US technology and asset-light economics: NVIDIA

Revised practical article; source details and figures are dated to their cited filing or event, with current evidence reviewed through 24 September 2026.

A company can outsource factories and still require enormous amounts of capital. NVIDIA is a useful case because its business is often described as “fabless” or asset-light, yet its filings show material investment in research and development, working capital, supplier commitments, and data-centre infrastructure. The phrase is relative. It cannot replace reading the cash-flow statement and risk disclosures.

NVIDIA’s FY2026 Form 10-K reports revenue of $215.9 billion, up 65% from the prior year, and says growth was driven by its compute and networking platforms for accelerated computing and AI [1]. That is the management-reported growth story. Our owner-earnings question is narrower: after funding the business needed to preserve its competitive position and unit volume, how much earning power might remain for owners over time?

Start with the filing, not a “cash flow” headline

The FY2026 filing reports net income of $120.067 billion, stock-based compensation expense of $6.386 billion, depreciation and amortization of $2.843 billion, and capital expenditures of $6.1 billion [1]. It also reports year-end accounts receivable of $38.466 billion and inventory of $21.403 billion, both substantially higher than their year-earlier balances. The cash-flow statement records a $15.399 billion use from receivables and an $11.324 billion use from inventories during the period, net of acquisitions [1].

These are not reasons to call the business weak. A fast-growing hardware platform can require larger inventory and customer receivables. They are reasons not to equate profit with current cash available to shareholders. Cash-flow quality requires a bridge through accruals, receivables, inventory, payables and obligations. Dechow, Kothari and Watts show that earnings and cash flow behave differently across settings and periods [2]. For NVIDIA, the filed cash-flow statement already includes the period’s reported working-capital movements. If an analyst normalizes those movements, the model should name the normalization and avoid deducting them twice.

The filing also reports $95.2 billion of outstanding inventory purchase and long-term supply/capacity obligations at 25 January 2026 [1]. A large portion relates to inventory purchases. Some obligations may support future growth, some protect the supply chain, and their exact cancellation or settlement terms vary. It would be wrong to subtract the whole balance from one year’s owner earnings as if it were a debt repayment due immediately. It would also be wrong to ignore it. The analyst should trace timing, contractual terms, inventory assumptions, and demand uncertainty.

Reinvestment is not the same as capex

The company spent $6.1 billion on capital expenditures in FY2026 and said it expects that amount to increase in FY2027 to support future business growth [1]. A cross-statement check helps explain the rounded figure: the cash-flow statement separately reports $6.042 billion of purchases related to property, equipment and intangible assets in investing activities, and $101 million of principal payments on those assets in financing activities. Together, $6.143 billion rounds to $6.1 billion at one decimal [1]. Keep the two filed lines separate in a cash-flow reconstruction; the arithmetic is a reconciliation to the rounded MD&A disclosure, not proof that management's measure has an identical definition or that either component is maintenance rather than growth. The report does not label the full amount “maintenance” or “growth.” The owner-earnings estimate needs a separate evidence-based view of facilities, equipment, software, network, engineering infrastructure and other costs required to maintain the existing business.

For a technology company, capital intensity can sit in more than property and equipment. Research and development is expensed under the relevant accounting rules, but future product competitiveness may depend on ongoing research. Manufacturing capacity may be supplied by external foundries, yet non-cancelable supply commitments create a cash and demand exposure. A valuation that subtracts only property capex while assuming the current R&D advantage lasts forever may overstate distributable earnings. A valuation that capitalizes all research spend as maintenance without explaining the assumption may overstate the required reinvestment. Show the range and the supporting evidence.

Stock compensation deserves an owner lens

NVIDIA’s FY2026 cash-flow statement adds back $6.386 billion of stock-based compensation as a noncash adjustment [1]. The statement is correct: issuing shares does not use cash at the grant date. The owner’s cost is dilution over time or the cash spent on repurchases that offset it. The 10-K discloses share repurchases, tax withholding and awards in the shareholders’ equity note. Those are separate mechanics, and the analyst should not call the add-back “free cash.”

Use one internally consistent approach. An earnings-based calculation that retains the reported stock-compensation expense should not add it back again. A cash-flow approach that adds it back should model diluted shares, award expense, net issuance, or the repurchases required to maintain an owner’s proportional claim. Treating an add-back as cash while also freezing the share count creates a quietly optimistic owner-earnings estimate.

Growth concentration is a risk input

NVIDIA reports the Data Center platform as compute and networking infrastructure sold to a range of cloud, hyperscale, enterprise, startup and public-sector customers. The filing says three direct customers accounted for 25%, 18% and 13% of accounts receivable at year-end FY2026 [1]. That is a concentration of the receivable balance; it is not by itself a revenue concentration percentage. Keep those denominators distinct.

The same 10-K identifies demand, product transitions, export restrictions, supply availability, product commitments, and customer spending on data-centre infrastructure as material risks. A valuation should test a slower product ramp, lower gross margin, customer timing, excess inventory, delayed acceptance, and a reduction in supplier commitments. These are business scenarios, not guesses about next quarter’s share price.

What we can and cannot conclude

The filing supports a clear statement: NVIDIA’s extraordinary reported FY2026 earnings came with significant reinvestment and working-capital commitments. It does not, by itself, tell us exactly how much of its spending is necessary to maintain the current competitive position and unit volume. An owner-earnings range requires a multi-year operating model, detailed project and infrastructure context, and an explicit stock-compensation treatment.

The valuation work should then ask what a reasonable return might look like if growth normalizes. Ohlson’s valuation research links market value to book value and expected earnings under an explicit model [3]. That is a useful reminder to reconcile the current earnings base with future reinvestment and the equity claim. A high-quality growth company can still be a poor investment if its price assumes an outcome that the business cannot sustain. This article does not estimate a fair value or recommend a trade.

“Asset-light” describes ownership, not the whole capital system

A fabless chip designer avoids owning the most expensive fabrication plants, but the physical work still happens. The cash requirement has moved across company boundaries into foundry capacity, advanced packaging, memory, testing, board assembly, logistics, and customer infrastructure. A shareholder analysing the designer should ask which obligations are on its balance sheet, which sit in supplier contracts, and which risks appear only when scarce capacity is contested.

Qiao and Wang compare 470 semiconductor companies across 1999–2018 and find that the specialized fabless-foundry model is associated with improved pure efficiency and a reduced effect from semiconductor business cycles in their sample [4]. That is evidence for the economic value of specialization across the industry; it is not evidence that every fabless company is low-risk or that its supply commitments are optional. The study also notes that boundaries between integrated and specialized models can blur. NVIDIA's disclosed purchasing obligations are one issuer-specific reason to trace the operating system beyond owned property.

The practical question is who funds each bottleneck. A foundry may commit wafer capacity; an assembly partner may need advanced packaging capacity; a supplier may require deposits; a distributor may hold inventory; a cloud operator may finance the data centre; and a customer may delay payment until a system is installed. These arrangements affect working capital and the timing of cash even if NVIDIA owns none of the fab or server building. Map the chain from customer order to chip design, wafer start, packaging, system build, delivery, acceptance and collection. Each stage may have a cancellation right, take-or-pay commitment, prepayment or inventory risk.

A growing company can report strong margins and still face a cash squeeze if it pays suppliers earlier than customers pay it. Conversely, a temporary inventory increase may be an intentional buffer against constrained supply and a demand ramp. The analyst should compare inventory composition and aging, receivables by customer, supplier payables, deposits, purchase commitments and subsequent sales. Working capital is not inherently good or bad; the question is whether it supports profitable, collectible sales without exposing owners to an unpriced reversal.

Put the FY2026 figures into a coherent bridge

The FY2026 Form 10-K reports net income of $120.067 billion, stock-based compensation expense of $6.386 billion, depreciation and amortization of $2.843 billion and capital expenditures of $6.1 billion [1]. The arithmetic subtotal of net income plus those two noncash charges less reported capex is $123.196 billion. That number is not operating cash flow and not owner earnings. It has not incorporated reported working-capital movements, taxes already embedded in earnings, acquisitions, supplier obligations, or an estimate of maintenance investment.

The same filing records cash-flow uses of $15.399 billion from receivables and $11.324 billion from inventory during the year, net of acquisitions [1]. A naïve analyst might subtract both from the $123.196 billion subtotal and report $96.473 billion. That calculation is still not a valid owner-earnings figure: accounts payable and other operating balances also changed; total capex contains growth and replacement spending; and the company may need capacity investment to preserve its competitive position. Use the full cash-flow statement and working-capital notes to reconcile the reported period before estimating a normalized year.

A defensible workflow has three passes:

1. Reproduce filed cash. Transcribe reported net income, operating cash flow, cash capex, acquisitions, stock compensation, depreciation, working-capital movements and closing cash. Reconcile the cash-flow subtotals to the filing.
2. Explain the gap. Show how net income becomes cash from operations and how that cash becomes cash after total investment and acquisitions. Label each gap: receivables, inventory, payables, taxes, other accrued balances, purchase of property and equipment, or acquisition.
3. Estimate repeatable owner cash. Normalize a full operating cycle, estimate the renewal investment needed to maintain unit volume and competitive position, and state how expansion capex, supply reservations and dilution are treated.

If the analyst starts with net income, working capital must be estimated once in the owner-earnings bridge. If the analyst starts with filed cash from operations, the year's recorded working-capital effect is already in that figure. Adding it back or deducting it again without a stated normalization double counts. Dechow, Kothari and Watts explain why accrual earnings and cash flows differ across firms and time [2]; the practical response is to reconcile the difference, not select whichever number supports the preferred valuation.

Distinguish research investment from the accounting label

Research and development is expensed in the financial statements, but much of the economic benefit may last beyond the year of spending. Lev and Sougiannis find that capitalizing and amortizing R&D yields value-relevant information in their empirical setting [5]. That result does not mean an analyst should automatically add all R&D back to profit. It means the analyst should ask whether research spending builds a continuing capability, maintains the existing product franchise, or funds uncertain new products.

For NVIDIA, R&D supports architectures, software, developer tools, networking and future generations of accelerators. Some spending may defend current customer demand; other spending may create a new market or product cycle. The public filing does not divide the total into “maintenance R&D” and “growth R&D” in a way that an investor can mechanically apply. Use R&D as a continuing reinvestment requirement in the normalized case. A sensitivity can show the effect of a slower or faster growth rate in R&D relative to sales, but do not add current R&D back to owner earnings while assuming that the product advantage continues for free.

A simple analyst schedule can track annual R&D, R&D as a share of sales, new product introductions, gross-margin shifts, product mix, software ecosystem activity and the time from development to commercial contribution. The schedule should not claim that a patent count or launch announcement proves economic return. Look for evidence that engineering expense turns into customer adoption, repeat deployments, pricing power, or lower cost per delivered unit. If the investment is large but those outcomes weaken, the proper conclusion may be lower expected returns on future R&D, not a different accounting add-back.

Treat purchase commitments as a timed exposure

The filing reports $95.2 billion of outstanding inventory purchase and long-term supply or capacity obligations at 25 January 2026 [1]. This is a large exposure that deserves a schedule, but it should not be subtracted wholesale from one year's owner earnings. Some commitments correspond to inventory that will be sold; some reserve capacity needed to meet customer demand; timing and cancellation rights differ; and the corresponding revenue may arrive in a later period.

For each disclosed commitment, or each best-available category, record the amount, expected payment period, related asset or supply, cancellation or rescheduling rights, likely customer revenue, deposit paid, and downside recovery. Compare the schedule with inventory and demand. A three-month obligation to buy components that can be turned into a shipment under a signed order differs from a long-dated noncancelable capacity reservation tied to an uncertain market forecast.

The stress case should combine events that can happen together. If end customers slow deployments, inventories may age, orders may be rescheduled, suppliers may still expect payment, and receivables from delivered systems may stretch. Do not stress each line independently and then add them in a way that assumes every risk is unrelated. Show what would happen to cash and margins if utilization fell, lead times normalized, prices fell, or a customer delayed acceptance.

Model stock compensation as a claim on per-share cash

The $6.386 billion of reported stock-based compensation is a real component of employee remuneration even though the cash-flow statement adds it back as noncash [1]. There are two coherent ways to handle it. Keep the expense in earnings and use a diluted share count consistent with outstanding awards. Or add the expense back and explicitly model future shares or the cash repurchases required to offset dilution. Do not add it back, assume the share count is constant, and call the resulting number an owner-level cash estimate.

Li and Wyatt study employee stock options in equity valuation and show that ignoring dilution can distort the valuation of option grants, with larger bias in firms that are R&D intensive or heavy users of options in their sample [6]. For an investor, inspect annual grants, vesting, exercises, forfeitures and net share settlement. Compare gross shares issued with buybacks, not just the net share count. A large buyback can keep diluted shares roughly flat while returning less cash to owners than the headline repurchase amount suggests.

A per-share worksheet should include basic and diluted weighted-average shares, period-end shares, options and restricted units, expected future grants, repurchases, and any acquisition consideration payable in shares. Use the share count period that matches the earnings numerator. Run a case where compensation stays at the current revenue ratio and another where it rises or falls based on hiring, share price and grant practice. This makes clear whether apparent growth in total owner earnings translates into growth for each continuing shareholder.

Translate platform strength into testable scenarios

“AI demand is durable” is an investment thesis, not a model input. Translate it into customer deployments, units, average selling prices, gross margin, software or networking attachment, receivable days, inventory, supplier capacity, R&D, and renewal capital. For each driver, record the historical range, current evidence and a plausible downside. Then ask what combination of market growth and share retention is needed to support the price.

The downside case should not assume that R&D disappears when sales weaken. Maintaining engineering talent and software support may be necessary to recover competitiveness. Nor should it assume that every supplier commitment remains fixed; contracts may allow cancellations or rescheduling. Separate what the filing says from the analyst's probability judgment.

Valuation should follow the per-share owner-earnings range. Estimate a normalized base and downside, connect long-term growth to reinvestment and competitive durability, and test the market price against both. Ohlson's earnings and book-value framework is a reminder that valuation depends on future economics and accounting claims rather than a single cash-flow headline [3]. The analyst need not publish one exact intrinsic value if the range is driven by unresolved supply, maintenance or customer-demand evidence.

What should refresh the view?

Set concrete review triggers before the next filing: an unusual receivables increase relative to sales, inventory aging, change in gross margin, a material shift in purchase commitments, a large change in R&D intensity, slower customer acceptance, or diluted shares rising faster than earnings per share. A trigger should prompt a document review, not an automatic buy or sell. The workbook should keep the last reviewed date next to each assumption, and the analyst should state which new disclosure would narrow or widen the range.

The practical conclusion is straightforward: NVIDIA's FY2026 reported figures show substantial profit and a rapidly expanding operating footprint, alongside large working-capital uses, continuing research needs, stock compensation and supply obligations [1]. The business may create considerable value. To estimate what owners can claim, an analyst must model the full network of reinvestment and contractual commitments, then divide the result by the shares that will actually participate.

Read demand through customer cash, not announced investment

A large cloud-provider capital budget is not the same as a purchase order for NVIDIA products, and a purchase order is not the same as a collected invoice. Build a demand ladder with separately labeled evidence: customer capex guidance, purchase commitments reported by customers, orders acknowledged by suppliers, shipment, installation, acceptance and collection. The farther the evidence sits from collected cash, the more assumptions remain.

For the largest customer relationships, examine concentration where the filing discloses it, payment terms, receivable balances, geographic mix, distributor use, returns and customer financing arrangements. A customer may be creditworthy but still delay acceptance because a cluster is not ready, electricity is unavailable, or software integration is incomplete. The supplier may have shipped and recorded revenue before cash arrives. The analyst should compare sales growth with receivables and cash collections, and check whether any increase results from mix, timing, contract terms or a change in channel inventory.

The scenario should not turn a single customer concentration percentage into an automatic alarm. Large enterprise customers can have strong credit and high repeat demand. Concentration matters because a change in spending, deployment timing or negotiation power can affect volumes and margins quickly. A useful downside asks what the company would do with inventory and capacity if one or two customers paused deployments for a quarter: which products can be redirected, how quickly, at what discount, and with what supplier payment still due?

A decision table that connects evidence to action

At each annual and quarterly review, keep a compact table that tells the analyst what to reopen:

A red flag is not a conclusion. An inventory build can precede a major launch; a working-capital release can reflect healthy collections or supplier stress. The analyst should write the competing explanations and identify the next filing disclosure, customer evidence or management answer that would distinguish them. Then update the range instead of keeping a stale point estimate.

Workbook map

The Project 2 workbook preserves NVIDIA FY2026 net income, stock-based compensation, D&A, total capex, receivable and inventory cash uses, ending balances and disclosed purchase/capacity obligations in Filing Inputs. Evidence Vectors records the 25 January 2026 fiscal year-end separately from the 25 February 2026 SEC filing date and maps each input to its workbook row; page/note locators not yet transcribed remain identified as missing. Assumptions selects whether SBC remains expensed or is added back with dilution, and leaves maintenance investment, normalized working-capital reinvestment and diluted shares blank until supported. The NVDA Bridge therefore remains unavailable until the two required cash estimates are entered; its DCF also requires a dated price and share count. The new Profile Stress row applies a clearly hypothetical 20% cash shock and higher inventory/capacity and priority-claim indices; Appendix 04 reproduces it and warns that the index is not an NVIDIA forecast.

Technical appendix

• Appendix 04 — Listed-company profile stress cases

Technical appendices

The worked calculations and source boundaries below extend this article's examples. They keep assumptions and unresolved inputs visible so readers can reproduce the steps without mistaking an illustration for an issuer or sector estimate.

• Appendix 02 - Technology, working capital and asset-heavy reinvestment

References

[1] NVIDIA Corporation, Form 10-K for the fiscal year ended January 25, 2026, filed February 25, 2026, https://www.sec.gov/Archives/edgar/data/1045810/000104581026000021/nvda-20260125.htm.

[2] P. M. Dechow, S. P. Kothari, and R. L. Watts, “The Relation Between Earnings and Cash Flows,” Journal of Accounting and Economics, vol. 25, no. 2, pp. 133–168, 1998, https://doi.org/10.1016/S0165-4101(98)00020-2.

[3] J. A. Ohlson, “Earnings, Book Values, and Dividends in Equity Valuation,” Contemporary Accounting Research, vol. 11, no. 2, pp. 661–687, 1995, https://doi.org/10.1111/j.1911-3846.1995.tb00461.x.

[4] G. Qiao and Z. Wang, “Vertical Integration vs. Specialization: A Nonparametric Conditional Efficiency Estimate for the Global Semiconductor Industry,” Journal of Productivity Analysis, vol. 56, pp. 139–150, 2021, https://doi.org/10.1007/s11123-021-00616-x.

[5] B. Lev and T. Sougiannis, “The Capitalization, Amortization, and Value-Relevance of R&D,” Journal of Accounting and Economics, vol. 21, no. 1, pp. 107–138, 1996, https://doi.org/10.1016/0165-4101(95)00410-6.

[6] Y. Li and S. Wyatt, “Employee Stock Options, Equity Valuation, and the Valuation of Option Grants Using a Warrant-Pricing Model,” Journal of Accounting Research, vol. 43, no. 1, 2005, https://doi.org/10.1111/j.1475-679X.2005.00164.x.

Companion resources

The workbook and technical appendices are available with research membership. The video production guide maps this article to a planned episode; a published video is not yet available.