Building Dividend Portfolios

The AI investment cycle has created one of the largest capital concentration events in the history of technology. The five largest cloud and infrastructure companies are on track to collectively spend somewhere between six hundred and seven hundred billion dollars on AI infrastructure in 2026 alone. That capital has to pass through somewhere before it reaches anyone's product. Understanding where it passes through, and under what terms, is the most important financial question in the AI market right now. The answer does not live in the companies with the largest language models. It lives in the company with the largest lock on the hardware those models run on.

A useful way to think about the current AI cycle is to separate it into three distinct layers: the infrastructure layer, which provides the hardware and compute capacity; the platform layer, which provides cloud access, developer tools, and enterprise software built on top of that hardware; and the model layer, which provides the actual AI systems companies and consumers use. These three layers are not equally profitable, and they are not equally protected from competition. Most of the public narrative concentrates on the model layer. Most of the money concentrates at the infrastructure layer. That gap is the core financial story of this cycle.

The raw numbers confirm the direction of the spending. In 2026, Amazon is projected to spend around two hundred billion dollars on capital expenditure, the majority of which goes toward data center infrastructure. Alphabet is projected at roughly one hundred seventy-five to one hundred eighty-five billion. Microsoft is tracking above one hundred twenty billion. Meta is guiding between one hundred fifteen and one hundred thirty-five billion. Those four companies alone represent the vast majority of GPU procurement globally. The trajectory has not slowed. According to Goldman Sachs projections, total hyperscaler capital spending from 2025 through 2027 will reach approximately one point one five trillion dollars, more than double what was spent across 2022 through 2024.

What changed is not that AI became commercially important. What changed is that the companies with the largest balance sheets decided they could not afford to find out whether a competitor had figured out AI monetization while they had not. When competitive fear rather than measured return analysis drives capital spending, the supplier of the scarce input becomes extraordinarily valuable.

Nvidia generated two hundred fifteen point nine billion dollars in revenue in its fiscal year ending January 2026, up sixty-five percent from the prior year. Its data center segment, which houses the GPU business at the center of all AI infrastructure spending, contributed over ninety-one percent of that total. In the fourth quarter alone, data center revenue reached sixty-two point three billion dollars, up seventy-five percent year over year. Gross margins across the company have held in the low to mid seventy percent range. Net margins are approaching fifty-six percent. Free cash flow for the trailing twelve months stands at roughly ninety-seven billion dollars against capital expenditures of around six billion. These are not semiconductor company numbers. They are closer to what you expect from a monopoly software business with no marginal cost of production, except Nvidia manufactures physical hardware.

The reason this is financially possible is a software moat that took fifteen years to build. The CUDA programming environment, which developers and researchers use to write code that runs on Nvidia hardware, has become so embedded in the training and inference workflows of every major AI company that the switching cost is closer to rebuilding your engineering department than updating a vendor contract. AMD produces chips that are technically competitive in some benchmarks and less expensive. The market share data suggests this does not matter much. Engineers who have written hundreds of thousands of lines of CUDA-optimized code do not retrain their teams to switch to a lower-cost alternative when the infrastructure their models depend on is running at production scale. The software moat compounds the hardware moat, and the hardware moat compounds the supply constraints.

Demand continues to exceed supply. Nvidia guided its next fiscal quarter to approximately seventy-eight billion dollars in revenue, against analyst expectations of around seventy-two point six billion. The company shipped its first next-generation Vera Rubin samples to customers last week and has noted supply constraints as a headwind even to its gaming segment. When a company is capacity-constrained at seventy-eight billion dollars in quarterly revenue and its forward price-to-earnings multiple sits around twenty-five to twenty-nine times on a consensus basis, the market is pricing in a meaningful deceleration. Whether that deceleration arrives, and when, is the central tension in Nvidia's valuation.

Microsoft occupies a structurally distinct but financially compelling position. Its Intelligent Cloud segment, which houses Azure, grew twenty-eight percent year over year in the most recent quarter with Azure itself up thirty-nine percent. Operating margins for the company as a whole held above forty-seven percent despite capital expenditures that hit thirty-seven point five billion dollars in a single quarter. The commercial remaining performance obligation, which is essentially contracted future revenue, reached six hundred twenty-five billion dollars, with roughly forty-five percent attributed to OpenAI commitments. That figure is a measure of revenue visibility that very few companies at any scale can point to. Microsoft also reported an eighty billion dollar backlog of Azure orders it cannot currently fulfill due to power constraints at its data centers, which is a demand signal more than it is a supply failure.

The mechanism through which Microsoft converts AI investment into recurring revenue is the Copilot product family embedded across Microsoft 365, GitHub, Dynamics, and the Azure platform. The company now reports one hundred fifty million monthly active users across its Copilot products, with adoption growing fifty percent quarter over quarter in its most recent disclosure. The financial logic is straightforward: Microsoft is attaching a margin-accretive product to enterprise relationships it has maintained for decades. The incremental cost of adding AI capability to an existing Microsoft 365 contract is low compared to the pricing power it creates. The enterprise does not need to onboard a new vendor, negotiate new security reviews, or migrate data. It expands an existing agreement.

The private market valuations in the model layer tell a different story. OpenAI ended 2025 with approximately thirteen billion dollars in actual revenue and is projecting a fourteen billion dollar operating loss in 2026. The company expects cumulative losses of forty-four billion dollars through 2028 before any path to profitability. It is currently in discussions for a funding round that could raise up to one hundred billion dollars at a valuation between seven hundred thirty and eight hundred thirty billion, according to reporting from The Information and Reuters. The gap between revenue and valuation implies either that the model layer will eventually generate extraordinary margins at scale, or that current valuations reflect competitive positioning rather than financial fundamentals.

ChatGPT's share of global AI web traffic tells part of the story. Similarweb data shows it declined from eighty-six point seven percent in January 2025 to sixty-four point five percent in January 2026 as Google's Gemini captured substantial ground, aided by the Apple Intelligence integration announced in January. A thirty-billion-dollar consumer service losing twenty-two percentage points of category share in twelve months while burning over a billion dollars a month is a financial profile that requires a very specific set of assumptions about the future to justify its current valuation.

Anthropic represents a meaningfully different case within the model layer. The company grew from one billion dollars in annual run-rate revenue at the start of 2025 to fourteen billion dollars by February 2026, a trajectory that is among the fastest in enterprise software history. It raised thirty billion dollars in a Series G at a three hundred eighty billion dollar valuation in February 2026, with its customer base concentrated in enterprises and developers rather than consumer subscriptions. The enterprise concentration matters financially because it implies stickier revenue, higher average contract values, and lower customer acquisition cost relative to consumer-facing services. Whether a revenue multiple approaching thirty times on a three-hundred-eighty-billion-dollar valuation can be sustained depends on whether the enterprise market for AI models develops the kind of switching costs that software markets typically require to hold pricing.

The tension across the model layer is that switching costs at this level are genuinely low. A company running workloads on Claude can migrate to GPT-4 or Gemini in weeks. The API structures are standardized enough that the friction is real but not prohibitive. This structural feature will compress margins in the model layer over time in ways that the current private valuations do not fully reflect.

The hyperscalers funding this infrastructure cycle are doing so at capital intensity ratios that would have seemed implausible two years ago. Capital expenditure as a percentage of revenue has reached forty-five to fifty-seven percent across the largest players. For context, these ratios resemble utility companies or industrial firms, not the asset-light software businesses most analysts use as their mental model for big technology companies. Several major hyperscalers are now projected to see dramatic free cash flow compression in 2026. Alphabet's free cash flow is projected to fall by close to ninety percent this year, according to Pivotal Research estimates. Amazon may turn free cash flow negative. Barclays projects Meta's free cash flow declining roughly ninety percent as well.

This matters because Nvidia sits on the other side of those purchases. The hyperscalers' GPU procurement creates Nvidia's revenue. When entities with total combined capex budgets approaching seven hundred billion dollars are your customers and are described as supply-constrained rather than demand-constrained, the financial protection on your revenue is structural rather than cyclical.

The free cash flow compression at the hyperscalers has introduced something unusual in large-cap technology financing. Several of these companies raised significant debt in 2025 to fund the AI buildout, with aggregate debt issuance across the group reaching around one hundred eight billion dollars. Alphabet's long-term debt quadrupled in 2025. These are companies with decades of cash-funded operations now accessing debt markets at scale. The companies selling equipment to these buyers inherit a degree of demand certainty from the debt commitments made to fund the purchases.

One underappreciated second-order effect involves the energy constraint. Google recently acquired Intersect Power to secure power capacity. Microsoft's eighty-billion-dollar unfulfillable Azure backlog exists because of power limitations at data centers, not because of hardware shortages. Companies that can help hyperscalers secure power, build more efficient cooling systems, or deploy next-generation chips with better performance-per-watt ratios stand to benefit disproportionately. Nvidia's next generation Vera Rubin platform is described as delivering ten times more performance per watt than the current Blackwell architecture, which directly addresses the constraint that is bottlenecking the most significant source of demand in the market.

Custom silicon development by the hyperscalers themselves is the most credible long-term risk to Nvidia's position. Google has TPUs, Amazon has Trainium, Microsoft is developing its own AI accelerators, and Meta is building custom hardware. These are not hobby projects. They involve billions in R&D and significant engineering talent. The constraint is that custom silicon is typically optimized for specific inference workloads at the scale of the company that built it. For the general-purpose training workloads that consume the most compute and command the highest prices, merchant silicon from Nvidia continues to outperform most custom alternatives in the near term. The timeline for custom silicon to become a meaningful substitute at frontier training scale is probably measured in years, not quarters, but it is a real ceiling on the market share Nvidia can hold over a decade-long horizon.

For Microsoft, the constraint is enterprise conversion. Having six hundred twenty-five billion dollars in remaining performance obligations is a striking demand signal, but it is also a reminder that contracted backlog and recognized revenue are different things. The Copilot product line has shown strong adoption metrics, but adoption metrics and billing metrics are not the same thing. Enterprise procurement cycles are slow, budget processes are annual, and AI tool deployment often runs into security reviews, data governance challenges, and integration complexity that extends the time between signing and generating revenue. The margin profile of the AI infrastructure buildout also pressures cloud gross margins in the near term, which Microsoft has disclosed explicitly.

The model layer companies face the most acute constraint, which is the gap between operating losses and the operating environment those losses require to close. OpenAI projects cumulative losses of forty-four billion dollars through 2028. Anthropic's revenue growth is extraordinary but its capital requirements are also substantial, and the company has raised over forty-three billion dollars in total funding in a competitive landscape where no single player has demonstrated durable pricing power. The path from strong revenue growth to strong profitability requires either significant margin expansion through efficiency gains, or successful demonstration that enterprise customers will pay premium prices over time rather than defaulting to the lowest-cost model that meets their requirements.

The geopolitical constraint applies specifically to Nvidia. Export controls on advanced chips to China have removed a meaningful revenue market, and the company has explicitly excluded China data center revenue from its guidance. Chinese domestic chip development is proceeding under significant government investment, and the timeline for domestically competitive alternatives remains genuinely uncertain but not infinite.

The most financially durable position in the current AI cycle belongs to the company at the base of the stack that profits from all capital spending regardless of which model wins, which application captures the most users, or which cloud platform ends up with the largest share. Nvidia generated ninety-seven billion dollars in free cash flow over the last twelve months on capital expenditures of six billion. Its customers are collectively spending close to seven hundred billion dollars in 2026, a meaningful portion of which flows directly to Nvidia's revenue line. Its forward multiple, at roughly twenty-five to thirty times consensus earnings, implies the market expects a substantial deceleration that the current order book and supply constraint commentary does not obviously support.

Microsoft is the most financially credible winner at the platform layer. It has committed revenue, accelerating cloud growth, existing enterprise distribution, and an AI product that attaches to customer relationships rather than requiring new ones. The capital expenditure pressure on free cash flow is real but manageable, and the commercial backlog provides unusual revenue visibility for a company of its scale.

The model layer is where the financial questions are least resolved. Revenue growth at OpenAI and Anthropic has been extraordinary by any historical measure. Anthropic's trajectory from one billion to fourteen billion in annual run-rate revenue in approximately twelve months has no obvious comparable in enterprise software history. But extraordinary revenue growth and durable profitability are different outcomes, and the structural feature of the model market, which is that switching costs are relatively low and competition is intense, creates real uncertainty about how much of that revenue growth will eventually convert to earnings.

The capital currently flowing into AI infrastructure is the largest technology investment cycle in history by most measures. The companies that sit at the point through which that capital must pass, under conditions where supply remains constrained, are the ones with the most legible financial advantages in the near term. The questions about what those advantages look like in five years are real. The questions about what they look like in eighteen months are somewhat less open.

—

This analysis is for educational purposes. It does not constitute investment advice or a recommendation to buy or sell any security. Investors should conduct their own due diligence and consult financial advisors.