NVIDIA: Who Pays For the Next Wave?

Evidence & Assumptions

Detailed support for the same dated report. Read Disclosures and Limitations.

Amounts are US dollars unless stated otherwise. This record presents the financial evidence, scenarios, business forecasts and rating rationale. Valuation dates, evidence cutoff, forecast windows and eventual publication date remain distinct.

Why 4/10 — Trim

Decision: 4/10 — Trim. At $222.27, reduce direct NVIDIA exposure in favor of SPY while allowing a smaller holding within a diversified approach. Our judgment is that SPY offers the preferable balance of prospective return and risk. NVIDIA's growth and competitive strengths are exceptional; the evidence supporting the price's long-duration cash requirements is less secure. We find no adequately supported expected advantage for taking this much company-specific customer-funding and platform risk. That is a qualitative allocation judgment, not a measured forecast of underperformance.

Dates and horizon. Reference price: September 18, 2026 regular close, $222.27 [S001]. The latest NVIDIA financials end July 26. The ownership horizon follows the ten-year operating case through September 2036, with continuing value beyond it. The next two reporting cycles test the mechanism; the unchanged 90-day evaluation does not turn the DCF into a short-term target. No later events or price observations are backdated into this decision.

Why SPY Is the Baseline

SPY spreads exposure across S&P 500 companies and still participates in NVIDIA's success: the retained September 16 holdings record shows 7.96% NVIDIA and 38.03% information technology. [S006] Those are dated weights, not fixed future allocations. Choosing SPY reduces direct dependence on NVIDIA's margins, recoveries and replacement cycle; it does not remove AI, technology, valuation or market risk. Other index constituents also finance or depend on the AI buildout. The advantage is partial diversification, not an independent hedge against an AI downturn.

We compare USD total returns over the same ten-year ownership horizon, including distributions and consistent corporate-action treatment. SPY's ongoing fund expenses reduce its investor return; personal taxes and trading costs are excluded from the modeled comparison and can matter to actual reallocations. The record contains no sufficiently developed forward SPY valuation to claim a numerical expected return spread. The 11% NVIDIA discount rate is an analyst hurdle, not SPY's forecast return. SMH and hypothetical cash remain supplemental comparisons; SPY controls this rating.

What Earns Weight in the Decision

NVIDIA's reported trailing sales of $302.970bn and operating profit of $197.579bn show an extraordinary existing franchise. Eleven comparable quarterly revenue outcomes exceeded management's original midpoint, and the supply-chain and customer record support further growth. These are stronger evidence than a future market-size headline. [S011] [S012] [S015] [S027]

The price nevertheless represents about 45 times trailing cash after an illustrative stock-compensation charge. The middle case requires first-year sales of $590bn and year-ten sales of $1.270tn at a 50% operating margin, yet gives approximately $190 at the retained 11% hurdle. Nearly half the operating value comes after year ten; its original reinvestment convention implies roughly 48% incremental capital returns. The coherent addition case reaches $235, but falls to $224 or $218 when continuing investment earns 20% or 15%. The sensitivity removes most or all of its price cushion without assuming that NVIDIA stops being a successful business. Evidence: valuation · terminal capital · addition capital.

The case to trim therefore rests on more than the middle value being below the quote. NVIDIA is extending customer credit and taking correlated ecosystem exposures while actual competing deployments challenge share and pricing. Public evidence does not substantiate enough of the long-term customer recovery and retained system economics to offset those demands. Missing support for those assumptions weighs adversely, rather than merely leaving the old score unchanged.

Missing Evidence Has an Adverse Weight

This describes evidence absent from the reviewed public record, not proof that internal records do not exist, a disclosure violation or misconduct. Some future outcomes cannot exist yet; that limits how much of the long-term upside we can justify paying for now. Unknown remains unknown as a factual finding, but it weighs against ownership in the rating. Evidence: financials · customer funding · mature returns · replacement evidence · support model.

The Strongest Reasons to Keep a Smaller Position

• Demand is real, and NVIDIA has repeatedly executed. Paid adoption, the CUDA/software ecosystem, cluster reliability and supply expansion support further profitable growth. Falling compute cost can stimulate useful demand faster than efficiency reduces equipment needs. The $331 favorable case remains a reasonable scenario, not dismissed because its probability is unmeasured. [S015] [S027] [S066] [S116] [S117]
• Customer economics are not uniformly poor. CoreWeave reports mid-20s contribution margins on fully ramped contracts, including depreciation, and selected A100/H200 renewals support continuing usefulness. These observations deserve more weight than blanket claims that equipment becomes worthless. They still omit representative owner cash after corporate overhead, financing and full investment recovery. [S147] [S148] [S151] [S157]
• Funding resilience can bridge expansion. Microsoft's cash surplus, major customers' resources, OpenAI's $122bn committed round and borrowing capacity weaken an immediate funding-crisis thesis. External finance can fund sound growth. Nscale's prospective Anthropic contracts also support future NVIDIA demand, though required deployment financing was not binding in the retained filing. [S136] [S137] [S139]–[S146]
• A rival GPU need not eliminate NVIDIA's economics. Microsoft's AMD systems include NVIDIA networking; NVLink Fusion extends that route. Groq technology and broader systems can defend inference demand. Lost sockets, retained revenue and retained profit are different. The amount and margin of that defense are not yet established across a fleet. [S096] [S163]–[S165]
• The balance sheet can absorb setbacks. Cash and marketable debt, ongoing operating cash and a pause in buybacks provide a substantial near-term cushion. The six-month stress still leaves about $52bn after a full call on the existing $3.5bn guarantee. This does not cover every later obligation, but rules out treating all gross commitments as immediately payable debt. [S011]
• The valuation objection is serious. At a 9% hurdle, the unchanged middle case is about $255; the developed addition case is $235 at 11%, before tougher continuing-capital tests. A lower rate or more durable margins could make trimming costly, while SPY can also be expensive and technology-heavy. The required return is a judgment, not a measured law. These alternative valuations and benchmark risks could make trimming the wrong decision.

These arguments prevent a robust case for 3 — Reduce or an extreme rating. They justify continued smaller ownership despite our preference to reduce the concentrated exposure at the stated price.

Resilience and How Value Could Reach Owners

The linked customer stress gives approximately $124 even with the first two middle years intact: later sales and margins weaken, working-capital release is blocked, investment recoveries disappear and net support consumes cash. The original $76 setback is earlier and more severe. These are scenario values, not forecast prices or floors. Power delays, memory costs, Taiwan concentration, export controls and adaptation spending can interact with customer funding. We do not subtract the $105bn guarantee cap as a certain present loss. Evidence: environment · operating bridges · support model.

Value reaches owners if customers accept and pay for systems, NVIDIA retains profit after ordinary reinvestment and net support, and that cash supports distributions, net share reduction or productive retained investment. The disclosed dividend and buybacks provide actual routes, subject to board choices; stock compensation and strategic spending consume part of cash. The DCF values cash capacity, not an entitlement to immediate distributions. A future sale can recognize improved cash prospects, but neither the ten-year model nor a product launch guarantees a date of price convergence. Continuing-capital assumptions matter because much of the value lies beyond the explicit decade.

Why Four, and What Would Move It

Why not 5 — Consider Other Opportunities? The ownership concern is actionable: substantial customer credit and contingent ecosystem support coexist with a price that needs long-duration extraordinary economics. Missing mature cash-recovery and recovery/overlap evidence directly weakens those assumptions. Given SPY's partial diversification while retaining NVIDIA exposure, the balance supports trimming rather than only withholding new purchases until another opportunity appears.

Why the distinction matters. Lack of a buying advantage is not sufficient by itself to justify four. The affirmative case is the combination of demanding long-term cash economics and observable expansion of credit and support whose recovery depends on related customers and demand. Rapid growth can explain rising receivables; investment-grade terms are not evidence of late payment, and expansion funding deficits do not prove unprofitable mature deployments. Those favorable explanations were weighed alongside profitable contracts, renewals and strong NVIDIA earnings. We still modestly favor SPY's broader earnings base at this price, but partial diversification alone would not establish the score. This is a security-level judgment, not an assumption that a reader has an oversized position.

The strongest objection remains valuation sensitivity: the unchanged middle case is about $255 at 9%, versus $190 at 11%, and the favorable path gives $331 at 11%. These conditional values do not establish which assumptions will occur. A five is a defensible competing interpretation if one gives more weight to durable margins and financing resilience. We retain four because customer-credit and support dependence weaken the price's demanding cash assumptions; we cannot claim quantified expected underperformance of SPY.

Why not 3 — Reduce? The unfavorable comparison is not robust enough under reasonable competing assumptions. Actual profits, funding capacity, renewal evidence, retained networking content, the $235 addition case and the $331 favorable case can support materially better outcomes. The discount-rate judgment can change the valuation. The record does not establish widespread defaults, disappearing platform value or poor returns across all customer cohorts. It therefore supports a limited reduction, not a confident larger one or exit.

Toward 5, then 6 — Consider Buying: credible representative collections and customer returns, repeat purchases supported by earned cash, resilient system gross profit and demonstrated net support recovery would first remove the case to trim. Buying then requires a supported expected SPY-relative advantage and adequate price resilience. The existing approximately $170 review level is a sensitivity under unchanged economics, not a target or automatic order.

Toward 3 — Reduce or lower: persistent late collection, funding gaps, unrecovered support, weaker renewal economics or commercially meaningful losses of workload and system profit would strengthen the adverse comparison. One forecast miss does not assign a score. Improving near-term funding also does not by itself establish a decade of customer-funded growth.

4/10 · Trim

NVDA · $222.27 Regular close · 18 September 2026

Trim direct NVIDIA exposure at the stated reference price. Today's price assumes customers can keep earning enough from AI to fund new systems, and that NVIDIA can sustain growth without much heavier investment. The public evidence does not yet establish either. Exceptional profits, funding and platform strengths support a smaller diversified holding.

Investment stance 4 out of 10 — Trim. The balance favors reducing direct exposure, while continued ownership of a smaller position remains reasonable.

Four favors reducing exposure; it is not an exit judgment, a probability or a fixed portfolio weight.

Why now Competing chips have real deployments. NVIDIA can retain system value, but the price leaves little room for weaker margins or costly customer support.

What could prove us wrong Profitable adoption and richer systems could justify adding even at our 11% requirement. Funding progress and repeat GPU use give this case substance.

What changes the view Representative customer cash returns, better collections and resilient system profit could remove the case to trim. Persistent credit deterioration, unrecovered support or sustained workload losses would favor a stronger reduction.

Our measurement window 90 calendar days from actual publication; research revisited roughly quarterly. Long-term values are conditional, not 90-day targets.

A device business became a computing platform.

NVIDIA designs processors, networking and computing systems, together with software that makes them useful. Founded in 1993, it first built its business around graphics. CUDA, introduced in 2006, made its processors accessible for general-purpose parallel computing. Deep learning greatly enlarged that opportunity. Mellanox, acquired in 2020 for $6.9bn, strengthened the networking required to connect many processors. Together, these developments made large computing systems central to the business. P03 [S003], P04 [S004].

Customers buy chips, boards, complete systems, networking and software or services. NVIDIA outsources wafer manufacturing and much assembly. Fabrication capacity, memory and packaging are therefore purchased inputs, even though the company does not own the foundries. Revenue is earned when the relevant contractual delivery or service obligation is met. The company’s order is not necessarily the final cloud user’s profitable project.

A common share is a residual ownership claim after obligations to employees, suppliers, creditors and other claimants. It provides no right to receive a fixed percentage of current free cash flow on demand. The board decides distributions and reinvestment; market price can diverge substantially from an investor’s estimate of value. Repurchases reduce shares only to the extent they exceed new issuance and other changes. NVIDIA’s 2024 ten-for-one split changed units, not each holder’s proportional ownership. F01 [S011], F02 [S012].

The questions that can change the decision

1. Does paid, productive use finance repeat purchases? This distinguishes durable demand from construction supported indefinitely by new capital.
2. Does cash catch up with sales? Receivables, inventory, customer terms and investment gains explain why revenue and net income are insufficient by themselves.
3. Can NVIDIA retain exceptional economics through new product cycles? Alternative silicon, memory costs, power and customer bargaining determine the answer.
4. What future success does today’s price require? A strong business can be an unattractive incremental purchase at a sufficiently demanding price.

Activity classification and its limits

Under TK Investment Research’s activity taxonomy, the company coverage label is OTHER-03: disclosed activities cannot be separated sufficiently to prove a 70% revenue threshold . This describes a classification limit, not a finding that NVIDIA’s business is unknowable. Its identifiable device design and related support fit TEC-03 ; separately sold developer and infrastructure software fit TEC-02 ; identifiable end-user engineering applications fit TEC-01 . Data Center is a customer market containing several products, so its 92.5% quarterly share does not establish any one activity’s share. F01 [S011], F02 [S012].

Economic activities, applicable methods and unresolved boundaries

Materiality remains unknown for undisclosed software, services and purchased system content; no missing percentage is labeled immaterial. Device and platform risk is material to the entire revenue base. Customer support is material by potential effect regardless of recognized revenue. The financial chapter reconciles consolidated results. We apply software and fabless tests together to the evidence, while valuing the consolidated cash flows once.

For the uncovered activities, the economic questions are concrete: who owns the infrastructure; who pays during low utilization; whether another customer can replace committed demand; when a guarantee becomes payable; and how much value survives if the customer fails. Public contract descriptions establish some terms but do not permit a reliable expected-loss estimate. That uncertainty reduces conviction in a precise value and supports limiting exposure.

Financial reconstruction and capital discipline

GAAP means standard U.S. accounting rules. Gross profit deducts product costs from sales; operating profit also deducts research, selling and administrative costs. Operating margin is that profit per dollar of sales, before financing and income tax. Cash flow differs because customers pay and the company invests on different schedules from accounting recognition.

NVIDIA's operating profitability is exceptional. Its funding model now deserves as much attention as its income statement. Investment gains increase reported earnings, longer customer terms delay collections, and strategic investments and guarantees tie more NVIDIA capital to the ecosystem buying its products. These facts can coexist with real demand. They do not establish fabricated sales or customer defaults.

Three controls govern this reconstruction: financial investments are separated from cash; investment gains are separated from operating earnings; and contractual commitments are separated from current debt and expected losses. Amounts are in US dollars, millions unless a table says billions. The latest financial period is the quarter and half year ended July 26, 2026, released August 26. Fiscal years end in January of the named year. Source references F01–F37 identify the documents, dates and actual reading scope in the source appendix.

Five years, the current half year and a comparable trailing period

The five-year numerical series comes from the comparative audited statements in the FY24 and FY26 annual filings. The current half year and its comparable prior period come from the latest 10-Q. The trailing twelve months, or TTM, are calculated as FY26 plus H1 FY27 less H1 FY26. TTM is a historical measurement, not a forecast or an annualization of the most recent quarter. Pre-June 2024 shares are adjusted for the ten-for-one split. F01 [S011], F02 [S012], F03 [S013].

Consolidated earnings and cash generation, $m

Free cash flow equals operating cash flow less the two equipment/intangible rows. It does not deduct acquisitions, equity purchases, dividends or all payments for acquired technology. The FY26 Groq cash payment of $13,000m is separate from the acquisitions row; another $2,944m appears in H1 FY27 financing cash flows. Consequently, reported FCF does not measure cash remaining after all strategic investment. FY23 stock compensation is $2,709m in the cash-flow statement versus $2,710m in the expense note; the reported distinction is retained. FY22 and FY23 pre-split diluted EPS of $3.85 and $1.74 become $0.385 and $0.174 after the split.

Cash reconciliations, $m; each equation balances

Twelve quarters and the management forecast record

This table uses GAAP results and the preceding quarter's original revenue guidance midpoint, rather than analyst consensus or revised estimates. Original releases F05–F16 are the controlling sources. All listed guidance bands were ±2%. The eleven comparable outcomes exceeded their midpoint; that is a historical execution record from an unusually strong growth period, not a forecast of continuing beats.

Quarterly results and preceding revenue guide, $m

The latest two quarterly revenue figures sum to $177,836m, $1m below filed H1 revenue of $177,837m. Quarterly net income likewise sums to $118,009m versus filed $118,010m. The TTM uses the filed half year; no fabricated adjustment closes the rounding difference. Summing rounded EPS would introduce a further measurement error.

The forecast record includes exceptions that revenue beats alone conceal. FY25 Q1 call guidance envisaged Blackwell production shipments in Q2 and a Q3 ramp; the next call's mask-change discussion moved production shipping to Q4, when Blackwell eventually delivered $11b. Before the H20 restriction, FY26 Q1 GAAP gross-margin guidance was 70.6% ±0.5 percentage point; actual margin was 60.5% after a $4.5b inventory and purchase-obligation charge. FY26 later recovered to a 75.0% fourth-quarter GAAP margin. F26 [S036], F25 [S035], F23 [S033], F10 [S020], F07 [S017].

Management outlook at August 26, 2026; not research assumptions

Source: F05 [S015] outlook and F17 [S027] prepared remarks pp 3,7. Management's longer forecast horizon arrived with a lower margin premise. A research case that adopts the growth but retains the old margin needs an explicit independent justification.

Business mix and what growth does not disclose

Reported market-platform revenue, $m

FY26's $85,441m revenue increase reconciles to Data Center $78,551m, Gaming $4,692m, Visualization $1,313m, Automotive $655m and OEM $230m. Data Center supplied about 92% of the increase and rose from 39% of FY22 revenue to 90% of FY26 revenue. This is a mix bridge. Disclosed data do not allow a reliable separation of units, realized price, networking attachment, rack content and application utilization. Equal revenue growth and GPU-unit growth must not be assumed. F02 [S012] Note 16; F03 [S013] Note 17.

FY27 market reporting changes to Data Center and Edge Computing. Q2 Data Center $89,023m equals Hyperscale $48,710m plus ACIE $40,313m; Edge adds $7,198m to reach $96,221m. ACIE means AI Clouds, Industrial and Enterprise. One company moved from ACIE to Hyperscale in Q2 after a business-model change, with comparatives recast. Edge cannot simply replace Gaming in an unchanged historic series.

The GAAP reportable segments are a different grouping: Compute and Networking revenue $88,299m plus Graphics $7,922m. Their Q2 operating profits of $62,696m and $3,899m total $66,595m, exceeding consolidated operating profit by $2,861m. That difference includes $2,027m stock compensation and $834m other unallocated costs. Omitting these costs would overstate consolidated economics. Segment assets are not disclosed in a form supporting reliable segment returns on capital. F01 [S011] Note 13.

Earnings quality and the cash-collection test

H1 FY27 net income to operating cash flow, $m

H1 operating cash flow was 63.1% of net income, versus 85.6% for FY26. Q2 alone was 40.3%, partly reflecting the timing of two estimated federal tax payments, compared with generally none in Q1. That seasonality does not erase the substantive receivable build. Receivables increased from $38,466m in January to $63,059m in July. Five direct customers accounted for 70% of the ending balance. NVIDIA describes 90-day to one-year terms for certain investment-grade customers making large data-center deployments. Subsequent collection and changes in aging are the critical tests; a higher receivable balance alone does not demonstrate misconduct. F01 [S011] cash-flow statement, Note 7 and MD&A.

Inventory of $31,575m comprises $11,341m raw materials, $13,377m work in process and $6,857m finished goods. Finished goods declined from $8,774m at year-end while inputs and work in process increased. That is consistent with procurement and production expansion, but it does not establish final demand. H1 provisions of $1,600m for inventory and $501m for excess purchase obligations remain economic costs. Customer advances rose from $160m to $2,800m, a cash benefit much smaller than the receivable outflow.

Deferred revenue also reconciles: opening $2,572m + additions $17,709m − recognized revenue $13,869m = ending $6,412m. Current $4,616m plus long-term $1,796m equals the same ending amount. These balances are obligations to provide future goods or services, not additional current-period earnings.

Normalization rules

Sources: F01 [S011] Notes 1–11; F02 [S012] accounting policies and audit opinion; F05 [S015] non-GAAP reconciliation. No privately negotiated customer contract, detailed aged receivable schedule or subsequent collection ledger was available in the public documents reviewed.

Capital structure, shares and financial assets

Balance-sheet development, $m

FY22 balances come from the January 30, 2022 comparative column of the FY23 annual filing, p56 [S102] (also source GH06). FY22 equity reconciles as assets $44,187m less liabilities $17,575m = $26,612m.

The table's historical marketable-securities category contains different security types. At July 26, cash $22,443m plus marketable debt $34,143m provides $56,586m of relatively liquid resources; after debt carrying value, the conservative net-cash measure is $23,220m. Total public equities are $47,740m : $42,783m in marketable securities plus $4,957m in Other assets. The latter is locked through December 2027; the filing also identifies $36.9b subject to short-term sale restrictions. Nonmarketable securities are another $51,157m. These investments have different valuation and liquidity risks and can weaken when ecosystem customers need support. The $3,259m equity-method infrastructure-financier balance is inside nonmarketable securities, not an additional asset. About $1.7b of overseas cash/debt securities could trigger unaccrued foreign or state taxes on repatriation. F01 [S011] Notes 5–6 and MD&A.

Debt reconciles from $8,468m opening carrying value + $24,896m net proceeds + $2m inferred amortization = $33,366m. Face principal $33,500m less $134m unamortized discounts and issuance costs yields the same total. The $2m is an arithmetic inference, not a separately extracted payment. June issuance was $25b across seven tranches. A $25b commercial-paper program was undrawn; unused issuance capacity is not cash or a committed bank facility. Existing notes have nonfinancial covenants, with compliance reported at the January year-end. The new indenture was not independently contract-read; no numerical covenant-headroom claim is made.

Debt principal reported at July 26, 2026 and contractual coupons; maturity calendar years, $m

Sources: F01 [S011] Note 9; F02 [S012] debt disclosure. This is the July 26 schedule. The $1b 2026 tranche matured September 16, before the reference stock quote; subsequent repayment was not independently verified. It is not asserted to remain outstanding on September 20. An ordinary $1b cash repayment reduces both cash and principal equally and does not change net cash. Notes are senior unsecured. The approximately $31.4b July market value of debt is distinct from contractual principal repayment.

H1 FY27 shares and compensation reconciliation, millions of shares

H1 cash repurchases were $39,044m, compared with $39,845m in the equity accounting roll-forward; settlement and excise-related amounts prevent treating these as the same cash measure. The accounting amount divided by 203m shares implies roughly $196 per share for the period, not today's price or evidence of attractive value. Unrecognized award compensation was $19.4b over about 2.6 years. At current ending shares, the new $0.25 quarterly dividend would cost approximately $24.1b annually, subject to declarations and share changes. Historical FY26 dividends of $974m therefore substantially understate the forward run rate. F01 [S011] equity statements and Notes 2,12.

Historical accounting returns are high, but a precise sustainable return on operating capital is not established here. Financial assets also sit inside Other assets: FY26 year-end included $4,840m of long-term public equities in addition to its marketable category. Fully separating operating capital requires consistent treatment across endpoints. Expensed R&D, rapid growth, late-year transactions and contingent ecosystem commitments further limit a conventional balance-sheet denominator. The valuation's future reinvestment assumptions must therefore stand on their own evidence.

Commitments, guarantees and downside funding

The following schedules use billions , not the millions used above. Amounts are contractual or conditional over different horizons. They are not current debt balances or estimates of losses. The latest 10-Q distinguishes main commitments, additional arrangements and guarantees; this reconstruction keeps those categories separate. F01 [S011] Note 10, pp 17–20.

Main commitments, $bn

Additional arrangements, $bn

Supply commitments rose from $119b a quarter earlier to $279b. Some terms permit adjustment, cancellation or rescheduling at a cost before firm orders; the public aggregate does not quantify the full cancellable share. Normal product costs eventually financed by sales form part of these amounts. Subtracting the entire supply schedule as a current incremental loss would overstate the burden. Assuming that cancellation provisions remove the burden would understate it.

Under the additional $36b AI-cloud arrangements, providers buy NVIDIA systems and NVIDIA commits to cloud services which providers may sell to third parties at more advantageous rates. Commitments reduce with capacity use and contingent revenue sharing may apply. This can help create productive capacity, while linking the supplier's economics to its customers' later utilization. The $20b third-party leases are expected to be reassigned; expected reassignment is not completed risk transfer.

Guarantees are a separate contingent category

The additional option covering approximately 3.8 GW is discretionary and is not added to current guaranteed exposure. Expected loss cannot be inferred from the maximum cap without borrower credit, project readiness, contractual recovery and cash-flow analysis. The first future phase's timing also matters: a multi-decade contingent guarantee is not equivalent to a present demand for the full amount.

Illustrative six-month liquidity stress; assumptions, not a probability forecast; $bn

Operating cash already reflects supplier and operating payments, so supply commitments are not subtracted again. The illustration demonstrates a substantial immediate cushion under these stated assumptions. It does not bound multi-year risk from weaker collections, investment losses, new commitments or the later $105b guarantee. Continued buybacks, acquisitions or further strategic calls would use more of the cushion. Buybacks are discretionary; the dividend assumption is explicit rather than treated as an unavoidable debt claim.

Governance and whether incentives follow the new risks

Latest proxy evidence identifies founder influence, performance incentives and disclosed relationships. Huang beneficially owned 870,604,104 shares, or 3.58%, as of March 23, 2026, including interests defined in the proxy's trust and foundation footnotes. This is substantial influence, not majority voting ownership. The board determined all directors other than Huang independent; Stephen Neal serves as independent lead director and the company has no board chair. Long service warrants scrutiny but does not itself disprove independence. F04 [S014] pp23–36.

Incentive and oversight evidence

Groq was a nonexclusive technology license and employee transfer, not a conventional purchase of its products, customer contracts or equity business. Accounting recorded $14.4b goodwill and $2.5b developed technology with a five-year life; $13b was paid initially and $4b including imputed interest was due within a year. Future amortization, ongoing research expense and upfront cash have different economic roles. No demonstrated investment return is inferred from the purchase accounting. F02 [S012] Note2.

The crypto-related securities and derivative proceedings remain allegations, not adjudicated findings. The latest filing reports no accrual because management regarded loss as reasonably possible but not probable and could not estimate a range. This is the issuer's accounting assessment, not legal clearance or a research estimate of zero loss. Antitrust, export and other regulatory matters require their own event evidence. F01 [S011] Note10.

The decision-relevant governance question is whether capital and credit limits remain disciplined as growth increasingly involves investments, extended customer terms and guarantees. Public disclosures show the expansion. They do not provide a representative cohort of realized customer returns, a complete receivables-aging and subsequent-collection schedule, or a probability-weighted loss estimate on the major guarantees. Those omissions limit confidence in a single precise estimate of fully distributable cash or sustainable return on capital.

What the original quarterly filings add

Targeted review of all nine original 10-Qs in this twelve-quarter window confirms that inventory reserves are an ongoing economic cost. Net quarterly inventory and excess-purchase-obligation expense was $442m in FY24 Q3 and $279m, $823m and $759m in FY25 Q1, Q2 and Q3. FY26 Q1 expense was approximately $4,864m; Q2 and Q3 were $522m and $44m. FY27 Q1 was approximately $997m; Q2 was $808m. Only the FY26 Q1 and FY27 Q1 estimates derive from rounded gross disclosures. FY25 Q2 charges were primarily associated with low-yielding Blackwell material; the FY26 Q1 H20 charge split into $1.9b inventory and $2.6b purchase obligations. Removing the exceptional export shock does not justify adding back ordinary launch, yield and forecast-error costs. F29 [S039]–F37 [S047], gross-margin MD&A and working-capital notes.

Disclosed customers above the receivable concentration threshold accounted for 22% of A/R in FY24 Q3, 49% in FY25 Q2, 56% in FY26 Q2 and 70% in FY27 Q2. These are changing sets of direct customers, not a constant cohort or a measure of final demand. Customer letters are not permanent identities. Geography also changes basis: FY26 Q3 replaces billing location with direct-customer headquarters and recasts comparatives. Combining those series would misstate geographic trends. F29 [S039], F31 [S041], F34 [S044]–F37 [S047].

Guarantee exposure was already developing before the latest large arrangement: FY26 Q3 reported an $860m facility guarantee with $470m partner escrow, versus $3.5b and $712m in FY27 Q1. Different obligation categories and horizons prevent treating every historical commitment headline as the same debt-like series. F35 [S045] Note 9; F36 [S046] Note 8.

A FY24 useful-life estimate extension for NVIDIA's own equipment increased Q3 operating profit by $41m and nine-month operating profit by $107m. This was small relative to the revenue inflection and says nothing definitive about customers' GPU lives. Across all nine quarterly Item 4 disclosures, management reported effective disclosure controls and no material quarterly internal-control change, while a phased ERP upgrade continued. These are management conclusions, not nine separate audit opinions. F29 [S039] Note 1; F29 [S039]–F37 [S047] Item 4.

Demand, customer funding and competitive advantage

On August 10, NVIDIA announced memoranda with six major financial firms aiming to mobilize more than $500bn of third-party infrastructure capital over time. Final agreements remained a condition in that announcement. This could broaden customer financing and reduce dependence on NVIDIA’s own balance sheet. It is not $500bn of funded cash, NVIDIA revenue or an independently verified return on deployed equipment. The model assigns no automatic sales increase or guarantee release to the headline. P09 [S010].

Paid AI demand is visible. The unresolved investment question is whether its cash returns can support the next wave of infrastructure commitments, and how much of the resulting economics NVIDIA can retain.

Cloud providers recognize growing revenue, Microsoft reports paying Copilot users, and specialized GPU operators have operating infrastructure. These are signs of commercial demand. Cloud revenue also includes non-AI services, and AI workloads can run on other processors. The investment case depends on converting that demand into cash after equipment and operating costs.

Follow the customer's cash

The economic chain runs from an end user to an application or model provider, then to a cloud operator, systems vendor and chip supplier. Some buyers own equipment directly. The customer on NVIDIA's invoice may therefore differ from the cloud operator bearing the investment risk and the application customer ultimately funding it. Adding spending at every link counts the same demand repeatedly. Investments flowing back from suppliers or cloud providers to their customers make the cash path more complex.

Can the customers fund the spending?

The large customers have substantial operating cash and access to financing. Several are nevertheless spending beyond current cash generation. That combination supports continued infrastructure purchases while leaving demand exposed to financing conditions and the returns on completed projects. Company-wide cash strength cannot establish the profitability of an individual AI deployment.

The comparison below starts with operating cash, subtracts cash equipment spending and then identified finance-lease principal. Operating rent and lease interest already included in operating cash are not deducted again. The figures precede dividends, acquisitions and most financial investments. They cover different periods and businesses, so they should not be added or treated as a ranking of AI returns.

Cash after equipment and identified lease principal, USD billions

Amazon's equipment figure is net of proceeds and incentives; its final deduction is financing-obligation principal. Its less seasonal trailing-year result is −$9.511bn after those repayments. Alphabet separately prepaid $0.835bn on future finance leases; the filing does not explicitly reconcile that payment to the cash-flow captions used here. Oracle discloses $0.186bn of finance-lease cash without a principal split. Its separate $0.830bn net repayment of capital-expenditure financing reduces the displayed amount to −$6.226bn before lease principal. CoreWeave's other debt repayments are outside the table. These qualifications keep the comparison from overstating its precision.

Cash balances need qualification. Microsoft's reported $76.843bn includes $3.8bn restricted under a supplier agreement. Alphabet's broader $242.474bn cash-and-securities balance includes $87.063bn of equities, including $80bn of SpaceX shares under short-term sale restrictions. The table therefore shows Alphabet's cash and debt securities separately from equities. Meta's $90.260bn already excludes restricted cash, including infrastructure escrow expected to be released in 2028–30. These investments and escrow remain assets, but cannot all fund immediate purchases. CF04 [S141], CF06 [S143], CF07 [S144].

Amazon has financing capacity alongside a growing burden. It received $66.998bn of long-term borrowing proceeds in the first half. A separate $17.5bn delayed loan was undrawn at June 30 and had to be drawn by September 30; its status at the research cutoff was not verified. Existing operating and finance leases require $10.424bn in the second half of 2026, with another $4.018bn scheduled on uncommenced leases and $23.452bn of unconditional purchases. After June 30, Amazon issued $25.0bn of U.S. dollar notes for general corporate purposes, maturing in 2029–2066, and paid the remaining $21.3bn OpenAI investment commitment. The issuance supports access to financing and adds future debt service. Its face value is not net proceeds, and these events do not provide a complete September cash reconciliation. CF01 [S138].

Future commitments constrain flexibility. Microsoft's next fiscal year includes $9.250bn of debt principal and $13.203bn of lease payments. Alphabet reports $200.7bn of short-term purchase and other commitments. Meta reports $53.52bn of noncancelable commitments due in 2026. Oracle's remaining fiscal year includes $3.839bn of lease payments and $5.449bn of unconditional purchases. These categories differ and overlap ordinary operating or capital spending; adding them into a single “funding gap” would mislead. Uncommenced leases further extend the burden, often over many years. CF03 [S140], CF05 [S142]–CF08 [S145].

Advances help liquidity before services are earned. Oracle identifies $11.363bn of customer prepayments with a significant financing component in operating cash. Excluding that inflow gives −$16.759bn after equipment, before lease principal. The advances are real funding with future delivery obligations. CoreWeave's undrawn facilities are favorable evidence, but draws depend on financed assets and contracts; debt principal of $4.413bn falls due in the second half of 2026. Its deferred-revenue growth is not a sufficient basis to classify all customer funding as financing-like. CF08 [S145], CF09 [S146].

Available cash and borrowing capacity support continued spending; funding gaps and future obligations leave collections exposed to financing conditions. These disclosures do not establish a loss rate for NVIDIA's receivables or support commitments. Cash earned from delivered services and available for repeat purchases would provide stronger evidence of durable funding.

Do established deployments earn enough to fund another generation?

The public evidence supports profitable operations at some established deployments. It does not yet show representative cash returns after recovering equipment, corporate overhead and financing. “Mature” can mean fully running, not that the original contract has ended or all capital has been repaid.

CoreWeave's chief financial officer said in February that its fully ramped contract portfolio generated contribution margins in the mid-20s and EBITDA margins around 70%. EBITDA means earnings before interest, tax, depreciation and amortization. That is a statement about operating contracts, stronger than a forecast for newly signed business. It remains management-reported: no cohort size, distribution of results or reconciliation to cash retained after all costs accompanied the statement. CR01 [S147].

In CoreWeave's illustrated calculation, contribution margin is revenue left after direct data-center costs and technology/infrastructure expenses, including equipment depreciation. Its example deducts $27 and $48 from $100 revenue, leaving $25. Corporate sales and administrative costs, interest and tax are outside that displayed calculation. Depreciation allocates equipment cost over its accounting life; it is not proof that the purchase has already been recovered in cash. The separate five-year cash-return example assumes 80% asset financing at 6.5% and is explicitly illustrative. CR02 [S148].

What the evidence establishes

The cash accounts preserve the distinction. CoreWeave reported $3.663bn first-half operating cash and $14.117bn cash equipment/software purchases. A $1.365bn increase in deferred revenue and $607m receivables release supported operating cash. These flows are real, but mix advances, collections and expansion. The filing reports one operating segment, without separating stabilized-contract cash from new deployments. Its committed contracts also generally require payment irrespective of physical utilization: weak application use may reach the operator through credit or renewal risk before current billings decline. CR03 [S149], CR08 [S154].

Nebius provides a useful cross-check. Its AI cloud reported a 49.7% Q2 adjusted EBITDA margin, which excludes depreciation and financing costs. Its advertised one-year-ten-month payback for new deals uses forecast costs and future capacity, including sites not yet built; the method excludes prepayments. Group operating cash of $2.246bn included $1.197bn of additional deferred revenue. Neither the margin nor the payback estimate is a completed investment return. CR06 [S152], CR07 [S153].

The favorable operating evidence supports continued customer procurement and weakens a blanket claim that deployments cannot pay. Representative cash recovery across completed contracts would strengthen the demand and customer-funding assumptions; sustained collection or renewal shortfalls after stabilization would favor higher working-capital needs or support losses. The present disclosures do not quantify those effects on NVIDIA.

Reciprocal transactions require careful interpretation

NVIDIA invested $2bn in CoreWeave in January 2026. Separately, a September 2025 order initially valued at $6.3bn provides NVIDIA access to residual unsold capacity through April 2032, subject to conditions. Equity and a capacity contract are distinct exposures; neither is an immediate customer receivable or proof that all capacity would otherwise remain idle. The combination nevertheless supports the operator's expansion and makes independent end-user cash generation a critical test. E11 [S058]–E13 [S060]

OpenAI's March 31 closing update reports $122bn of committed capital, superseding the February $110bn headline for the round; the amounts must not be added. Its approximately $4.7bn revolving credit facility was undrawn at that announcement. A closed funding round strengthens financial capacity, but committed capital, available borrowing and cash received remain different quantities. Reported revenue of $2bn per month demonstrates scale without disclosing audited cash profit. NVIDIA remains foundational to training and much inference while OpenAI diversifies silicon. R01 [S136]; February history: E14 [S061]

Anthropic's May $65bn round includes $15bn of previously committed hyperscaler investment. Its stated revenue run rate exceeds $47bn, which is not audited annual revenue or profit. Its Amazon and Google/Broadcom capacity arrangements demonstrate that expanding AI demand can favor alternative architectures. Supplier-supported financing is therefore neither uniquely NVIDIA's model nor sufficient evidence of accounting impropriety. E15 [S062]

Where the platform protects NVIDIA, and where buyers can switch

Our judgment is that NVIDIA has a valuable system advantage, strongest for changing workloads and teams that cannot economically build another stack. It faces credible substitution among large model providers and repeat inference workloads. This is already affecting procurement choices; it need not cause declining NVIDIA sales. More AI work, less NVIDIA share of new capacity and more NVIDIA networking content can occur together. The consequential question is how much revenue and profit the company retains from each expansion.

Competition by buying decision: current evidence and consequence

These conclusions rank mechanisms, not market shares. Framework portability helps: PyTorch's HIP backend reuses CUDA-facing interfaces and newer ROCm support closes specific gaps. It does not eliminate custom-kernel work, numerical validation or operational differences. Google offers its own compiler, interconnect and serving stack; its November 2025 native eager PyTorch discussion was development/early-preview work, not universal drop-in compatibility. E18 [S065], E32 [S079], E45 [S092]

What is operating, and what is still a commitment

Selected primary commercial evidence available before the publication cutoff

Cerebras' loan disclosure makes the cash boundary concrete: by June 30, $86.327m of noncash billings and other permitted credits had reduced the OpenAI loan, with no cash loan repayments. That cumulative amount is not a reconciliation of the quarter's $56.8m revenue and must not be subtracted from it. The arrangement converts earlier funding into services and other credits; recognized revenue does not mean the same amount of new cash arrived that quarter. E39 [S086], Notes 3 and 10.

Competitor financing also deserves scrutiny. Broadcom's XPV program includes a backstop with a maximum approximately $29bn at full deployment; its amount grows with deliveries and falls with payments. No amount had been paid at the August quarter-end. AMD's conditional customer warrants and Cerebras' customer-funded buildout similarly show that procurement economics can include financing and incentives. Hardware diversification does not necessarily diversify exposure to the same ultimate lab customers. E16 [S063], E39 [S086], E48 [S095]

The favorable counterexample is equally concrete. Nscale's September filing describes Anthropic agreements for future NVIDIA Rubin infrastructure with aggregate payments up to approximately $44.6bn. Required financing was not yet subject to binding commitments, and delivery conditions matter. This is substantial contracted demand, not installed capacity or recognized NVIDIA revenue. Alongside Anthropic's operating alternatives, it supports multi-sourcing with continuing NVIDIA growth, rather than wholesale abandonment. R02 [S137]

Adaptation can preserve revenue, at a cost

NVLink Fusion lets custom accelerators use NVIDIA interconnects and rack infrastructure with optional NVIDIA processors, networking and software. Losing a GPU socket need not mean losing every dollar of system content. But no realized attachment rate shows how much revenue or profit is retained, and rival network suppliers compete for that spending. We therefore distinguish GPU share, total NVIDIA content and operating profit. E42 [S089]

The Groq arrangement demonstrates both capability and expense. NVIDIA's FY2026 filing describes a nonexclusive technology license and employee hires: $13bn paid and $4bn payable including imputed interest, with no Groq customer contracts, existing products or equity purchased. Its Rubin/LPX architecture combines GPUs with specialist inference technology. That gives NVIDIA a response to specialization, while requiring integration and continued development. The old payment must not be deducted twice; future adaptation is also not free merely because the platform is strong. E49 [S096], E47 [S094]

These mechanisms affect the named valuation assumptions differently. Retaining fewer new installations lowers the revenue path; retaining interconnect and processor content can offset some loss. Lower net pricing or more purchased component content can compress gross margin. Additional engineering, integration and customer support reduces operating margin unless scale or other savings compensate. The middle path's eventual 50% operating margin still means $50 of operating profit per $100 of sales: it assumes substantial lasting pricing power after these costs. Public evidence supports that as a possible outcome, not a uniquely demonstrated margin. The valuation chapter tests these operating drivers. (Evidence: valuation)

A simple illustration shows why demand growth does not settle profitability. A business with $100 of sales and $50 of total operating costs earns $50. A 10% price cut at unchanged volumes and costs leaves $90 of sales and $40 profit: profit falls 20% and margin becomes 44.4%. This is arithmetic, not a NVIDIA forecast; volumes, manufacturing savings and richer system content can offset it.

Selected renewals also strengthen the defense. CoreWeave's September 17 presentation reports an A100 renewal into 2029 with three-year pricing aligned to typical one-year terms, and an H200 three-year renewal above its original price. These two unnamed, company-selected examples demonstrate continuing commercial usefulness, not fleet-wide returns. Longer life can improve customer cash recovery and preserve familiar software, while delaying replacement purchases for a fixed workload. Our sales cases must therefore depend on new useful demand as well as replacement. E43 [S090]

What older equipment and rival deployments reveal

Replacement demand depends on what customers can earn from the next machine and recover from the last. Amazon’s 2025 accounts shortened a subset of server and network depreciation lives from six to five years as technology advanced, adding $1.4bn to depreciation. Its general schedule spans five to six years. CoreWeave uses six years for technology equipment and reported no material long-lived-asset impairments in 2023–25. Nebius moved from four to five years in 2026, citing usage and utilization commitments. These are accounting estimates, not measured profitable lifetimes; an asset can remain recoverable without earning an adequate return on its original investment. Amazon [S155], CoreWeave [S156], Nebius [S166].

CoreWeave’s selected renewals provide stronger commercial evidence. One A100 contract extends into 2029 at three-year pricing comparable with typical one-year terms; an H200 renewal carries a premium to its original contract. That rebuts automatic worthlessness when a new chip arrives. It does not establish fleet-wide recovery, and the A100 architecture’s 2020 launch does not date every installed unit. Customers can also buy new generations while retaining older capacity. Longer use can improve recovery of the original equipment cost while delaying replacements. Faster upgrades can lift NVIDIA’s near-term sales even when customers rely on outside financing; durable repeat demand depends on customer returns, funding access and continued NVIDIA purchases. Renewal and expansion examples [S157].

Evidence that constrains the assumptions

Competition is already commercial. Anthropic reported over one million Trainium2 chips training and serving Claude. Meta reported hundreds of thousands of its own MTIA chips in production, including recommendation-model training; its later generative-AI chip deployments remain distinct plans. Cerebras recognized $56.8m of quarterly OpenAI service revenue, although that includes pass-through costs and can settle earlier customer financing. AMD also published named production deployments on Azure and at Meta. These observations show viable alternatives, without proving equal economics for every workload or treating internal use as merchant chip revenue. Anthropic [S158], Meta [S159], Cerebras [S160], AMD [S161].

New-capacity awards matter before installed NVIDIA equipment is replaced. Anthropic’s up-to-2GW AMD agreement and OpenAI’s expected fourth-quarter Helios start support a competing pipeline. They are not completed deployments. Nor is manufacturing a rack the same as a customer earning cash from it. AMD’s dated customer milestones [S162].

NVIDIA has a concrete defense beyond future product promises: Microsoft’s AMD-based ND MI300X v5 machines specify a dedicated NVIDIA Quantum-2 CX7 network connection for each accelerator. AWS’s Trainium4 design also selects NVLink Fusion, but that future design does not establish present shipment revenue. Fusion’s other NVIDIA components are optional, and AMD’s Helios uses its own networking. Retained system value is therefore credible and contested. Microsoft specifications [S163], Trainium4 design [S164], Fusion components [S165].

This evidence supports keeping both margin pressure and retained-content upside in the valuation. It does not identify a representative retirement curve or justify a precise new market share. Longer profitable use could reduce replacement sales, while NVIDIA networking in rival-chip systems could preserve revenue. Neither effect is yet measured across a representative fleet.

What the performance evidence does establish

Two specific benchmark claims, with different questions and limits

These results concern different models, systems and denominators, so they cannot select a universal winner. They do identify why workload and power constraints matter. Combined with operating customer deployments, the evidence is sufficient to conclude that alternatives can win real work and NVIDIA can remain highly competitive in others. MLPerf supplies structured tests, but this review did not reconstruct its database or execute a local benchmark. E21 [S068]

A further issue is the pending Hugging Face acquisition: approximately $11.9bn purchase consideration plus up to $1bn retention, expected in first-half 2027 subject to approvals. NVIDIA commits to maintaining an open platform supporting other silicon. Our inference is that distribution and integration opportunities coexist with developer-neutrality and regulatory risks. Closing, synergies and customer reactions are unproven; the transaction is not treated as completed. E34 [S081]

Useful work and observed commercial prices

Customer return requires cash collected or costs demonstrably avoided after power, cooling, networks, engineering, support, financing and equipment refresh. Training must include completion time and failed runs. Inference must specify quality, latency, context length, batching, cache use and concurrency. The relevant denominator is a correctly completed useful task. A GPU-hour or token alone cannot provide it.

Efficiency creates competing effects. If useful output per accelerator doubles with unchanged task demand, required capacity halves. If useful tasks triple simultaneously, capacity rises 50%. These are arithmetic examples, not forecasts. Quantization, smaller models and caching reduce requirements; reasoning and agent loops can increase computation per task. Actual paid workloads and utilization decide the balance.

An illustrative customer's break-even condition

Consider an operator selling one standardized GPU-equivalent unit of service. Every input below is an analyst assumption, not an observed NVIDIA device cost, cloud contract or customer return. Assume $40,000 installed hardware/network cost allocated per unit; zero resale proceeds; $6,000 annual fixed facility, staff and support cost; $0.50 variable power/service cost per paid hour; and an 8% annual capital charge. A four-year capital-recovery schedule costs about $12,077 a year, including recovery of the original purchase and the financing/required-return allowance. This is equivalent to four level annual payments at 8%; it does not assume any actual customer has those loan terms. Do not add depreciation, principal or interest again to that charge.

At 70% paid utilization, 8,760 calendar hours produce 6,132 paid hours. A net realized price of $4 per paid hour generates $24,528 annual revenue. After $3,066 variable cost, $6,000 fixed cost and the capital charge, the illustrative pre-tax surplus is $3,385. “Paid utilization” means hours billed and collected; it is different from a chip's computational busy time.

Assumed customer economics; one change at a time, USD per unit

The formula is: break-even price = variable cost per paid hour + (annual fixed cost + annual capital-recovery charge) ÷ paid hours. The base breaks even at about 59% paid utilization at $4; at $3 it requires about 83%. Useful life here means commercially productive service life, not an accounting depreciation election. No ramp delay, income tax, bad debt, working-capital timing, residual value or separate equity-return premium is modeled. This simplified levelized test is therefore not a full project valuation or debt-covenant calculation.

The illustration explains mechanisms; it does not underwrite an actual CoreWeave contract or resolve the ramp-versus-mature-profitability question above. Lower collected prices or paid use can remove cash for the next purchase. Under take-or-pay contracts, physically idle equipment may still generate billings; customer credit and renewal timing determine when weaker application demand reaches operator cash. Early obsolescence can impair recovery, while attractive upgrades can accelerate replacement purchases. Public asking prices below are not the assumed operator's collected prices. Actual customer inputs are needed to establish realized returns.

Public asking prices, September 21, 2026 UTC; USD

These are advertised prices; actual availability and negotiated transaction prices remain unknown. This snapshot demonstrates choices, not a falling-price trend or spare capacity. A 20% reduction in paid utilization, from 80% to 64%, raises fixed cost per paid hour by 25% before price cuts; this sensitivity is not an estimate of observed utilization.

Productivity evidence also resists simple conclusions. METR's early-2025 study found experienced developers took 19% longer using AI on its selected tasks. Its February 2026 update describes selection and measurement problems in later estimates and believes newer tools likely help more, without a robust magnitude. Neither the old slowdown nor enthusiastic adoption establishes today's universal return. E22 [S069]

Outside observations, expectations and falsification

The public-discussion sample covers English-language posts from January 2026 onward: two dated LocalLLaMA product threads and two separate stocks-investor threads. Two additional Hacker News discussions provide context only because their exact dates were unavailable. Ranking is uncontrolled and authors are unverified. No sentiment percentage, market share or purchase estimate follows from this small convenience sample.

Product discussions describe both NVIDIA familiarity and alternatives constrained by memory, budget or privacy. Investor discussions debate demand timing and alleged price increases. These populations answer different questions. Reports of increases above 15% remain an unverified realized-pricing lead here; duplicated reporting is not independent confirmation, and a system-price increase need not become an equivalent NVIDIA margin increase. E26 [S073]–E31 [S078], E33 [S080]

The five durable decision tests (Evidence: thesis tests) connect these observations to sales, margins, collections, replacement demand and support costs. Public evidence leaves customer-level utilization, private-lab margins and comparable production costs unresolved.

Stronger backing, with more obligations to deliver.

OpenAI’s March 31 update materially strengthens the financing counterargument. It reported closing the round with $122bn of committed capital , broader institutional participation, a roughly $4.7bn revolving facility undrawn at closing , and revenue of $2bn per month . The $122bn updates February’s $110bn announcement; adding the two would count the same financing twice. Committed capital is not a disclosed unrestricted cash balance, the revolver is borrowing capacity, and monthly revenue is neither audited profit nor free cash flow. Amazon, NVIDIA and SoftBank remained strategic anchors, alongside other investors. The update reduces the force of a near-term funding-shortage argument without proving independence from suppliers or the ability to finance every planned deployment. R01 [S136].

Nscale’s September 18 filing makes NVIDIA’s relationships with one infrastructure provider more concrete. Different contracts create different claims on cash; their face amounts cannot be added into one loss estimate. R02 [S137].

Nscale transactions and their treatment at the research cutoff

Nscale also disclosed August 25 agreements for Anthropic to use NVIDIA Vera Rubin infrastructure, with aggregate payments to Nscale of up to $44.6bn . This is substantive evidence against a simple claim that alternative chips displace all NVIDIA demand. It is future cloud activity across four tranches, subject to acceptance and availability. Nscale explicitly said it had no binding financing commitments for the required deployments at the prospectus date. Delays or sustained service failures can permit termination. The contract total is therefore neither installed capacity nor funded NVIDIA revenue. R02 [S137].

These arrangements support continued NVIDIA demand while tying its finances more closely to customers. Stronger customer backing deserves weight in the favorable case; conversion into paid, productive systems remains the test. None of these disclosures justifies adding headline contract values directly to NVIDIA sales or subtracting all support amounts from equity value. The valuation below separates cash timing, acquired assets, services received, existing commitments and potential recoveries. Contract-level gaps constrain confidence and the support assumptions.

Rates, supply and the conditions for deployment

NVIDIA enters the next quarter with expanding manufacturing capacity and a less forgiving financial environment. The investment question is whether customers can turn delivered equipment into useful, paid computing quickly enough to justify their capital commitments. Higher rates, electricity constraints and NVIDIA’s growing financial support for its ecosystem make collection and customer returns increasingly important. Supplier expansion supports the opportunity; it does not establish the profitability of every proposed AI campus.

A higher hurdle for earnings growth

The Federal Reserve raised its target range by 25 basis points to 3.75–4.00% on September 16. It maintained ample-reserves operations and reinvestment, which supports monetary implementation but does not promise rising equity prices. The ECB also raised its deposit rate to 2.50%, effective September 16; Taiwan held its discount rate at 2%. These decisions affect financing in different markets, with U.S. rates most immediately relevant to NVIDIA’s valuation and dollar-funded customers. [M01] [S104] [M02] [S105] [M17] [S120] [M18] [S121]

U.S. Treasury nominal par yields: annual percentages, 92-calendar-day comparison

These indicative yields are not holding-period returns. September 18’s ten-year real Treasury yield was 2.68%. Higher competing yields increase the hurdle for a premium equity valuation, although stronger expected economic growth can simultaneously support earnings. A hypothetical dollar received ten years away falls about 8.6% in present value when its discount rate rises from 10% to 11%; that illustration is not an estimated NVIDIA price sensitivity. [M03] [S106] [M04] [S107]

NVIDIA demonstrated financing access by issuing $25 billion of fixed-rate notes in June. Existing coupons do not reset when the Fed moves. The more immediate transmission runs through new customer borrowing, refinancing and returns on reinvested cash. Every $10 billion of accessible cash that fully resets by one percentage point changes annual pretax interest by approximately $100 million; actual timing and balances matter. Broad credit indices cannot establish whether an individual AI tenant can meet its obligations. [M27] [S130] [M05] [S108]

Inflation, currencies and energy

August consumer inflation was 3.4% year over year, with core inflation at 2.4% and energy up 16.3%. Payrolls grew 162,000, unemployment was 4.1%, and July real consumer spending was approximately flat. These releases describe mixed household conditions, not a direct forecast of hyperscaler spending. Gaming and workstation demand face a more immediate household-income channel; data-center demand depends more on funded budgets and profitable workloads. Consumer-price measures also differ from semiconductor input costs. [M06] [S109] [M07] [S110] [M08] [S111]

The latest official weekly foreign-exchange observation used here is September 11: NT$31.64 and ₩1,340.30 per U.S. dollar. Dollar-priced equipment exposes overseas buyers to local purchasing-power changes, while supplier costs and overseas engineering expenses create partly opposing effects. Published hedges do not reveal a complete net currency sensitivity. Separately, EIA’s August Brent average was $91 per barrel, $7 above July. Energy shocks transmit mainly through customer operating costs, construction and logistics; they should not become an arbitrary percentage cut to NVIDIA’s gross margin. [M09] [S112] [M12] [S115] [M16] [S119]

Manufacturing is a sequence of qualified capacities

TSMC raised its 2026 capital-spending outlook to $60–64 billion and reported $40.2 billion of second-quarter revenue. Its commentary still identified CoWoS as the principal advanced-packaging route and described alternative substrate development as unfinished. Micron reported more than $1 billion of HBM4 revenue, a rapid ramp and expected memory tightness beyond 2027; additional Singapore packaging capacity was expected in the first half of 2027. These are meaningful supplier commitments and operating observations, but neither discloses NVIDIA’s complete allocation, yield or qualification schedule. [M13] [S116] [M14] [S117]

Deployment constraints and the evidence needed to resolve them

NVIDIA’s $279 billion of supply and capacity commitments therefore have two interpretations: access to scarce production and exposure if demand or product mix changes. They span multiple periods and contractual terms; the total is neither a current loss nor guaranteed revenue. Taiwan concentration remains consequential even when mainland-China accelerator shipments are restricted. Israel-based networking activity adds another continuity exposure. No numerical probability of conflict or immediate substitute-capacity assumption is justified. [M12] [S115]

Megawatts require dates, permits and economics

FERC’s June action sought changes to large-load integration processes at six grid operators. It did not grant every proposed campus immediate electricity access. PJM had already reduced its 2027 peak-load forecast by 2.3% relative to its prior forecast after refining expected additions. EIA’s September outlook also lowered its 2027 West South Central electricity forecast, partly reflecting a Texas connection pause. These observations caution against equating all requested megawatts with eventual equipment demand. [M15] [S118] [M22] [S125] [M16] [S119]

PORTS illustrates the distinction. DOE confirmed a March groundbreaking; OpenAI expected the first 800 MW in 2028, with wider development dependent on generation, transmission, financing and permits. NVIDIA’s initial guarantees cover approximately 4.25 GW and have a cumulative $105 billion payment cap, subject to lease-readiness and default conditions. Optional additional capacity is separate. Recovery through re-letting or sale can reduce payments; reimbursement from the defaulting tenant is a correlated exposure. This cap should not be deducted as immediate debt or counted as certain future revenue. [M20] [S123] [M21] [S124] [M26] [S129]

Cooling claims also require a site boundary. NVIDIA describes a 45°C liquid-cooling design capable of dry heat rejection in suitable climates, while PORTS’ final water requirements remain unresolved. Closed-loop cooling does not prove zero whole-site water consumption. At a constant 100 MW of IT load, a $10/MWh electricity change costs $8.76 million annually before facility overhead. A delayed power connection can matter more than that price change. [M19] [S122] [M20] [S123]

International expansion broadens the opportunity without removing these conditions. NVIDIA’s September 9 Australian announcement describes a partner buildout of up to 2 GW by 2027. Participating providers would operate the facilities, while NVIDIA supplies its computing platform, networking, software and support. The announcement does not disclose a NVIDIA dollar purchase commitment or guarantee, establish that the entire capacity is operating, or quantify incremental orders. It supports the breadth of deployment plans, while leaving the same funding, completion and utilization tests in place. [M32] [S135]

Policy and security-specific risks

U.S. H200 licensing eligibility is conditional, not unrestricted China access. Chinese policy and customer acceptance can still prevent sales, and NVIDIA reports that the applicable inspection-related tariff cannot be passed through to customers. The narrow 25% semiconductor tariff also has U.S.-use exceptions; applying it to every NVIDIA shipment would be wrong. The base case should recognize China sales only when licensed, accepted and economically attractive. [M10] [S113] [M11] [S114] [M12] [S115]

The proposed $11.9 billion Hugging Face acquisition, announced in September with up to $1 billion of employee equity retention, adds approval, integration and open-model regulatory exposure. Expected closing is in the first half of 2027; benefits remain contingent. Securities litigation and the China Mellanox investigation also remain relevant despite no quantifiable loss estimate. Shareholder voting provides another distinction: the 2025 supermajority-removal amendment failed its outstanding-share threshold, while the 2026 approval was advisory. Available evidence does not establish current investor positioning or option-implied probabilities. [M25] [S128] [M12] [S115] [M29] [S132] [M30] [S133]

The central interpretation is continued funded AI expansion with localized delays and demanding valuation hurdles. A harsher case combines persistent energy pressure, expensive funding, slower collections and obsolete inventory; a favorable case requires improving deployment and independent customer cash generation. Falling Treasury yields alone would not cure weak AI demand. The next discriminators are customer budgets, collections, accepted systems and energized sites, alongside September 30 PCE, October 2 employment, October 14 CPI and the October 27–28 and December 8–9 Fed meetings. A delay becomes lasting impairment when financing disappears, buyers substitute or economics no longer support completion. [M06] [S109] [M07] [S110] [M08] [S111] [M23] [S126]

Price the cash that success could leave for owners.

At $222.27 and 24.3bn diluted shares, the equity claim costs about $5.40tn. July’s actual ending 24.147bn shares give $5.37tn. We use diluted units consistently. Trailing free cash flow after equipment spending and related principal was $126.886bn. Charging the $7.241bn stock-compensation add-back gives an illustrative $119.645bn cash proxy: a 2.22% yield, or 45.14 times cash. This describes current economics, not sustainable forecast earnings. F01 [S011], F02 [S012], P01 [S001].

Operating profit after an illustrative 17% ordinary tax is $163.991bn, implying 32.9 times that earnings basis. Investment and timing explain the difference from the cash proxy; they are not independent fair-value estimates. Trailing net income includes $30.552bn pretax equity gains, excluded from operating valuation. Product mix, growth, acquisition accounting and capital needs prevent importing an unadjusted peer multiple.

How to read the values

We estimate ten years of operating profit after tax, add depreciation, and deduct capital spending and extra working capital—the cash tied up in receivables and inventory after supplier/customer funding. This leaves cash before financing. Stock compensation stays inside operating expenses. Fixed 24.3bn shares are the corresponding cash-equivalent compensation convention: no second dilution charge or buyback advantage.

Each annual cash flow is discounted at year-end using 11%. This required return is the compensation we demand for time and business risk, not a stock-return prediction. A continuing value estimates cash after year ten. These are conditional present values as of September 20, using rolling years ending September 2027–2036. The reported $302.970bn trailing sales and asset balances are dated July 26 and carried forward as proxies for September opening conditions. No complete September balance-sheet rollforward is asserted. No scenario probabilities are assigned.

Conditional present values · ten annual periods · 11% discount requirement

The assumptions doing the work

Gross margin is sales less product costs, divided by sales. Operating margin also deducts research, sales and administrative costs. A 50% operating margin means $50 operating profit per $100 sales before financing and tax. This remains an exceptional level of profitability; the large future revenue number alone does not establish how demanding the path is. The middle path grows at an 8.89% compound annual average from year one to ten. Growth starts near 24% and slows toward 2%. The July trailing-revenue proxy predates the September model start, so it is not an exactly aligned ten-year growth base.

What has to happen for the cash paths to work.

The first year can be reconciled to management’s outlook. The later years require judgments about customer spending, NVIDIA’s share of that spending, system content and replacement cycles. These bridges expose those judgments; they do not turn conditional values into a forecast distribution.

From fiscal guidance to the first rolling year

The first model year runs from September 21, 2026 through September 20, 2027. NVIDIA’s fiscal 2027 has 53 weeks, including a 14-week fourth quarter. Reported first-half sales were $177.837bn; the third-quarter guide is $108bn, plus or minus 2%. We assume $128bn for the fourth quarter. This produces fiscal 2027 sales of $413.837bn. Applying management’s preliminary 70% fiscal 2028 growth expectation gives $703.523bn. The fourth-quarter estimate and quarterly timing below are research assumptions. F01, Note 1 and statements [S011]; F05, guidance [S015]; F17, prepared remarks [S027].

The $590bn first-year bridge

We spread revenue evenly within each quarter solely to convert dates. The $128bn fourth quarter implies 18.5% sequential growth, or approximately 10.1% growth in the average weekly sales rate after adjusting for its extra week. That assumes product and supply expansion converts into accepted shipments. With the same timing assumptions, a fourth quarter of $118bn or $138bn gives first-year revenue of about $570bn or $611bn. Changing fiscal 2028 growth to 60% or 80% gives approximately $566bn or $615bn. Moving shipments toward later or earlier quarters gives approximately $573bn or $611bn. These are separate sensitivities, not a probability range.

What the later sales path requires

The middle path grows from $590bn in year one to $1,270bn in year ten: an 8.89% compound annual average over nine intervals, with faster growth early in the path. A trillion-dollar revenue figure is not, by itself, proof that the path is demanding. We test its requirements in three ways, each describing the same revenue rather than adding a new market.

Illustrative customer-channel decomposition, USD billions

This diagnostic holds Data Center at 92.5% of sales and splits it 55% hyperscale and 45% AI cloud, industrial and enterprise customers, near the latest mix. Holding those weights is our assumption. A hyperscaler can rent a neocloud’s capacity; the NVIDIA equipment sale is counted once. Adding cloud-service revenue, customer capex and equipment revenue would count different stages of the same spending chain. F01, segment disclosures [S011]; F17, customer channels [S027].

Required annual spending on addressable AI systems and networking

Divide Data Center sales by the assumed share to obtain these requirements. The pool excludes buildings, power, unrelated general-purpose IT and resale of cloud services. The share choices are hypotheses, not measured market shares, and the resulting pool is not an independently forecast market size. Near-term hyperscaler capex has a broader asset boundary. It cannot establish this decade-long denominator. Paid customer growth supports the funding argument; production use of rival accelerators makes retained share a live question.

A separate system-content check uses management’s approximately $40bn of Rubin revenue opportunity per gigawatt as a conversion convention. Year-five and year-ten Data Center sales then imply about 23.82 and 29.37 gigawatts’ equivalent of annual system sales. At assumed content of $30bn or $50bn per equivalent gigawatt, the year-ten requirement becomes 39.16 or 23.50. These are not forecasts of new grid capacity: generation, realized prices and replacement of existing systems change the relationship. CPU and networking content are already inside the system figure. F17, system-content discussion [S027].

If half the year-ten equivalent sales replace old systems, replacements would represent 14.68 gigawatts’ equivalent annually. Supporting that pace needs an eligible installed base of about 44, 73 or 103 gigawatts’ equivalent under three-, five- or seven-year replacement cycles. Those are hypothetical requirements, not a measured fleet. Longer equipment use can improve customer recovery but may slow replacement sales. Public evidence does not yet establish the spending pool, retained share and replacement base needed to make this our statistically preferred decade forecast.

From gross profit to operating profit

Current gross margin is approximately 75%; management’s recorded outlook steps down to 71–72% in fiscal Q4 and 72–73% in fiscal 2028. Our first-year 64% operating margin combines 72.5% gross margin with operating costs of 8.5% of sales. Later margins are research assumptions. F01, statements [S011]; F17, margin outlook [S027].

The middle path’s cost structure, USD billions except percentages

Holding the first-year margin constant, $680bn of additional sales would add $435.20bn of operating profit. Gross-margin erosion removes $133.35bn and rising operating-cost intensity removes $44.45bn. Thus operating profit rises by $257.40bn, from $377.60bn to $635bn, despite the 14-point margin decline.

The gross-margin fade allows for stronger customer bargaining, competing accelerators, purchased components and system mix, and transition costs. We cannot reliably assign a precise margin loss to each. A 62% gross margin still assumes substantial differentiation. Research and support costs can also grow faster than sales: the bridge triples operating-cost dollars while sales rise 2.15 times. NVIDIA reported rapidly rising research-compute costs. Ordinary research services and stock compensation remain inside operating expenses. They are not deducted again as investment or customer-support losses. F01, MD&A operating expenses [S011].

This explains what the 50% operating-margin assumption means; it does not prove that 50% is the most likely outcome. More durable pricing and software value support a higher path. Competition, customer bargaining or compute costs could produce a lower one. The same-hurdle tests in the valuation chapter make that uncertainty consequential.

Which obligations consume modeled cash?

Commitment face values and expected losses answer different questions. The model already pays for products, ordinary research services and infrastructure through margins, working capital and capital spending. It separately values July’s financial investments. Its middle path assumes no additional net customer-support losses after useful services and recoveries. That is a meaningful assumption, not an absence of exposure. F01, Note 10 [S011].

Cash treatment and the boundary of the model

These are July schedules except the identified subsequent guarantee. A future investment closing is not cash already spent in July, and an individual partner arrangement may already sit inside a reported aggregate. The support analysis therefore uses incremental net losses after offsets, rather than adding every public announcement.

A joint failure of customer economics

The joint test keeps the first two years unchanged, then lowers sales by 20% and operating margins by ten percentage points from the middle path. Year-three revenue becomes $680bn after $730bn in year two. The model receives no working-capital cash release from that decline: collections and stranded inventory need not improve when customers weaken.

From year three it also charges $10bn annually of incremental, after-tax net support losses, after research use, resale, revenue sharing, reassignment and other recoveries. This severity assumption excludes costs already reflected in margins, capital spending or working capital. Extending it into continuing value assumes support remains economically necessary after current contracts expire. Recovery on existing financial equities falls to zero once. This connects lower sales, pricing, collections, support recoveries and investment values without turning every commitment into immediate debt. No output from this test is a price floor.

The addition case must clear the same hurdle

A stronger countercase preserves the same first two years, then raises sales by 20% and operating margins by three percentage points relative to the middle path. It uses the same 11% required return. In year ten, revenue is $1,524bn and operating margin is 53%. It requires a successful supply ramp, sustained paying utilization and renewal demand, and less erosion of NVIDIA’s pricing and system value. Those conditions are plausible; they are not assured.

These competing paths support reducing the position. The modest addition-case cushion is vulnerable to more demanding capital needs, while material missing customer-return and support evidence weighs against concentrated ownership. Strong existing profits, funding access and favorable paths support a smaller holding, so the evidence does not justify a broader reduction. Evidence: rating rationale.

The annual middle path

Cash before financing = operating profit − ordinary tax + depreciation − capital spending − extra working capital. USD billions, rounded for display. The first year requires $57.406bn additional working capital; capital spending at 3.5% of sales exceeds the trailing 2.47%.

Strong growth, fading advantage · USD billions · rolling years ending September

The opening-period approximation. The $302.970bn revenue base is the reported trailing year ending July 26, 2026, retained as a proxy for September opening conditions. It is not September trailing revenue. To illustrate the timing difference only, remove 55/91 of fiscal Q3 FY26 sales ($57.006bn) and add 56/91 of Q3 FY27 guidance ($108bn), assuming uniform sales within each quarter. That gives $334.977bn, reduces first-year working-capital investment from $57.406bn to $51.005bn, and adds $0.237 per share if everything else is held fixed. Guidance is not actual revenue; this is not a full rollforward of cash, receivables, inventory or investments and therefore is not a corrected September valuation. The published core values retain the disclosed July proxy. FY26 Q3 filing [S045]; current guidance [S015].

Test business economics before changing the hurdle

The 11% requirement is a judgment, approximately six points above the 5.01% ten-year Treasury yield. It was chosen before the original model outputs. Debt is less than 1% of enterprise value, so the difference from an equity hurdle is small; this is not a measured cost of capital or fitted historical risk coefficient. M03 [S106], M04 [S107].

Business economics at the same 11% requirement

All-years changes are mechanical sensitivities, including continuing economics; they are not independent revenue forecasts. In particular, the all-years upside implies $708bn first-year sales, above our $590bn timing bridge, and adding five margin points would imply 69% first-year operating margin. That requires stronger supply/timing and lower operating costs than the middle case. The final row is a more coherent addition case: it leaves the first two years unchanged and raises sales 20% and margins three points from year three. It reaches $1,524bn sales and 53% operating margin in year ten and gives $235.42 at the same 11% requirement. The approximately 5.9% excess over the quote is a conditional value difference, not an expected holding-period return.

CoreWeave’s reported mature-contract margins and Microsoft’s AMD machines with NVIDIA networking strengthen this favorable case. Amazon’s cash-spending gap, large future lease commitments and CoreWeave’s reliance on financing constrain it. These findings change the weight of the evidence, but supply no representative fleet, retained-share or cash-return denominator from which to estimate a precise new revenue or margin path. We therefore retain the same conditional cases and 11% requirement. Funding comparison (Evidence: customer funding); contract economics (Evidence: mature returns); replacement and competition (Evidence: replacement evidence).

Continuing economics and return sensitivity

We assume 3% nominal company growth after year ten. This is not a market-share estimate: markets growing more slowly would require share gains. The middle formula implies roughly 48% return on incremental operating capital. Its asset-light economics therefore remain demanding well beyond the detailed forecast. About half of operating value comes from years after ten, where small assumption changes compound. A direct terminal reinvestment check (Evidence: terminal capital) tests lower returns on new capital.

Same strong-growth path, different required returns and permanent growth

Holding the first ten years unchanged, replace all terminal net capital spending and working-capital investment with after-tax operating profit × 3% growth ÷ an assumed return on new capital. At 20% that return gives $181.90 per share; at 15%, $177.04. The replacement formula increases continuing reinvestment; it is not added on top of the old capital charge. These conditional checks lower value but still assume a profitable continuing franchise.

Developed addition case · same first ten years and 11% hurdle

The developed addition case retains the first ten years of its original sales and cash flows in all three rows. Only the continuing reinvestment requirement changes: after-tax operating profit × 3% growth ÷ assumed return on additional capital replaces the original net capital-spending and working-capital charge. The $235.42 case therefore depends on capital efficiency as well as growth and margins. The 20% and 15% tests are not measured future returns. The broader favorable $331.46 path remains disclosed, but neither favorable path has an established probability.

From obligations to cash available for owners

The opening asset bridge adds $56.586bn cash and debt securities, reserves $25bn for operations, includes 80% of $47.740bn public equities and 50% of $51.157bn private investments, then deducts $33.5bn debt principal. The net allowance is $61.8565bn, or $2.55 per share. Recovery percentages are judgments. July balances are carried forward because intervening movements are not fully known. Operating lease costs remain in margins; no second lease-debt deduction is made.

The commitment map above identifies costs already paid through margins, capital spending or working capital. Future equity commitments acquire assets: excluding both outlay and acquired value assumes an equal-value exchange, while any shortfall belongs in incremental net support. That premise covers the disclosed $25bn committed investments; discretionary future support must also be tested if needed to sustain sales. Neither promise of repayment nor face amount establishes recovery.

One correlated customer stress · first two middle years unchanged

The first row changes only sales and margins from year three. The second also withholds the $10bn working-capital release otherwise generated by sales falling from $730bn to $680bn; charges $10bn of additional net support annually in years three–ten and $10.3bn in year eleven growing 3% thereafter; and removes the $63.7705bn allowance for existing equities once. Net support is after useful services, resale, reassignment and recoveries, with no extra tax credit. It includes any residual guarantee losses within that assumed cost, not a further $105bn face-cap charge.

The last row additionally ties up $15bn in receivables/inventory in year three, recovering that extra cash once in year five. This delay is separate from the blocked release and does not incur a second expense. Its approximately $0.09 per-share effect is small because it is temporary; permanent loss would cost more. All stress magnitudes are analyst illustrations, not calibrated loss rates. The original $75.86 setback starts earlier and is more severe over the long run: year-ten sales are $625bn versus $1,016bn in this operating stress, with higher tax, capital-spending and working-capital rates and lower continuing growth. Moving only the same operating shock from year three to year one changes $130.11 to $122.63. The remaining difference comes from the other path assumptions, not timing alone. The joint support and recovery charges are separate from that controlled comparison.

An immediate $25bn zero-value cash loss changes value by $1.03 per share; a hypothetical immediate unrecovered $105bn guarantee loss changes it by $4.32. Such arithmetic neither dates nor predicts a call and cannot bound associated sales losses. The pending Hugging Face purchase plus maximum retention totals $12.9bn; an upfront zero-value sensitivity is $0.53 per share, without assuming consideration already paid or free acquired earnings. Do not stack these illustrations on the joint stress without identifying a separate, nonduplicated loss. F01 [S011], M25 [S128], M26 [S129], R02 [S137].

What would justify $222.27?

At middle margins, taxes, capital needs and 11%, 17.20% more sales throughout the forecast matches the quote: $691bn in year one and $1.49tn in year ten. This solves one model, not a unique market forecast. Better margins, a lower hurdle or longer growth can substitute. At 9%, the unchanged middle path gives $255. If comparable cash per share doubles at an unchanged price-to-cash multiple, price doubles; if the multiple halves, price is unchanged. Business growth alone does not determine stock returns. We have no supported exit multiple or calibrated 90-day price distribution.

Why 4/10 · Trim

Trim direct exposure; retain a smaller holding where consistent with diversification. At $222.27, even a stronger-growth scenario leaves little upside if sustaining that growth costs more than expected. We cannot yet establish whether customers broadly earn enough cash to sustain their spending, how promptly NVIDIA will collect what it is owed, or how much financial support it could recover if customers struggle. Reported mature operating margins, financing access and networking content preserve a credible ownership case, but do not justify merely waiting for another opportunity before reducing concentrated exposure. Evidence: rating rationale.

Why not 5/10 — Consider Other Opportunities? The price's demanding economics and material evidence gaps, together with concentrated customer and support exposures, support an actual reduction. Why not 3/10 — Reduce? Strong profits, funding, renewals, the $235 addition case and $331 favorable path can support ownership under reasonable competing assumptions. The negative comparison is not robust enough for a larger reduction. No probability-weighted expected return is asserted. Evidence: rating rationale.

With unchanged economics, a price near $170 reopens the valuation review, about 10% below the middle conditional value. It is neither an order nor a forecast. Evidence of customer cash returns, collection and retained system economics would first weaken the case to trim; buying requires stronger evidence and an attractive price. Persistent credit deterioration, unrecovered support or economically meaningful workload losses would favor 3 — Reduce or lower. Predeclared decision tests (Evidence: thesis tests) remain separately observable.

The S&P 500 fund SPY is the primary return comparison (7.96% NVIDIA on September 16); the semiconductor fund SMH supplies industry context (22.39% NVIDIA on September 9). Both retain NVIDIA exposure. Cash uses fixed 4.14% annual simple accrual over the actual interval between equity marks. The full-90-day illustration is 1.0208%; no purchased instrument or guaranteed outperformance is implied. These comparisons describe opportunity costs; no transaction is assumed. P06 [S006], P07 [S007], M03 [S106].

Predictions that can be wrong.

The assessment covers 90 calendar days from publication. Each of the five business forecasts has a numerical pass condition. The allocation judgment will be assessed against investment returns and business developments, without a numerical pass threshold. Management guidance, long-term scenario values and share returns are evaluated separately.

1. Sales reach $108–116 billion. For Q3 FY27, we expect revenue within this inclusive range, versus $96.221 billion in Q2. Management guides to $108 billion ±2%; our higher upper end allows another shipment beat. Any result outside our range is a miss, including a larger positive surprise.
2. Gross margin remains 73–75%. Q3 GAAP gross margin stays in this inclusive band, versus 75.0% in Q2. This allows some memory and transition pressure around management’s 74% ±0.5 percentage-point outlook. A result outside the band is a miss.
3. Cash conversion recovers. Q3 operating cash flow reaches at least 40% of quarterly revenue, from 25.0% in Q2. Our judgment expects some recovery after Q2 tax timing and a large receivable build. Below 40% is a miss and would challenge the case for retaining exposure.
4. Customer credit stops stretching rapidly. Quarter-end receivables divided by Q3 revenue, multiplied by the actual number of days in the quarter, stay at or below 65 days. Q2 was about 60. Above 65 is a miss; the measure tests cash tied up in sales, not whether a customer has defaulted.
5. Growth contains a customer’s interest burden. CoreWeave’s calendar Q3 net interest expense is no more than 25% of revenue, versus 24.85% in Q2 ($640 million / $2.575 billion). This predicts that growth keeps the financing burden from worsening materially. Above 25% is a miss and strengthens the funding concern. Passing challenges that concern, but does not establish ability to repay debt or independently fund new equipment.

Cash conversion uses quarterly operating cash flow on a comparable accounting basis. If the next filing revises first-half cash flow, the calculation uses the revised comparable figure; both it and the original $74.421bn will be shown with their disclosure dates. A result that cannot be calculated reliably is unobservable. Later restatements receive a separate assessment, preserving the initial result and the 40% threshold.

CoreWeave’s test uses the GAAP line “Interest expense, net” divided by quarterly revenue, with expense as a positive cost and any net interest income as a negative cost. This includes accounting accruals; it does not measure cash interest, principal repayment or total funding capacity. Baseline: CoreWeave’s Q2 release [S057].

Definitions and deadlines

Tests 1–4 cover NVIDIA Q3 FY27, expected with the November 17 results. Revenue and margin use reported GAAP amounts. Quarterly operating cash flow equals nine-month year-to-date cash flow less comparable six-month cash flow; cash conversion divides this quarterly amount by quarterly revenue without annualizing. Receivable days equal net quarter-end accounts receivable divided by quarterly revenue, multiplied by the actual days between fiscal quarter ends. This measures cash tied up in sales, rather than overdue invoices or defaults. Reported quarter dates and units govern the calculation; thresholds remain fixed.

Test 5 covers CoreWeave’s calendar quarter ending September 30, 2026. Results must appear in a filing or issuer release by the fixed 90-day endpoint. Missing data are unobservable. A value on an inclusive boundary passes; a value outside a forecast band misses even when economically better. Restatements are assessed separately with their disclosure dates, preserving the original assessment.

The revenue forecast follows management’s guide with a modest beat allowance; margin allows a wider band around guidance. Cash conversion and receivable days test the credit concern. CoreWeave’s interest burden tests whether growth is increasing financial strain. One quarter can challenge a mechanism without establishing durable customer returns.

Five tests of the durable investment thesis

Quarterly forecast passes do not establish a decade of strong economics. The following decision tests cover the next two reporting cycles: NVIDIA Q3/Q4 FY27 and counterparties’ next two published quarterly updates. They are evidence rules, not additional numerical forecasts. Record the disclosure date and period; evidence arriving after the original 90-day endpoint receives a later, separate assessment.

Evidence that changes the operating assumptions

Oracle illustrates the funding distinction: its August quarter’s $23.103bn operating cash flow includes an $11.363bn increase from customer advances carrying a significant financing component. These are real receipts that improve funding capacity; they do not by themselves establish cash earned from delivered services. Ordinary advances also retain their economic value. Neither GAAP classification nor a mechanical subtraction proves durable self-funding. Oracle Q1 FY27 10-Q, cash-flow statement and Note 1 [S056].

No single favorable headline automatically changes the score. A future qualitative review must state which assumptions changed and how the evidence affects the investment case. It may support 3 — Reduce, continued 4 — Trim, 5 — Consider Other Opportunities or, with sufficient positive relative advantage and confidence, 6 — Consider Buying. Missing evidence remains factually unresolved but weighs adversely in the rating. These judgments do not alter the independently scored business forecasts.

How the investment action will be assessed

The starting price is the first regular-session close strictly after publication. The ending price is the last regular-session close at or before publication plus 90 × 24 hours in UTC. NVIDIA, SPY and SMH use matching U.S. sessions and verified historical USD closing prices. The measured interval is the elapsed UTC seconds between these closes divided by 86,400; it can be shorter than the full 90-day window. The $222.27 valuation reference is not the performance starting price.

Total returns include distributions held as cash, before individual fees and taxes, with consistent adjustments for splits and other corporate actions. Ordinary dividends and ETF cash distributions count when their ex-date falls strictly after the starting session through the ending session, including earned amounts not yet paid. Buying at the close on an ordinary ex-dividend date does not earn that dividend. Special dividends, stock distributions and other corporate actions follow their verified entitlement terms. SEC entitlement guidance [S009].

The assessment will show NVIDIA’s return, its difference from SPY in percentage points, SMH’s return, and maximum peak-to-trough decline in daily closing total-return wealth. Drawdown therefore excludes intraday movements. Halts, delistings and unavailable prices will be identified with any known proceeds; missing marks will remain explicitly limited or unobservable, without substituting prices or benchmarks.

The cash comparison uses a fixed hypothetical annual rate of 4.14%, accrued simply over the same interval: 4.14% × elapsed days / 365, before tax and cost. A full 90 days would earn 1.0208%; the matched comparison can be shorter. This is an opportunity-cost convention, without assumed transactions or portfolio weights. Any actual replacement investment needs its own decision and comparison, including costs.

The allocation assessment evaluates our recommendation to trim. It evaluates the consequences of reducing direct exposure alongside subsequent business evidence, without inventing a transaction or portfolio weight. It is not an unconditional exit recommendation or a scored investment success rate. The five business forecasts retain their separate numerical tests and original outcome definitions.

We aim to revisit the research roughly quarterly, with earlier reassessment warranted by material credit, product, regulatory or funding developments. Subsequent research does not extend the original outcome deadline. Business evidence disclosed after that deadline receives a separately dated assessment.

Sources, judgment and what remains unknown.

This study connects NVIDIA’s reported profits with customer funding, cash collection and the capital required for future growth. The valuation shows the operating results each investment case needs; it does not assign probabilities to those cases.

Scope of the research

The limits that matter to the decision

The main gaps are cash returns from representative mature customer deployments, the timing and profitability of equipment replacement, and NVIDIA's retained system economics. Public terms also leave uncertainty about overlapping commitments and recoveries if a customer fails. These omissions weigh against ownership according to their importance and increase the qualitative weight assigned to adverse scenarios. They do not establish that customer returns are poor or that guarantees will be called. Evidence: rating rationale.

Comparable used-equipment transaction prices and customer-level utilization were not available. Primary short-interest data were unavailable at retrieval, so the analysis uses no exact current positioning estimate. Nasdaq source record [S008]. Direct discussions with procurement owners and operators could help answer some of these questions; this study used public evidence only.

Dates and disclosure

The reference price is $222.27, the September 18 regular-session close; NVIDIA’s latest accounts end July 26. Counterparty periods appear beside their figures.

As confirmed September 21, 2026, the publisher has no NVIDIA holdings or options, planned trades, compensation, or business or personal relationships involving NVIDIA or this research. Prepared with AI assistance.

Read the Source Register →