NVIDIA’s growth is spectacular. The harder question is what customers earn from it.

Separate evidence from conclusions

A company-research worksheet, not a forecast.

The reported development

NVIDIA’s second-quarter fiscal 2027 announcement reported $96.2 billion in revenue for the quarter ended 26 July 2026, up 18% sequentially and 106% year over year. The release is dated 26 August. Those are company-reported figures; they demonstrate supplier growth, not the return earned by every buyer of that infrastructure.

The distinction matters because the AI investment chain contains multiple businesses. A chip supplier sells equipment, a cloud operator sells capacity, and an application provider sells an outcome to an end customer. Strong demand at the first stage can coexist with uncertainty about margins at the last.

Follow the conversion from equipment to useful work

Our research framework separates four steps: equipment ordered, capacity commissioned, capacity used and output paid for. These stages are not interchangeable. An announced purchase is not proof that a data centre has power, cooling and networking ready; an operational cluster is not proof of profitable utilisation.

For each company update, write down which stage the announcement actually supports. A delivery milestone reduces a different uncertainty from a contract with an end customer. Avoid adding the same commitment at several stages and presenting it as additional demand.

The questions a company profile cannot answer alone

Ask how concentrated the customer base is, how quickly customers can deploy equipment and how financing affects their obligations. Then ask whether revenue depends on training large models, serving existing users or a mixture. These are diligence questions, not claims that a particular customer is in difficulty.

Software can change the economics in both directions. More efficient inference can lower the compute needed per task, while lower prices can encourage more tasks. Neither outcome can be inferred from a hardware shipment figure alone. A useful research update identifies which evidence would distinguish these possibilities.

A watchlist with an explicit uncertainty

Build a quarterly evidence ledger containing reported revenue, disclosed margins, delivery commentary and customer deployment statements. Keep management forecasts separate from completed results. Link each entry to a dated filing or official release, and resist the temptation to fill missing utilisation data with a precise estimate.

The investment question is not simply whether AI grows. It is which participant captures durable returns after equipment, energy, financing and model-serving costs. This feature offers a way to interrogate that question; it does not supply a valuation target or a recommendation to buy NVIDIA.

Map the chain from an announcement to recurring use

An AI infrastructure announcement often combines several stages that deserve separate scrutiny. A proposed project, a contracted order, installed equipment, an operating service and a satisfied repeat customer are different pieces of evidence. Our analysis framework is to mark the stage a statement actually supports, then identify the next document needed to move the assessment forward.

For a proposed facility, look for details about its intended scope and dependencies. For an equipment purchase, establish who buys from whom. For operation, look for evidence that the service is available and serving the stated workload. For recurring demand, ask what would demonstrate continued use. Do not treat a partner logo as proof of all five stages. This is a proposed due-diligence sequence, not an allegation about any named project.

The distinction matters when reading NVIDIA coverage because readers may be evaluating very different things: chip demand, a cloud operator's business, a customer's application or a listed security. A positive observation at one stage can be relevant without answering the whole investment question. Make the unit of analysis explicit before drawing a conclusion.

Worked example: two imaginary AI service businesses

Consider two fictional operators that each report $100 of revenue from an AI service. Operator A spends $60 delivering it and B spends $85, leaving $40 and $15 respectively before other costs. These numbers are deliberately invented to illustrate the method; they are not NVIDIA results or estimates for any company. Equal revenue can coexist with very different room to cover the rest of the business.

Now suppose A's revenue comes from a single temporary project while B's comes from customers who renew. That additional fact could change your interpretation, but it still would not establish profitability, cash generation or future growth. You would need to inspect contract duration, collection, replacement spending and other expenses. The useful habit is to ask which missing variable might reverse the apparent conclusion.

Applied to real company research, this means putting each number next to its definition, period and source. Keep reported results separate from our calculations and management expectations. If a company does not disclose a particular metric, say so. Inventing an AI-specific margin from broad company totals would create precision without the evidence needed to support it.

A company research worksheet with four questions

First: who is the paying customer? Identify the direct buyer and avoid assuming that buyer's customers have the same economics. Second: what is being sold? Hardware, software access, capacity and consulting can imply different obligations. Third: what evidence supports repeat demand? A new order is useful information, but renewal, continued use and expansion answer other questions.

Fourth: what constrains delivery? List only constraints supported by the specific sources, then mark any broader scenario as a hypothesis to investigate. For example, asking whether an infrastructure plan depends on power availability is a research question; claiming that a particular project is delayed for that reason requires evidence about that project. This distinction prevents a plausible sector narrative becoming a fabricated company fact.

Keep the worksheet short enough to revisit after the next filing. One column should contain the disclosed fact, one its source and date, and one the unanswered question. Add an interpretation only after completing those columns. Readers should be able to challenge the inference without having to reconstruct the evidence from a collection of headlines.

How to read the next announcement without overreacting

When a headline arrives, classify it first: historical result, contracted activity, product availability, partnership or forward-looking target. Then compare it with the previous disclosure using the same measurement and period. An increase in one category cannot automatically replace missing evidence in another. In particular, a large future target should remain labelled as a target.

Use a two-sided conclusion. Explain what the announcement strengthens, and which uncertainty it leaves unresolved. The purpose of this company series is to help readers understand how AI businesses work. It does not assign a share-price target, claim access to private contracts or recommend a transaction. A defensible article can finish with a sharper unanswered question rather than an exaggerated certainty.

Sources & method

Linked sources support the documented facts. Evaluation frameworks are CryptoXAI analysis; hypothetical examples are labelled. No hands-on benchmark results are claimed. An expansion date records an editorial change, not a new verification of every source.

• NVIDIA Q2 fiscal 2027 results, 26 August 2026 ↗