topology as well as headline link speed; buying faster ports alone does not solve every bottleneck.
The company reported second-quarter 2026 revenue of $3.036 billion, up 37.7% year over year, with a 45.4% GAAP operating margin. These are company-wide results, not a statement that all revenue came from AI networking. Product-level opportunity and consolidated financial performance should remain distinct.[7]
A timely development came on October 7, when Arista announced rack-scale Ethernet architectures for scale-up and scale-out AI fabrics. This extends the story beyond conventional connections between servers. An announcement establishes a product direction, however, not the future revenue, customer adoption or profit contribution of that offering.[8]
Scale-up and scale-out networks serve different communication needs and can coexist. Ethernet adoption is also not synonymous with Arista capturing every order. Competing systems, customer-designed architectures and other interconnect approaches can influence procurement and margins.
Evaluate customer concentration, repeat deployments, competitive differentiation and cash generation. Networking spending does not have to rise faster than accelerator counts in every architecture. Port speeds, utilization and network design can improve efficiency, so the relationship between compute growth and supplier revenue must be demonstrated rather than assumed.
6. AVGO: custom accelerators and networking silicon
Broadcom participates in custom AI accelerators and networking semiconductors. Large customers can justify specialized designs when workload scale makes performance, power efficiency or total cost attractive. Custom hardware can complement general-purpose accelerators rather than replacing them across every type of computing.
In its fiscal third-quarter 2026 release, Broadcom reported AI semiconductor revenue of $16.7 billion, up 221% year over year. Total revenue was approximately $29.6 billion. The company also operates infrastructure software and other semiconductor businesses, so its overall valuation includes more than this AI exposure.[9]
Custom silicon creates close technical relationships, but concentrated customers can possess substantial bargaining power. A large program can drive growth while increasing dependence on one customer’s roadmap. Design changes, deployment delays or a supplier change may affect several reporting periods.
Development wins should be distinguished from production shipments and realized cash. A successful design must progress through manufacturing, packaging, qualification and deployment. Specialized hardware also needs usable software; theoretical performance per watt is only one component of the customer’s actual economics.
Follow the breadth of production programs, networking demand, cash conversion and customer dependence. Broadcom’s connection to both compute and communications strengthens its relevance, but also overlaps with risks elsewhere in this watchlist. Multiple products sold into the same spending cycle do not eliminate exposure to a slowdown in that cycle.
Demand growth is only the first valuation question
The six companies require different valuation approaches. Comparing them using one forward P/E can obscure capital intensity, cyclicality and accounting differences. A generator replacing equipment, a memory manufacturer expanding fabrication capacity and a semiconductor designer have different reinvestment demands.
| Business | Useful valuation lens | Common analytical trap |
|---|---|---|
| CEG / VST | Contract cash flows, debt and required investment | Treating all generation as exposed to spot prices |
| VRT | Organic growth, sustainable margins and cash conversion | Equating backlog with completed profitable revenue |
| MU | Midcycle earnings and returns on new capacity | Capitalizing peak margins as permanent profits |
| ANET | Durable growth, customer breadth and competitive returns | Confusing Ethernet adoption with guaranteed market share |
| AVGO | Program durability, cash flow and business mix | Ignoring concentration behind rapid AI growth |
Consider an illustrative company earning $10 per share at a 40-times multiple, giving a $400 price. If earnings grow 25% to $12.50 but the multiple falls to 30 times, the resulting price is $375. The business improved, yet the modeled share price declined 6.25%, excluding dividends.
This simple example explains why the best infrastructure narrative is not necessarily the best investment at every price. Investors need an estimate of what operating outcome is already required by the valuation. Excellent execution may merely satisfy that expectation, rather than exceed it.
Six stocks can still share one major risk
Each company has distinct operations, but large cloud customers influence several layers simultaneously. A delay in data-center deployment can affect equipment orders, networking purchases, memory demand and expected electricity consumption, though the timing and contractual protection differ.
Efficiency introduces another shared uncertainty. Better chips, models and utilization can reduce infrastructure required per unit of useful AI work. Lower costs may stimulate enough new usage to expand total demand, but the outcome is not automatic. Compute growth and electricity consumption need not increase in a fixed ratio.
Nor do bottlenecks migrate in a predictable sequence. A project can face power, packaging and network constraints at the same time. Solving one may expose another, while overbuilding can remove scarcity and reduce supplier pricing power. The framework is most useful when updated with operating evidence.
| Scenario | Industry development | Evidence that matters |
|---|---|---|
| Expansion with discipline | Demand grows and capacity arrives at attractive returns | Utilization, collections and stable profitability |
| Growth with lower returns | Volumes increase but competition and spending absorb gains | Pricing, margins and free cash flow |
| Deployment slowdown | Projects slip and customers defer purchases | Order changes, funding and contract protections |
Turn the theme into a repeatable research process
Choose a few measurable indicators for each company before reviewing the next quarter. Separate management targets from reported results, and distinguish organic expansion from acquisitions. For project announcements, record the expected operating date and spending obligations rather than only the headline size.
Then identify what would invalidate the thesis. A routine stock-price decline is different from persistently weaker returns on investment, a lost strategic customer or a funding shortfall. Maintaining that distinction can reduce both panic selling and the temptation to average down whenever a familiar ticker becomes cheaper.
The full investment case requires necessity, durable value capture and a reasonable price. Missing a previous winner does not remove any of those requirements.
Frequently asked questions
Are these six companies alternatives to Nvidia?
They provide different exposures to the wider AI system. Some supply complementary infrastructure, while some products compete with parts of Nvidia’s platform. This is not a claim that six stocks replicate its economics or returns.
Which are the power stocks?
Constellation and Vistra generate electricity. Vertiv supplies power-management and thermal infrastructure rather than serving the same role as a generating fleet. Their business models and investment requirements differ.
Does doubling AI compute mean doubling every supplier’s sales?
No. Efficiency, equipment prices, architecture, product mix and replacement cycles change the relationship. A larger computing market can coexist with falling prices or lower spending per unit of useful output.
Why is Micron different from a software growth stock?
Memory manufacturing requires substantial capital, and supply-demand imbalances can sharply change pricing and profits. AI demand can support growth without removing those cyclical characteristics.
Does holding all six provide diversification?
It adds business-model variety but retains shared exposure to AI investment and major customers. Portfolio diversification also depends on position sizes, other holdings and how losses might occur together.
Does this article identify the cheapest stock?
No. It offers a research framework supported by dated disclosures. Ranking current value requires updated prices, consistent earnings estimates, debt information and assumptions about sustainable returns.
Official sources and methodology
Research checked October 8, 2026. Financial periods differ across companies. Figures are selectively included to illustrate the thesis, not to rank performance across unmatched periods. Announced projects and products are distinguished from realized earnings. No unidentified social-media post is treated as verified investment evidence.
- Constellation: completion of Calpine acquisition, January 7, 2026.
- Constellation and Google: nuclear capacity agreement, October 6, 2026.
- Vistra and Meta: nuclear power agreements, January 9, 2026.
- Vertiv: second-quarter 2026 earnings release filed with the SEC.
- Micron: high-bandwidth memory product information.
- Micron: fiscal fourth-quarter and full-year 2026 results.
- Arista: second-quarter 2026 results, August 4, 2026.
- Arista: rack-scale AI architectures, October 7, 2026; AI networking overview.
- Broadcom: fiscal third-quarter 2026 results, September 2, 2026.
Disclaimer: Educational information, not personalized investment advice or a recommendation to buy any security. Forecasts, projects and market conditions can change. Investments involve risk, including loss of principal.