6 September 2026. Built from NVIDIA's SEC filings — 10-Ks for FY2017 to FY2026, 10-Qs through fiscal Q2 2027 (filed 26 August 2026), 8-Ks through 3 September 2026 — supplemented by primary outside sources: peer company results releases, the US Bureau of Industry and Security and Federal Register, SIA, WSTS, SEMI and NERC. Not a valuation and not a recommendation.
NVIDIA commits to TSMC wafers, CoWoS packaging and HBM memory more than a year ahead of demand, has Foxconn, Wistron and Fabrinet assemble the racks, and sells them at a 75.0% gross margin. It discloses no unit counts and no prices. What it does disclose is the economics of a dollar: of every $100 of revenue in the quarter ended 26 July 2026, $25.0 was cost of revenue, $8.7 was operating expense, and $66.2 was operating profit.
The customer problem is capacity. Training or serving a large AI model needs far more computation than one processor can deliver, so the work is spread over thousands of chips that must behave as though they were a single machine. Splitting the work is the easy half. Moving data between the chips fast enough that they do not sit idle is the hard half, and that is what NVIDIA actually sells. The chips come with it.
This has changed what leaves the building. Through FY2024 NVIDIA mostly shipped HGX boards, which a server maker built into a machine. The FY2026 10-K describes Data Center offerings as "typically delivered to customers as rack-scale systems, subsystems, or modules," and attributes the year's gross-margin decline to the model transitioning "from offering Hopper HGX systems to Blackwell full-scale datacenter solutions." NVIDIA now sells the rack, not the part — which is why the rack is the right unit to think in.
The cycle starts more than a year before revenue. NVIDIA commits to wafer capacity at TSMC, to CoWoS advanced packaging and to HBM memory, citing "extended lead times of more than 12 months" for which it has "paid premiums, provided deposits, and entered into long-term supply agreements and capacity commitments." Contract manufacturers assemble the systems. NVIDIA then bills a direct customer that is usually not the operator of the machine — an ODM, OEM, system integrator or distributor — which is why 76% of Data Center revenue billed to Taiwan-headquartered customers was attributed to end customers in the United States and Europe in FY2026. Cash arrives roughly sixty days later, against a $63.1bn receivable balance, and the Q2 FY2027 10-Q notes that "for investment-grade customer purchases, we have and may in the future provide longer payment terms ranging from 90 days up to one year."
None of this is recurring revenue. "Most of our sales are made on a purchase order basis," the FY2026 10-K states, and customers "can generally cancel, change, or delay product purchase commitments with little notice to us and without penalty." Contractual obligations extending beyond a year were $3.2bn at 26 July 2026 — against $96.2bn of revenue in that quarter alone. There is no backlog cushioning this business in any accounting sense, which matters more than any other single fact on this page.
One further change is worth flagging because it removes information. From Q1 FY2027 NVIDIA reports only Data Center — split into Hyperscale and AI Clouds, Industrial & Enterprise — and Edge Computing, into which Gaming, Professional Visualization, Automotive and OEM have been folded. Gaming can no longer be seen separately, and the FY2017–FY2026 platform series cannot be bridged forward from the filings alone. Software is described but never measured: NVIDIA sells AI Enterprise and vGPU licences, concedes in its own risk factors that "we may fail to sell any meaningful standalone software or services," and discloses no software revenue figure at all. Every disclosed dollar is hardware.
The industry converts scarce manufacturing capacity into AI compute, running from EDA and IP through TSMC's leading-edge fabrication and CoWoS packaging, through HBM memory from SK hynix, Micron and Samsung, to accelerator designers, then to ODM and OEM assemblers, then to clouds and model developers. The useful question is not how large the industry is. It is which link keeps the profit — and each participant's own reported margins answer it without anyone's interpretation.
| Stage / company, own most recent report | Revenue | Gross margin | Operating margin |
|---|---|---|---|
| Micron — memory (FQ3'26) | $41.5bn | 84.6% | 80.4% |
| SK hynix — memory (2Q26) | ₩79.3tn | n/d | 76% |
| NVIDIA — accelerator design (FQ2'27) | $96.2bn | 75.0% | 66.2% |
| TSMC — foundry and packaging (2Q26) | $40.2bn | 67.7% | 60.3% |
| Broadcom — custom accelerators (FQ3'26) | $29.6bn | n/d | 54% |
| AMD — Data Center segment (Q2'26) | $6.7bn | n/d | 31.3% |
| Marvell — custom silicon (FQ2'27) | $2.7bn | 53.1% | 16.8% |
| Dell ISG — systems (FQ2'27) | $31.8bn | n/d | 15.0% |
| Supermicro — systems (FY2026) | $39.1bn | 10.8% | n/d |
Each figure comes from that company's own results release for the period shown. Fiscal periods differ and margins are GAAP unless the company reports otherwise. Broadcom's 54% is company-wide GAAP operating margin, not an AI-only figure; SK hynix does not disclose a gross margin.
The pattern is not the one the industry narrative implies. The best margins in the chain are not at the accelerator designer — they are at the memory makers, whose 76% and 80% operating margins exceed NVIDIA's 66.2%, and at TSMC. The worst are wherever a participant can be replaced with capital and effort: Dell's infrastructure business at 15.0%, Supermicro at a 10.8% gross margin across its 2026 fiscal year. Value accrues to whoever controls a physically constrained step, and it is thin everywhere else. This single table explains more about NVIDIA's position than any market-share statistic would.
Scale is not in question. Data Center revenue of $89.0bn in a single quarter compares with roughly $25.6bn of quarterly AI-relevant silicon at its three nearest rivals combined — Broadcom's AI semiconductor revenue of $16.7bn, AMD's Data Center segment at $6.7bn, Marvell's data-centre revenue at $2.2bn. Intel's $6.3bn data-centre business is overwhelmingly Xeon CPUs; its Q2 2026 release does not mention an accelerator product at all.
Three barriers genuinely bind. The scale-up domain is the first: making 72 GPUs act as one accelerator over NVLink is not something a merchant network vendor currently replicates, and it is precisely why a rack rather than a chip is the sellable unit. Pre-purchased supply is the second — $279bn committed to manufacturing capacity and memory, which no entrant can match inside two years when TSMC's chief executive says on his own earnings call that "our packaging capacity is so tight that now it's limiting my customers' growth" and SK hynix's results release says "customer demand exceeds supply capabilities." The software installed base is the third, and the softest.
Two barriers leak, and the evidence is specific rather than speculative. The scale-out network layer has already gone to Ethernet, where Broadcom's Tomahawk switches sit across the hyperscalers building their own accelerators — and NVIDIA's own answer, Spectrum-X, is an Ethernet product rather than a defence of InfiniBand. CUDA does not bind at the top of the market either: AMD's Data Center segment moved from a $155m operating loss to $2.1bn of operating income in four quarters, and OpenAI has separately committed to 6 gigawatts of AMD Instinct and 10 gigawatts of Broadcom-built accelerators of its own design, while Anthropic contracted for up to a million Google TPUs. Customers with their own kernel engineers demonstrably can leave. Whether the long tail of enterprises can is not verifiable from outside evidence.
That is not a case against the position, but it does locate it precisely. NVIDIA's own scarce assets — the rack-scale interconnect and the developer installed base — are real, and outside evidence supports both its claim to accelerator dominance and its claim that demand exceeds supply; the latter is corroborated at four independent points in the chain (TSMC on packaging, SK hynix on memory, Dell's $95bn AI backlog, and $166bn of combined hyperscaler capital expenditure in one quarter). But those assets are engineering leads rather than physical scarcity, which makes them formidable and contestable at the same time, in a way TSMC's fabs are not. Outside evidence contradicts NVIDIA on one claim outright: China is not an addressable opportunity, and the company's own guidance now assumes no China data-centre compute revenue at all.
Revenue went from $60.9bn in FY2024 to $130.5bn in FY2025 to $215.9bn in FY2026, and reached $96.2bn in the single quarter ended 26 July 2026 — up 106% year on year and 18% sequentially, with guidance of $108.0bn ±2% for the following quarter.
Almost none of it is acquired, which is unusual enough to state plainly. NVIDIA has made exactly one revenue-additive acquisition of scale in a decade: Mellanox, closed April 2020 for $7.13bn, which contributed 10% of FY2021 revenue and created the networking line. Everything since bought capability, not revenue. The December 2025 Groq transaction — $13.0bn at closing plus $4bn payable within a year — was a non-exclusive technology licence plus employee hires, and the FY2026 10-K states that "no customer contracts, existing products, or equity interests were purchased." The Hugging Face agreement signed 2 September 2026, at roughly $11.9bn plus up to $1.0bn of retention equity, has not closed and post-dates every figure here. FY2026's 65% growth is organic; so is FY2027's.
Compute revenue rose 59% to $162.4bn in FY2026. Selling racks instead of boards raises the revenue NVIDIA captures per unit of end-customer capacity, because it now supplies the CPUs, switches and interconnect a system builder used to supply.
Networking grew 142% to $31.4bn in FY2026 on NVLink compute fabric for GB200 and GB300 systems plus Ethernet and InfiniBand. It is the Mellanox asset compounding six years after purchase, and it is the part of the rack a customer is least able to source elsewhere.
Microsoft, Amazon, Alphabet and Meta together spent $166bn in calendar Q2 2026; Meta guides to $130–145bn for the year; AWS said on 26 August 2026 it plans to deploy 2 million additional NVIDIA GPUs across 2027–28. This is a budget decision made annually by four companies, not a contracted revenue stream — the largest single dependency in the business.
Blackwell, then Blackwell Ultra from Q2 FY2026, then Vera Rubin entering production shipments in Q3 FY2027. Each generation raises performance per rack and resets the competitive comparison — but the ramp risk recurs every year, and FY2025 gross margin was hit by "inventory provisions for low-yielding Blackwell material" on exactly this pattern.
Gaming, visualisation and automotive, now consolidated, grew 27% year on year to $7.2bn — 7.5% of revenue. NVIDIA has warned supply constraints are a headwind here, because the same wafers earn far more in a data centre.
Chinese revenue fell from $25.0bn in FY2025 to $19.7bn in FY2026, and Hopper and H200 shipments were each under 1% of Data Center revenue in Q2 FY2027. This is not a growth lever today; it is a removed one.
Gross margin is set by two things NVIDIA controls and one it does not. It controls mix — a full rack carries a different margin from a board — and it controls how much inventory it writes down. It does not control the price of HBM memory. All three moved in FY2026: margin fell 3.9 points to 71.1% as the model shifted to full-rack solutions and as a $4.5bn charge landed on H20 inventory and purchase obligations after the April 2025 export-licence requirement. Provisions totalled $7.2bn against $1.5bn of releases, a net 2.6-point drag.
The first half of FY2027 reversed the charge effect, with margin back to 75.0% and a provision drag of one point. The next move is the informative one: NVIDIA guides Q3 FY2027 gross margin down to 74.0%, and the Q2 10-Q says supply availability and rising prices for memory and other components "may drive the prices for data center buildouts higher," describing the $279bn commitment increase as "primarily related to the procurement of memory." Input cost has turned from tailwind to headwind, and it is set by suppliers NVIDIA has no ability to substitute.
Operating leverage runs the other way and is why margins have held. R&D fell from 14.2% of revenue in FY2024 to 8.6% in FY2026 while rising 43% in dollars; selling and administrative expense fell from 4.4% to 2.1%. Roughly 42,000 employees, 31,000 of them in R&D, now support a revenue base that has trebled in two years. Operating margin reached 65.9% in the first half of FY2027 against a prior-decade peak of 37.3% in FY2022.
| $m unless stated | FY2020 | FY2023 | FY2026 | FY2027 H1 |
|---|---|---|---|---|
| Revenue | 10,918 | 26,974 | 215,938 | 177,837 |
| Data Center revenue | 2,983 | 15,005 | 193,737 | 164,269 |
| Gross margin | 62.0% | 56.9% | 71.1% | 74.9% |
| Operating income | 2,846 | 4,224 | 130,387 | 117,270 |
| Operating margin | 26.1% | 15.7% | 60.4% | 65.9% |
| Operating cash flow less capex | 4,272 | 3,808 | 96,676 | 69,987 |
Four caveats on comparability. Reportable segments changed from GPU/Tegra to Compute & Networking/Graphics in FY2021, so segment series do not span this table. The market-platform taxonomy was rebuilt in Q1 FY2027, so the FY2027 H1 Data Center figure is not defined identically to earlier columns. FY2023 operating income includes a $1,353m Arm acquisition-termination charge, roughly five points of margin. And two stock splits — 4-for-1 in July 2021 and 10-for-1 in June 2024 — make as-reported per-share figures non-comparable across these columns, which is why none are shown. Cash flow less capex is computed, not a reported line. Sources: FY2020, FY2023, FY2026 10-Ks; Q2 FY2027 10-Q.
FY2026 produced $102.7bn of operating cash flow against $6.0bn of capital expenditure; the first half of FY2027 produced $74.4bn against $4.4bn. This still requires almost no fixed capital of its own — the fabs belong to TSMC. But working capital has changed markedly: receivables went from $38.5bn to $63.1bn and inventory from $21.4bn to $31.6bn in six months, with raw materials alone going from $3.8bn to $11.3bn ahead of the Vera Rubin ramp. Days sales outstanding is around sixty, but with payment terms disclosed as extending to a year for investment-grade buyers, that number is a policy choice as much as a measurement.
Reported net income should not be read as operating performance from FY2026 onward. Other income of $24.1bn in the first half of FY2027 was 20.5% of net income and consists largely of unrealised marks on an equity portfolio that reached $99bn — which is why NVIDIA's own non-GAAP diluted EPS of $2.22 in Q2 FY2027 sits below its GAAP $2.46. Operating income is the comparable series.
| Use of capital, FY2017 to FY2027 H1 | $bn | What it says |
|---|---|---|
| Share repurchases | 135.6 | Suspended entirely FY2020–FY2022 while Arm was pending; authorisation raised $80bn in May 2026 |
| Purchases of equity securities in other companies | ~62.3 | Now the second-largest use of capital — buying stakes rather than building |
| Employee stock-plan tax withholding | 28.9 | Share-reducing, and larger than capital expenditure |
| Capital expenditure | 20.6 | Small: the manufacturing base is rented, not owned |
| Groq technology licence | 15.9 | Largest single commitment outside supply; no revenue acquired |
| Acquisitions, net of cash | 11.8 | Dominated by Mellanox in FY2021 |
| Dividends | 11.0 | Quarterly dividend raised from $0.01 to $0.25 in May 2026 |
Cash-flow basis, from the FY2017–FY2026 10-Ks and the Q2 FY2027 10-Q. The balance sheet changed shape alongside: total debt went from $8.5bn to $33.4bn on a seven-tranche $25.0bn senior note issue completed 18 June 2026 at coupons of 4.250% to 5.625%, against $99.4bn of cash and marketable securities. A company generating $70bn of free cash flow per half-year chose to borrow $25bn — a statement about intended deployment pace, not need. Pending and post-dating all figures above: the Hugging Face acquisition, ~$11.9bn plus up to $1.0bn of retention equity, signed 2 September 2026, expected to close in the first half of 2027.
NVIDIA is at or through its own prior peak on every axis that matters. Gross margin of 75.0% matches the FY2025 record; operating margin of 66.2% is a decade high against a prior peak of 37.3% in FY2022; revenue is growing 106% year on year. Nothing in this business is currently below trend — which is the necessary context for every margin and growth figure above, and the opposite of the situation in which the same numbers would be encouraging.
The company's own history shows the other end. In FY2019 a crypto-driven channel-inventory correction cut fourth-quarter revenue 24% year on year, took gross margin from 61.9% to 54.7% and produced $185m of DRAM and component charges; FY2020 revenue fell 7% and operating income 25%. In FY2023 gross margin fell to 56.9% and operating margin to 15.7%. Both episodes followed periods when demand looked entirely secure.
What is different today is that the transmission runs through the balance sheet rather than the channel, and it is far larger. Four companies fund most of the demand — $166bn of capital expenditure in a single quarter. Against that, NVIDIA holds $279bn of non-cancellable purchase and capacity commitments ($92bn due in the remainder of FY2027, $87bn in FY2028, $88bn in FY2029) and $31.6bn of inventory, while its customers can cancel purchase orders without penalty. A capital-expenditure pause would not reduce NVIDIA's obligations. It would convert them into charges, exactly as the $4.5bn H20 write-down did when one export rule removed one market. The industry precedent is Micron in FY2023: revenue down 49.5% and gross margin from 45.2% to minus 9.1% in four quarters. That is what happens to whoever holds the scarcity position when the cycle turns — and today those holders are the memory makers, to whom NVIDIA has just committed $279bn.
Three constraints bind on the supply side, all documented outside NVIDIA. TSMC's advanced packaging is limiting customer growth on the company's own statement, and only 10–20% of its $60–64bn 2026 capital budget goes to packaging, testing and masks. HBM is contracted forward, with SK hynix reporting long-term agreements with around ten customers. And electricity is the constraint furthest from anyone's control: NERC's 2025 Long-Term Reliability Assessment classifies 2026–2030 as high risk, with planned resources falling short of established criteria.
| Indicator | What it would tell you | Where published |
|---|---|---|
| Gross margin against guidance | Whether memory cost is passed on or absorbed | NVDA 10-Q and results release |
| Purchase and capacity commitments | Size of the forward supply bet — $95bn to $279bn in two quarters | NVDA 10-Q commitments note |
| Receivables, DSO and disclosed payment terms | Whether growth is financed by extending credit to buyers | NVDA 10-Q balance sheet and MD&A |
| Guarantee and equity-investment balances | How much end demand NVIDIA underwrites itself | NVDA 10-Q commitments and contingencies |
| Combined capex of Microsoft, Amazon, Alphabet, Meta | The demand budget, one quarter ahead of NVIDIA's revenue | Each company's 10-Q and release |
| Broadcom AI revenue against its own outlook | Whether custom accelerators take share at the frontier | Broadcom quarterly release |
| AMD Data Center revenue and segment margin | Whether a second merchant supplier becomes economic | AMD quarterly release |
| HBM pricing and margins at Micron and SK hynix | Where pricing power in the chain is moving | Micron and SK hynix releases |
| TSMC advanced-packaging capacity and capex split | When the binding physical constraint loosens | TSMC quarterly report and call |
| Changes to US advanced-computing export rules | Whether China reopens or closes further | Federal Register; Bureau of Industry and Security |
Cyclical risk is covered above. What follows is what cyclicality does not capture, ordered by how badly each compounds with the others.
At 26 July 2026 it held $99bn of equity investments and $25bn of further investment commitments; guarantees with maximum gross exposure of $108.5bn — $105bn of residual-value guaranties to SB Energy covering roughly 4.25 gigawatts leased to OpenAI for twenty years, plus $3.5bn of land, power and shell guarantees to AI clouds; and $36bn of six-year cloud service commitments. The 10-Q states the mechanism directly: "if AI clouds do not successfully sell committed capacity to third-party customers, we have agreed to purchase that capacity." In August 2026 NVIDIA signed memoranda with six capital providers to mobilise more than $500bn of third-party capital. Each arrangement is defensible alone. Together they mean a downturn arrives three times over — as lost revenue, as impaired investments, and as assumed obligations.
The Q2 FY2027 10-Q states that AI clouds and model makers "have significant demand for training and inference compute and currently lack the ability to secure long-term infrastructure contracts and investment-grade financing capacity." The buyers of the most expensive capital equipment in the world cannot finance it conventionally. That is why the guarantees exist, and it makes NVIDIA's revenue a function of credit conditions in a way a hardware business normally is not.
Two direct customers were 22% and 14% of FY2026 revenue; three receivable balances were 25%, 18% and 13% of the total. Those customers order on cancellable purchase orders while NVIDIA holds $279bn of non-cancellable commitments. The asymmetry is the risk, not the concentration by itself.
Roughly a fifth of first-half FY2027 net income came from unrealised gains on holdings whose value moves with the same AI capital cycle that drives the operating business. NVIDIA disclosed that a 10% decline in public equity holdings would have cost $3.9bn as at April 2026. The portfolio does not diversify the earnings stream; it levers it.
$15.9bn paid or committed for a non-exclusive technology licence and a team, producing $14.4bn of goodwill and a $2.5bn intangible with no acquired customer contracts, products or equity. The 10-K concedes the payments were "nonrefundable" and that "we may be unable to recover the associated costs." Its return is unverifiable from disclosure.
The FY2026 10-K no longer frames China as a risk but as an accomplished loss: NVIDIA was "effectively foreclosed from competing in China's data center computing/compute market," and states that this foreclosure "helped our competitors build larger developer and customer ecosystems to challenge us worldwide." A 25% tariff applies to H200 units imported into the US and "we have been unable to pass along any of the tariff to our customers." Chinese antitrust regulators published a preliminary finding on 15 September 2025 that NVIDIA's compliance with US export controls violated the conditions of their Mellanox approval, which reaches the networking business. The GAIN AI Act and the Remote Access Security Act are named in the filings as further exposure.
These are findings, not gaps in the research. Several would change how an investor sizes the business.
Unit volumes and average selling prices. Never disclosed, so growth cannot be decomposed between volume, price and content per rack from filings alone.
Software, services and DGX Cloud revenue. Described in the business section; quantified nowhere.
The maintenance-versus-growth split of capital expenditure, and the identity of the 22% and 14% customers.
Dollar amounts for several referenced investments. The FY2026 10-K describes the OpenAI arrangement only as being "finalized," with no figure; the Intel and xAI stake sizes are not disclosed in filings.
Whether the $1 trillion Blackwell-and-Rubin figure management cites on earnings calls has any contractual basis. Filed remaining performance obligations beyond one year are $3.2bn, which suggests not, and no filing repeats the number.
The share of hyperscaler AI compute now served by custom silicon. Neither Amazon nor Meta discloses deployed Trainium or MTIA scale, so the substitution rate cannot be measured from primary sources.
1. This is an AI data-centre systems business, not a chip business. The sellable unit is a rack of 72 GPUs, and the interconnect that makes them act as one machine is the part competitors have not reproduced.
2. The economic engine is turning pre-purchased scarce capacity — TSMC packaging and HBM memory — into integrated systems at a 75% gross margin, sold to a handful of buyers on cancellable orders.
3. The main growth lever is Data Center content per rack and networking attach, both structural. The pace at which they convert to revenue is set by four customers' annual capital budgets, which is not.
4. What could break the story is $279bn of non-cancellable supply commitments and $108.5bn of guarantees meeting a pause in customer funding — a sequence in which NVIDIA absorbs the correction rather than passing it on.
5. Monitor gross margin against memory cost, the commitments and guarantee balances, and hyperscaler capital expenditure. Those three lines describe the whole risk.