The investable AI infrastructure story has moved beyond a single question—how much silicon hyperscalers can buy—toward a more useful one: which parts of the stack can convert unprecedented capital spending into contracted revenue, cash flow and defensible returns? Nvidia reported $96.2 billion in quarterly revenue, including $89 billion from data centers, and forecast that its five largest customers will spend $1.3 trillion on AI infrastructure next year. Those figures establish the scale of the cycle. They do not establish that every supplier, builder or operator offers the same risk-adjusted opportunity.
This analysis uses reported company results, contract disclosures and market data from the supplied research set. Figures reflect the periods stated in those reports. That distinction matters because backlog, purchase commitments, revenue, earnings and cash flow are not interchangeable.
Where does demand have the clearest visibility beyond accelerators?
The direct answer: visibility is strongest where customers have placed orders, signed multiyear supply agreements or disclosed measurable consumption. Servers, storage, optical networking, cooling and power equipment are therefore more than ancillary trades; they are the physical bottlenecks through which AI spending must pass.
Compute systems and storage
Dell provides the clearest systems-level evidence. It reported $60.9 billion of AI server orders, a $95 billion ending backlog and more than $130 billion of AI server orders over 12 months. Management expects AI server revenue to more than triple to $74 billion for the full year. The practical implication is that original equipment manufacturers can capture growth even when customers diversify accelerators or design custom silicon.
Storage suppliers show a different form of visibility. Western Digital said it was effectively sold out for calendar 2026 on firm orders from its seven largest customers, with multiyear agreements extending into 2028–2029 for three of its five largest customers. Seagate has configurations and pricing fixed through 2027 and most nearline exabyte capacity allocated into 2028. Sandisk has 10 long-term agreements covering eight customers, with roughly two-thirds of fiscal 2028 volume locked in and price floors; even if every variable price lands at its floor, the agreements represent at least $93.9 billion of revenue. The “so what” is simple: contracted volume improves forecast quality, but it does not eliminate exposure to customer concentration, pricing resets after contract windows or a slowdown in data-center spending.
Networking and optical connectivity
Networking deserves a dedicated allocation screen because larger AI clusters increase the value of moving data between processors, memory and storage. Credo reported quarterly revenue of $479 million, up 115%, with 68% gross margin and a 48.2% operating margin. It expects fiscal 2027 optical revenue above $600 million, but its four largest customers generated 84% of revenue. Ciena reported an $8.5 billion backlog, guided to more than $10 billion by fiscal year-end and forecast 30% fiscal 2027 revenue growth; two customers nevertheless represented 42% of quarterly revenue.
Custom silicon is also moving demand beyond standard accelerators. Qualcomm’s Amazon partnership covers custom AI inference chips and optical connectivity across Amazon’s data-center network, with Amazon’s warrant vesting alongside as much as $60 billion in chip orders and related purchases. Inference is the likely long-duration workload: models may train in bursts, but deployed agents query infrastructure continuously. Investors should therefore track optical content, design wins and production ramps—not just accelerator headlines.
The facilities layer
Vertiv reported a $15 billion backlog against $10.2 billion of 2025 sales, guided 2026 sales to $13.8 billion–$14.2 billion and targets 20%–22% organic annual growth through 2030. The caveat is material: orders can be canceled or rescheduled, and much of the backlog converts within 12–18 months. Corning’s 80-million-mile high-density fiber contract with Verizon adds another visible networking build signal. At the commodity layer, copper reached a reported record $6.70 per pound and the futures curve moved into backwardation, while ore grades have fallen about 40% since 1991 and permitting can take more than a decade. Capacity expansion can therefore be constrained long after demand appears certain.
Is electric power the highest-conviction AI infrastructure layer?
The direct answer: power offers exceptional contract duration, but the investment case depends on start dates, customer credit and the share of earnings actually supported by AI contracts. A 20-year power agreement is not a 20-year earnings guarantee.
Constellation’s Meta agreement covers the clean-energy attributes of 1,121 MW from Illinois’ Clinton Clean Energy Center for 20 years beginning in June 2027. That exceeds its roughly 835 MW Three Mile Island restart agreement with Microsoft. Constellation also disclosed 920 MW of new long-term contracts with investment-grade customers on 15- to 20-year terms beginning in 2029–2032. Yet its 2026 earnings growth is currently driven more by the Calpine acquisition and market conditions; the AI contracts begin later. At roughly 25 times the midpoint of 2026 guidance, investors are paying for future visibility before most of it reaches earnings.
NextEra’s plan to restart Iowa’s 615 MW Duane Arnold nuclear plant illustrates both the opportunity and the timing risk. The project has access to a DOE loan guarantee of up to $1.9 billion, a 25-year power-purchase agreement tied primarily to Alphabet data centers and a planned operating date in early 2029. Alphabet has committed $7 billion to expand its Iowa data-center footprint. The asset could provide durable demand, but execution, relicensing and commissioning determine when cash flow arrives.
Equipment and transmission companies may monetize the build sooner. GE Vernova reported 116 GW of gas-turbine backlog and reservations, with at least 125 GW targeted by year-end and an annualized capacity path from 20 GW toward 30 GW by 2030. In India, GE Vernova T&D India secured an estimated ₹13,000 crore HVDC order, lifting order inflows to about ₹23,100 crore—nearly four times fiscal 2026 revenue—and expects a $1 billion data-center-related order from its U.S. parent. Odisha separately approved ₹15,948.7 crore to expand transmission capacity ahead of industrial projects that include data centers and semiconductors. These are global examples of grid investment becoming an investable AI-enabling theme.
A disciplined power screen should require five disclosures: contracted megawatts, contract term, customer credit quality, commercial-operation date and capital required per megawatt. It should also separate contracted generation from merchant exposure and identify regulatory or ratepayer dependencies.
Which AI applications are creating demand rather than merely consuming capital?
The direct answer: look for software vendors showing consumption growth, expanding remaining performance obligations, improving free cash flow and increasing large-customer wins. These metrics show that AI is attaching to existing workflows rather than remaining a research expense.
GitLab reported revenue of $286.3 million, up 21%, non-GAAP operating margin of 15%, net ARR growth of 42% and first-order growth above 100%. Deals worth at least $500,000 grew more than 150%, billings rose 24% and remaining performance obligations reached $1.2 billion. Paid consumption run rate exceeded $40 million, and management targets more than $100 million by fiscal year-end. MongoDB reported revenue of $771.8 million, up 30%, with Atlas revenue of $565.9 million, remaining performance obligations up 91% to $1.52 billion, 122% net ARR expansion and free cash flow equal to 18% of revenue. Vector search, operational data and managed AI tooling are turning infrastructure usage into platform expansion.
Cybersecurity offers the strongest cash-flow evidence because AI agents enlarge the identity, data and attack surfaces that enterprises must protect. Palo Alto Networks reported revenue of $3.41 billion, up 34%, next-generation security ARR of $9.1 billion, up 63%, and adjusted free cash flow of $4.41 billion, a 38.4% margin. Okta reported $805 million of revenue, 15% earnings growth and free cash flow of $227 million, up 40%; current remaining performance obligations rose 14%. By comparison, CrowdStrike’s revenue reached $1.5 billion over eight quarters, but its quarterly operating margin was -2% and its price-to-sales ratio was reported at 38, versus 13% operating margin and 10 times sales for Okta. Growth alone is not a sufficient quality screen.
Real-world adoption is still uneven. In a survey of 28 Indian healthcare leaders, 93% believed AI improves efficiency, 64% had piloted it and only 11% had moved it into production. Data quality was the leading barrier at 36%. That gap creates demand for data platforms, governance and security, but it also argues against treating announced pilots as immediate revenue.
How should a diversified technology sleeve be allocated?
The direct answer: use a core-and-satellite structure built around visibility and cash conversion, not an equal-weight basket of companies with “AI” in their strategy. An illustrative framework for a diversified technology sleeve is:
- Core demand owners and monetizers: integrated cloud platforms, profitable software vendors and cybersecurity companies with recurring revenue and positive free cash flow.
- Infrastructure picks and shovels: servers, storage, optical networking, cooling and electrical equipment with order books tied to identifiable capacity additions.
- Contracted power and grid exposure: generation, transmission and equipment suppliers whose contract start dates match the portfolio’s time horizon.
- International and sovereign-AI satellites: companies benefiting from regional cloud, semiconductor and grid investment rather than relying solely on U.S. hyperscaler budgets.
- Speculative capacity builders: neoclouds and development-stage operators sized according to leverage, financing needs and proof of utilization.
For an income-aware portfolio, the AI sleeve can be balanced with cash-generative assets rather than chased at any valuation. Chevron offered a 3.38% dividend yield, had raised its dividend for 39 consecutive years and generated $15.4 billion of adjusted free cash flow against $3.5 billion of dividend payments in the reported quarter. The Schwab U.S. Dividend Equity ETF held 102 companies meeting dividend-duration, dividend-growth, free-cash-flow and return-on-equity screens and yielded 3.05%. These holdings do not hedge every AI-specific risk, but they add income and business-model diversification.
How much neocloud and financing risk is too much?
The direct answer: treat neoclouds as leveraged infrastructure developers, not software companies. Their upside comes from scarce power and compute; their downside comes from debt, construction delays, customer concentration and underutilized capacity.
CoreWeave illustrates the trade-off. Revenue rose 112.5% to $2.575 billion, but free cash flow was negative $5.74 billion and long-term debt reached $27.56 billion. It had 1.5 GW of active power and targets 8 GW by 2030. IREN holds a $9.7 billion Microsoft contract but faces an estimated $25 billion–$30 billion of capital expenditure through the second quarter of 2027 and reported a $684 million quarterly net loss. Nebius is less mature but provides a contrasting validation signal: revenue grew 454% to $582.3 million, and its partnership making it Palantir’s preferred sovereign-AI infrastructure partner expands demand beyond hyperscalers; gross margin was only 18.81%.
Financing capacity is becoming part of the competitive moat. Bankers are seeking investment-grade ratings for OpenAI and Anthropic after potential public listings; Anthropic’s possible roughly $100 billion IPO could give an investment-grade issuer access to the $11.7 trillion corporate-bond market and reduce reliance on tens of billions of credit support from Nvidia, Oracle, Google and Broadcom. Amazon has arranged sterling funding for AI expansion, while Alphabet previously raised £5.5 billion in the currency. With the 10-year Treasury near 4.79% and one estimate putting AI-related corporate debt issuance at about $1.5 trillion this year, the spread between borrowing cost and deployed-capital returns matters as much as gross compute demand.
Before allocating to a capacity builder, require a clear view of contracted revenue, power cost, debt maturities, interest coverage, capex per megawatt, utilization, customer concentration and cancellation rights. Backlog should be treated as a pipeline, not cash flow.
Where can investors find U.S. and international diversification?
The direct answer: combine direct U.S. infrastructure exposure with selected sovereign-AI, grid and broad international positions. Mistral raised $3.5 billion at a $24.4 billion valuation in a Samsung-led round, reinforcing Europe’s sovereign-AI contender while remaining much smaller than leading U.S. rivals. Samsung’s involvement also links model development to chip manufacturing and cybersecurity demand. In India, transmission orders and state grid programs show that AI demand is pulling forward electrical infrastructure even where domestic model vendors are less prominent.
Broad international equity exposure can also temper U.S. valuation concentration. Vanguard’s Capital Markets Model forecasts 5.9%–7.9% annual returns for global stocks over 30 years versus 4.7%–6.7% for U.S. stocks. The Vanguard Total International Stock ETF holds 8,772 non-U.S. companies, trades at a reported trailing P/E of 18.44 versus 23.61 for the S&P 500 and yields 2.5%. It is not a pure AI-infrastructure vehicle, but it offers a lower-concentration way to participate in global capital formation.
What should decision-makers prioritize now?
The highest-quality opportunities are not necessarily the fastest-growing names. They are the companies that can document demand, fund the required capacity and retain enough pricing power to earn attractive returns on that capacity. The research set points to four priorities:
- Favor disclosed orders, multiyear contracts and customer credit quality over thematic exposure.
- Verify contract start dates so power and transmission investments match the intended holding period.
- Reward applications businesses that show consumption, remaining performance obligations and free-cash-flow growth together.
- Limit speculative neocloud exposure until utilization and financing milestones are visible.
The chip giants define the size of the AI build-out. The next layer of opportunity is broader: the systems that house accelerators, the optical networks that connect them, the power and cooling that keep them running, and the software that turns capacity into revenue. A portfolio built around those measurable links has a stronger research foundation than one built around the AI label alone.