NVIDIA sells accelerated computing systems: GPUs, CPUs, NVLink and network fabrics, reference rack designs and the CUDA software stack that ties them together. Data Center produced USD89.0bn of USD96.2bn Q2 FY2027 sales (nvda-q2-fy2027-release). The rest ("Edge": gaming, professional visualization, automotive) is a mature franchise. Manufacturing is outsourced to TSMC, memory makers, packaging houses and ODMs. External manufacturing does not remove capital needs: supply commitments, inventory, receivables, ecosystem equity stakes and customer credit support all use shareholder capital.
Revised 3 Nov 2026 for the research pipeline. Fiscal years end on the last Sunday of January, so FY2026 is roughly calendar 2025. Money is in USD millions unless stated. Fact IDs are in financials/facts.json.
Value-creation mechanism
NVIDIA creates value by turning a fixed, front-loaded R&D budget (USD23.6bn trailing, researchFY2026 etc.) into a new full AI computer system every year. It then secures enough of the world's scarce leading-edge wafers, advanced packaging and high-bandwidth memory to deliver that system before any alternative can. Customers pay for useful compute delivered soon, and judge the price against their cost per token and the revenue that idle capacity forgoes. The integration step creates most of the value: chip, interconnect, networking and CUDA libraries tuned together, sold into every cloud. Management sizes its revenue opportunity per gigawatt of AI capacity at about USD18bn for Hopper, USD25bn for Grace Blackwell and USD40bn for Vera Rubin (revenuePerGwRubin, nvda-q2-fy2027-call pp.4 and 10). Each generation captures more of the customer's capacity budget.
Supply orchestration is the second value step. NVIDIA commits ahead of demand: USD279bn of supply and capacity commitments at July 2026, mainly memory, up from USD95.2bn in January (supplyCommitments, supplyCommitmentsFY2026). This makes NVIDIA the bottleneck owner in a capacity-constrained market, which lets a 75% gross margin coexist with rising component costs. The third step is newer, and its value is unproven: NVIDIA lends its balance sheet to customers without investment-grade credit. That takes the form of equity stakes in AI labs (nearly USD50bn, aiLabInvestments), guarantees (USD108.5bn maximum) and extended payment terms (DSO 60 days, dsoQ2). This step creates demand and also absorbs capital and risk.
Business-model profile
Comparable firms are chosen by these traits, not by the industry label.
Growth vehicle: Organic. Revenue rose from 26,914 in FY2022 to 215,938 in FY2026, driven by data-center demand for its own platforms. Acquisition cash was 0.6% of revenue over FY2024–FY2026. Growth is organic: it comes from customers' capital spending on accelerated computing and is funded by the company's own operating investment. The model forecasts organic units and capacity, with no deal budget.
Capital intensity: 2.5% of sales, asset-light. Capital spending is about 2.5% of revenue because manufacturing is outsourced to TSMC, memory makers, packaging houses and ODMs. The capital growth uses sits elsewhere: inventory, receivables, supply prepayments and purchase commitments (tangible operating capital of 85.7bn at July 2026), plus ecosystem equity stakes and customer credit support.
Comparables. A comparable firm is a fabless platform supplier into a customer capital-spending cycle: revenue depends on a few large buyers' capex budgets, growth is organic and internally funded, and margins depend on a technology lead and a software ecosystem. Base rates must treat it as a very large supplier at the top of a spending boom. Not the label: "Semiconductors" mixes NVIDIA with capital-intensive foundries and commodity memory; it is a fabless systems and software platform exposed to its customers' capex cycle.
Declared in valuation/model.json (businessProfile). npm run check measures deal and capital intensity from the reported facts, and the outside view grades the forecast only against firms that share them.
Economic snapshot
Definitions:
- Trailing (TTM). FY2026 + H1 FY2027 − H1 FY2026.
- NOPAT. GAAP operating income × (1 − 17%), using the FY2027 guided tax midpoint, with stock compensation charged. Investment gains and interest are excluded.
- Balances. At 26 July 2026.
Capital base at July 2026:
Returns on that capital:
- Return on tangible capital, capitalizing leases: 191%.
- Return on capital employed, expensing leases (excluding the ROU asset): 204%. Including goodwill and intangibles: 157%.
- Including the ecosystem stakes as capital that now supports demand: 79%.
These are ending balances rather than averages, which flatters nothing because capital is growing.
Normalization: TTM capex is 7,474 against depreciation of 3,687. That is about twice D&A, which reflects growth capacity rather than under-maintenance, so no add-back is made. There are no minorities, and recent acquisitions are licenses and teams with no separately disclosed profit. The real normalization question is cyclical. The Q2 75% gross margin and 65% operating margin sit far above NVIDIA's own pre-AI record:
- FY2022: 64.9% gross, 37.3% operating.
- FY2023 downturn: 56.9% gross, 15.7% operating, including the Arm write-off and inventory provisions.
Stage 7 and the scenarios treat today's margin as a shortage-era level, not a through-cycle norm.
Economic segmentation
Reported segments are evidence, not the answer. In FY2026, Compute & Networking earned 130,141 of 139,297 segment operating income, or 93%, on a 67% segment margin. Graphics earned 9,156, a 41% margin (segmentComputeNetworkingOpFY2026, segmentGraphicsOpFY2026). Cut by how value is created and who decides the purchase:
The profit split between hyperscale and ACIE is an estimate. NVIDIA does not disclose margins by customer type. I assume similar gross margins, because hyperscalers' volume discounts and ACIE's extended terms and financing roughly offset.
The implication: about 90% of profit and nearly all of the value depend on one product cycle sold to two kinds of buyer:
- Five hyperscalers, all of which build their own accelerators.
- A set of capital-hungry AI labs and neoclouds whose purchases NVIDIA partly finances.
A sector label ("semiconductors") hides this. The demand is a capital-spending cycle, and the cycle's financing is partly NVIDIA's own.
Unit economics
Unit: one gigawatt of Vera Rubin AI capacity. Customers add capacity in gigawatt-scale AI factories, and management prices its opportunity per gigawatt. Revenue comes from management's sizing. The other lines are this repository's estimates, built from guided ratios.
Adding NVIDIA's financing changes the picture. OpenAI's ~12 GW is supported by part of the ~USD50bn in lab equity, about USD4bn per GW (nvda-q2-fy2027-call p.6). With that included, capital becomes about USD13bn per GW and the return about 160%. The NVIDIA-side unit is therefore not capital-constrained: its value depends on price per gigawatt, not on capital intensity.
The consensus question is the customer's unit, because their returns fund NVIDIA's price. Management puts a full gigawatt at about USD60bn of customer spend today (p.9). Servers are depreciated over five to six years at Amazon, Meta, Alphabet and Oracle (amzn-q3-2026, meta-q3-2026, googl-q3-2026, orcl-q1-fy2027). On that basis the customer carries roughly 8–10bn of annual depreciation per GW. Power at an assumed USD0.08/kWh adds about 0.7bn. Earning 10% pre-tax on USD60bn therefore needs about USD15–17bn of annual revenue per GW. If annual product cadence shortens the economic life to three years, the requirement rises to about USD22bn+.
Customer revenue is responding:
- Google Cloud +103% to 30.7bn in Q3.
- AWS +39% to 46.0bn.
But customers' capex now exceeds their operating cash flow:
- Alphabet free cash flow −9.1bn in Q3.
- Amazon trailing free cash flow −42.4bn.
- Meta −2.1bn in Q3.
- Oracle −5.4bn, while selling USD20bn of stock.
The customer unit is not yet self-funding at the margin.
Failure case: H20 for China. NVIDIA designed a product to fit export rules, then a regulator closed the market. The result was a 4.5bn inventory and purchase-obligation charge, about USD60m of licensed revenue, and no H200 revenue under the February 2026 license (inventoryProvisionsFY2026, nvda-fy2026-10k). The levers that work with customers (performance, supply, software) fail when the counterparty with the outside option is a government.
An earlier failure was FY2023. Inventory doubled (2,605 to 5,159, inventoryYeFY2023), the gross margin fell 8 points and operating income fell 58% while revenue was flat. Shipped capacity outran end demand.
Independent checks
- Segment accounts. Compute & Networking earned a 67.3% segment margin in FY2026, before unallocated stock compensation and acquisition costs. Group GAAP operating margin was 60.4%. The per-GW model's 64% sits between them, consistent with allocated fixed R&D.
- Customer budgets. The five hyperscalers' own 2026 guidance sums to about 810–830bn:
- Microsoft about 170bn.
- Alphabet 195–205bn.
- Amazon about 220bn.
- Meta 135–145bn.
- Oracle FY2027 about 90–95bn. This figure comes from a call summary and is unverified. NVIDIA puts it at "nearly 800bn" (p.3). NVIDIA's hyperscale revenue annualizes to about 195bn, roughly 24% of that capex. That is consistent with 40bn of NVIDIA content in about 60bn per GW, once land, buildings, power and these customers' own TPUs, Trainium and MTIA are excluded.
- Competitors' audited revenue. In the quarter to early August, Broadcom's AI semiconductor revenue was 16.7bn (+221%) and AMD's Data Center revenue was 6.7bn (+107%) (
avgo-q3-fy2026, amd-q2-2026). NVIDIA's Data Center revenue was 89.0bn (+117%). NVIDIA still takes about 79% of these three firms' AI data-center revenue. Custom silicon is growing about twice as fast from a smaller base. - Adverse allocation. Charge all trailing R&D (23.6bn) and all stock compensation (7.2bn) to the hyperscale unit alone. Its operating margin still exceeds 55%. No unit is unprofitable today, so the question is durability of price, not current profitability.
Where the consensus disagrees: one camp calls this a supply-constrained secular buildout; the other, circular financing. Three observations would settle it:
- Receivable days and receivables growth against revenue.
- The share of revenue from customers NVIDIA finances.
- Whether hyperscaler free cash flow turns positive while capex keeps rising.
Stage 11 freezes these as prospective checks.