NVIDIA Corporation

01 — Executive summaryThe best business in the market, priced for the base case

NVIDIA reported revenue of $96.2 billion for the quarter ended 26 July 2026, up 106% from a year earlier, at a 75.0% gross margin and a free cash flow margin near 60%. Three months earlier it reported $81.6 billion; three months from now management expects $108.0 billion. The operating evidence is not in dispute.

What is in dispute is what the buyer is underwriting. Three things have changed since we last looked, and all three concern the quality of demand rather than its level.

First, the unit of sale has moved from a chip to a system. Networking grows two and a half times faster than compute, because a rack-scale GB300 carries switches, optics and data-processing units alongside the accelerators. That widens the moat but makes revenue hostage to one customer's capital budget: a rack decision is worth $3 million, not $40,000.

Second, a meaningful share of reported profit does not come from operations. GAAP net income in the first quarter of fiscal 2027 was $58.3 billion against $53.5 billion of operating income, and the first-half gap between GAAP and non-GAAP net income was $18.6 billion, mainly unrealized gains on equity stakes. We value the operating business on non-GAAP earnings and treat the portfolio separately.

Third, the seller has become the financier. In one quarter NVIDIA announced a $500 billion GPU-backed lending platform with six private-capital managers, a reported $40 billion debt package behind a single customer's order, and a $12.9 billion acquisition. That is a legitimate way to grow a market, and also how a slowdown would transmit through credit rather than orders.

On our numbers the shares are worth $250 in twelve months, 9% above the current price, with a probability-weighted value of $246. That is Neutral, not negative: at 19 times our fiscal 2028 estimate the stock is not expensive. The problem is the shape of the distribution: the bear case, requiring only that financing for the marginal buyer tightens, costs 41%, while the bull case pays 53% and requires uninterrupted execution.

02 — The unit of saleA rack, a network, and a subscription to the standard

The useful way to understand NVIDIA in 2026 is to stop thinking of a semiconductor company and start thinking of a systems integrator with a monopoly on the interface. Three layers carry the economics; only one is a chip.

The first is compute. Blackwell Ultra is the volume product; Vera Rubin entered full production in the second quarter and is ramping at CoreWeave, Google Cloud, Azure, OCI and Nebius. Each generation is now roughly a year apart rather than two, and each is claimed to deliver a five-to-ten-times improvement in performance per dollar — which makes designing your own accelerator a moving target.

The second is networking, the layer the market still underrates. A rack is only as fast as its internal links, so NVLink switches, Spectrum-X ethernet and BlueField units are sold with the silicon rather than alongside it. Networking grew 199% year over year in the first quarter against 77% for compute, and Spectrum-X alone passed a $10 billion annualised run-rate.

The third is software and the instruction set. CUDA, Dynamo, NIM and the AI Enterprise subscription convert a hardware lead into a switching cost, and Vera — the first NVIDIA chip built for agentic workloads — extends that logic to the central processor, where management expects close to $20 billion of standalone revenue this year.

NVIDIA no longer competes on a component specification but on whether a customer's software runs, which is far more durable than a performance lead and why we treat 80% share as a floor rather than a peak. It is also why the Hugging Face acquisition is coherent despite the scrutiny it will attract.

03 — The financial recordTwo quarters that reset what a large company can do

Fiscal 2026 closed with $215.9 billion of revenue; fiscal 2027 will close above $400 billion on any reading of the guide. The table sets out what has been reported.

Two observations matter more than the growth rate. The first is that sequential growth is decelerating — 22%, 20%, 20%, 18%, and 12% guided. Return from here therefore depends on the level of revenue being sustained rather than on the rate of change accelerating, and the market has spent two years pricing the rate of change.

The second is that the EPS series tells a different story. GAAP earnings per share exceeded non-GAAP in each of the last two quarters — $2.39 against $1.87, then $2.46 against $2.22. A company whose GAAP earnings exceed its adjusted earnings is reporting gains its own management considers non-recurring.

04 — Segment: Data Center92.5% of the company, and both engines running

Data Center revenue was $89.0 billion in the second quarter, up 117% year over year and 18% sequentially, now 92.5% of the business. It is reported as one line, but the first quarter's disclosure shows the split: compute $60.4 billion, up 77%, networking $14.8 billion, up 199%.

That divergence is the most informative number in the filing. The marginal dollar of AI capital expenditure is increasingly spent on the interconnect rather than the accelerator, which is what happens when the constraint moves from raw compute to keeping tens of thousands of chips on one problem. NVIDIA also captures that spend at a higher margin: ethernet switching and optical interconnect have fewer credible merchant alternatives than GPUs.

The implication runs against the conventional bear case. The argument that hyperscalers will substitute their own silicon assumes the substitution is for an accelerator; if the unit of sale is a rack defined by its interconnect, a customer-built accelerator still has to be wired into an NVIDIA-defined fabric or rebuilt around a different one. Google's TPU is a genuine exception because it controls its stack end to end; Trainium and MTIA less so.

05 — Segment: Edge ComputingThe ballast, and the option nobody is paying for

Edge Computing is the smaller platform and, at $7.2 billion, 7.5% of second-quarter revenue, growing 27% year over year and 13% sequentially. It absorbs the former gaming, workstation, automotive and embedded lines — businesses where inference happens away from a data centre.

Financially it is a ballast — low-double-digit growth, lower margins than Data Center, and working capital on a seasonal pattern unrelated to the AI cycle. Where Data Center moved $13.8 billion sequentially, Edge moved $0.8 billion. Its value is as a clean read on non-AI demand, and that read is unremarkable: all of the company's growth is coming from one end market.

As an option it is more interesting than its revenue implies: DRIVE Hyperion, Isaac GR00T, Cosmos and Halos describe a coherent bet that the next compute cycle is robotic, but none of it is in our numbers.

06 — Margin and operating leverageA 75% ceiling, a 66% floor, and no further leverage to give

Gross margin expanded 14.4 points over five quarters and has sat at or just below 75% for four straight periods. Management guides the third quarter to 74.0%, the first guided sequential decline of the cycle, citing higher rack-scale systems content and the cost of ramping Vera Rubin. We would add a third: a customer base large enough, and concentrated enough, to negotiate.

The point is not the one-point decline. It is that 75% is a ceiling, not a waypoint. A company selling a proprietary accelerator with no second source can price to value; a company selling a rack, a network and a software stack is selling a bundle, and bundles are priced against their alternatives rather than against their cost. Every incremental point of systems content dilutes the margin of the silicon inside it.

Operating leverage tells a similar story. Operating income grew 147% year over year in the first quarter against 85% revenue growth, then 124% against 106%. Operating margin has settled near 66%, close to the practical maximum for a business spending $9 billion a quarter on operating expense, mostly engineering payroll. From here earnings grow at roughly the rate revenue grows.

07 — Quality of earnings$18.6 billion of first-half profit did not come from operations

GAAP net income in the first quarter of fiscal 2027 was $58.3 billion against $53.5 billion of operating income — exceeding it by $4.8 billion. With almost no debt and a 16% to 18% tax rate, that is impossible unless something below the operating line contributed. In the second quarter, GAAP net income of $59.7 billion compared with non-GAAP net income of roughly $53.9 billion. Across the half the cumulative difference is $18.6 billion.

The mechanism is unrealized gains on equity investments. NVIDIA holds stakes in companies that buy its products — model developers, neoclouds, infrastructure ventures — and those stakes are marked to market through the income statement as private valuations rise.

Three points follow. The gain is real economically but not in cash: it cannot fund a buyback, a dividend or a fab, so treating it as earnings inflates the payout capacity of the business. It is correlated with the operating business in the wrong direction: if AI capital expenditure slows, the valuations fall and the mark reverses, so the same event that reduces revenue would produce a charge. And it flatters the trailing multiple: on trailing GAAP earnings the shares trade at 29 times, a number that includes a portfolio return.

08 — Cash and capital returns$48.6 billion of free cash flow, and $235 billion of authority to spend it

In the first quarter of fiscal 2027 NVIDIA generated $50.3 billion of operating cash flow and $48.6 billion of free cash flow, a 60% margin and roughly double the year-earlier quarter. Capital expenditure was $1.76 billion. The model is asset-light: the wafer capacity, HBM4 stacks and CoWoS packaging the accelerators depend on sit on other companies' balance sheets.

That combination — high margin, low capital intensity, high cash conversion — is what allows the capital returns programme to be as large as it is. The company returned $19.5 billion in the first quarter and about $26.0 billion in the second, $45.5 billion in the half. The board added $80 billion of repurchase authority in May and $150 billion in September, leaving roughly $235 billion authorised through fiscal 2028, and raised the quarterly dividend from $0.01 to $0.25.

Two cautions. $235 billion is roughly 10% of the current market capitalisation — meaningful support for the share price, not a floor for it. And repurchasing stock at 19 times forward earnings is accretive only if the estimate is right; a buyback is a forecast expressed as a capital allocation decision.

09 — Who is buyingSix customers, then governments, then everyone else

Hyperscale cloud operators were more than half of Data Center revenue in the first quarter of fiscal 2027 — $38 billion, up 12% sequentially. Sovereign AI programmes contributed roughly $30 billion cumulatively across fiscal 2026, about 14% of that year's total revenue and triple the prior year.

Concentration is usually read as a risk, and it is one. But the composition matters more than the level. The six largest buyers are among the highest-quality credits in the corporate world, and their capital budgets are set by a competitive dynamic among themselves rather than by the returns on any single deployment. Microsoft cannot afford to be the hyperscaler that did not build enough capacity, whatever the payback looks like.

The sovereign layer differs. Saudi Arabia's HUMAIN, the UAE's G42, Japan's national programme, Korea's build-out with SK Telecom, NAVER and Brookfield, and a European programme of 35 announced supercomputers all share what commercial buyers lack: spending for strategic reasons that survive a poor financial return. That is patient demand, and also the hardest layer to underwrite — political, lumpy and denominated in export-control risk.

10 — The financing turnThe seller became the financier, and that changes what demand means

The most consequential thing NVIDIA did in the third quarter of 2026 was not a product launch. On 12 August it announced partnerships with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR to establish independent financing platforms intended to mobilise more than $500 billion of third-party capital for AI infrastructure, so an operator can borrow against the equipment rather than fund it from equity.

Read generously, this is the standard playbook of an infrastructure supplier; read strictly, there are three problems. Collateral quality: lenders typically model the useful life of advanced accelerators at three to four years and NVIDIA argues it is longer, so if the lenders are right the loan-to-value falls as generations turn over, shrinking the platform exactly when it is needed. Circularity: NVIDIA holds equity stakes in some of the companies buying its products and now participates in the structures that lend them money, so a decline in compute demand would be amplified through the capital structure rather than absorbed by it. Price discovery: nobody yet knows what a three-year-old Rubin rack is worth at auction, because none has been auctioned.

The reported $40 billion debt package behind one customer's order for roughly 360,000 Rubin-generation chips is the clearest test case. The demand is real; the question is whether the marginal customer can pay out of cash flow or only out of credit, which determines what happens to orders in a risk-off quarter.

11 — China at zeroA structural loss, not a cyclical one

China contributed roughly $17 billion a year to Data Center revenue at its peak. In the second quarter of fiscal 2027 it contributed nothing, and management has excluded China Data Center compute revenue from guidance for five consecutive quarters. The company wrote down $4.5 billion of H20 inventory and has not increased the charge.

The market has largely stopped discounting this, on the reasonable grounds that demand elsewhere has filled the gap. We think it should be marked as permanent rather than temporary. The substitution is going the wrong way: Huawei's Ascend accelerators have reportedly passed NVIDIA in installed base inside China, and industries that scale behind a trade barrier rarely give the market back when the barrier lifts. The enforcement perimeter is widening: in October 2026 a contractor linked to a server manufacturer pleaded guilty in a $2.5 billion case involving AI servers routed to Chinese end users. NVIDIA has not been accused of wrongdoing, but a fabless designer cannot fully control where its product ends up. And the licences granted are not producing revenue: approval for H200 sales to ByteDance, Alibaba and Tencent has not translated into material shipments. We assign the option no value.

12 — The second sourceThe threat is not AMD. It is the customer

NVIDIA holds roughly 80% of data-centre AI accelerator revenue in 2026; forecasts put that between 65% and 70% by 2028. The interesting question is where the lost share goes.

AMD's Instinct line is gaining from a small base and has a credible roadmap. But competing on a merchant accelerator means competing on the die, the interconnect, the compiler and the software ecosystem at once, against an incumbent shipping a new architecture every year. We expect AMD to take share and to be structurally less profitable doing it.

The real erosion is inside the customer. Google's TPU is a fully controlled stack and the only hyperscaler silicon that has demonstrably displaced NVIDIA at scale; Trainium and MTIA depend on the fabric around them, which limits how far substitution can go.

13 — The power constraintThe ceiling moved from silicon to electricity

For three years the constraint on the AI build-out was the accelerator. It is now the interconnection queue. United States data centres consumed roughly 176 terawatt-hours in 2023, about 4.4% of national electricity; published projections put the 2028 figure between 325 and 580 terawatt-hours, or 6.7% to 12% of the grid. NVIDIA's and Broadcom's order books do not depend on power being available, but the shortage governs how quickly orders convert into revenue.

That distinction is why we treat power as a demand-quality issue rather than a demand-level one. NVIDIA can book the order whether or not the buyer has the megawatts; what it cannot do is ship into a site that has not been energised. Every quarter a rack waits for a substation is a quarter in which the customer's return on that capital is negative and its appetite for the next order lower.

NVIDIA's response has been to buy the constraint: the SB Energy partnership at the PORTS-Pike campus in Ohio secures land, power and shell capacity rather than a customer relationship.

14 — ValuationNineteen times fiscal 2028, and a distribution that is not kind

At $229.28 the shares trade on 29 times trailing GAAP earnings, 24 times our fiscal 2027 non-GAAP estimate of $9.55, and 19 times our fiscal 2028 estimate of $11.92, against roughly 22 times for the S&P 500. A business growing revenue at 106% is not expensive on that arithmetic, and the rating below is not a valuation objection.

We build the model from the revenue line down: the $108.0 billion third-quarter guide, a fourth quarter modelled at $118.8 billion, and fiscal 2027 revenue of $404.6 billion. Three fiscal 2028 outcomes follow.

The base case assumes $540 billion of revenue, a 33% increase, with non-GAAP net margin compressing from roughly 57% to 53% as systems content dilutes gross margin and the tax rate normalises. That gives $286 billion of non-GAAP net income and $11.92 per share, which at 21 times yields $250. The bull case assumes $625 billion of revenue, a 55% net margin and a 24 times exit multiple, yielding $350; it requires Vera Rubin to ramp on schedule, the financing platform to work, and power to arrive on time. The bear case assumes $430 billion of revenue — a 6% increase, not a decline — a 47% margin as price competition appears, and 16 times, yielding $135. It requires nothing to go wrong with the product; only that the marginal buyer's credit tightens or power delivery slips a year.

Weighted at 25%, 50% and 25%, the value is $246, roughly 7% above the current price. We set the twelve-month target at $250, the base case rounded, and rate the shares Neutral.

15 — What the price requiresNot a miracle. Just the base case, and no multiple compression

Reverse-engineering the price is more informative than forecasting it. At 20 times, $229.28 requires $278 billion of fiscal 2028 non-GAAP net income. At our assumed 53% net margin that is $506 billion of revenue — 25% above the fiscal 2027 estimate and just below our base case. At 18 times the requirement rises to $309 billion of net income and $583 billion of revenue, above our base case and inside the bull case.

The market is not pricing a miracle; it is pricing the base case and a modestly above-market multiple. What the price does not tolerate is multiple compression. If the market decides that a company whose growth is decelerating from 106% to 30% deserves a market multiple, the same base-case earnings support $205 rather than $250, and nothing about the operating business would have to go wrong.

16 — The bull caseAn installed base with no second source

The strongest version of the long argument is not about growth. It is about switching costs that compound. Every rack NVIDIA ships takes the customer further into CUDA, into NVLink, into the AI Enterprise stack and into a training pipeline tuned for the hardware. The cost of moving that workload rises with every generation, and the generation cycle has shortened to a year. On this reading 80% share is not a peak to be eroded but an installed base still being laid.

The second pillar is price per unit of work. Each generation is claimed to deliver a five-to-ten-times improvement in performance per dollar. If that is even roughly right, the marginal cost of inference falls faster than any competing architecture, and a customer choosing on total cost of ownership keeps choosing NVIDIA.

The third is that the demand is not yet discretionary. Frontier training is a race among a handful of laboratories, none of which can afford to pause, and sovereign programmes are strategic rather than financial. Both will keep buying through a downturn that would stop a purely commercial buyer.

17 — The bear caseThe demand is real; the funding of it is not obviously permanent

The bear case does not require NVIDIA to lose a technology race, only that the way the current demand is being paid for turns out to be less durable than the demand itself.

A laboratory commits to buying compute, financed partly by equity raised at a private valuation, partly by debt arranged against the equipment, and partly by the supplier's own balance sheet. Each link is reasonable. Collectively they mean the revenue NVIDIA books is partly a function of capital NVIDIA and its investors have supplied, and when a market is financed this way the loop runs both ways: rising valuations justify more spending, and a fall removes the funding for the spending that supported them.

There is already a small example. On 8 October 2026, NVIDIA, Oracle and CoreWeave fell close to 3% in one session after a report that a major customer's annualised revenue was approximately $20 billion below the figure previously cited. Nothing about the product changed that day; what changed was the market's estimate of its customers' ability to pay.

The second leg is arithmetic. Sequential revenue growth has gone from 22% to 12% guided in four quarters, and year-over-year growth will mechanically fall as comparisons roll into quarters that already carried triple-digit increases. A stock re-rated as a compounder on its rate of change does not need to disappoint to fall; it needs only to normalise. On base-case earnings, a market multiple gives $205.

The third leg is the one the company cannot control. Power delivery and grid capacity are governed by utilities and regulators, not by NVIDIA's execution. A twelve-month slip in energising the sites behind the order book would not reduce the backlog; it would convert it into inventory and force customers to re-underwrite their payback periods.

18 — Risk matrixTwelve risks, ranked by probability and impact

Probability and impact are scored 1 to 5, where 5 is most likely and most damaging.

19 — CatalystsWhat would change our mind, and when

20 — ConclusionOwn the business at a different price

NVIDIA is the most important company in the compute economy and, on the operating evidence, one of the best businesses ever assembled. Revenue has gone from $44.1 billion a quarter to a guided $108.0 billion in seven quarters, gross margin has expanded 14 points to 75%, and a single quarter now generates $48.6 billion of free cash flow. The architecture cycle has shortened to a year.

None of that is what we are being asked to underwrite at $229.28. At that price the market has already accepted the base case: $540 billion of fiscal 2028 revenue, a 53% non-GAAP net margin, and a multiple above the market's. What remains is the shape of the distribution. The bull case pays 53% and requires uninterrupted execution on four fronts at once. The bear case costs 41% and requires only that the marginal buyer's access to credit tightens, or that the multiple normalises.

Our target is $250, about 9% above the current price, and our rating is Neutral. We are not arguing that the shares are overvalued on the earnings the company is likely to produce; we are arguing that the price already contains them, and that the $18.6 billion of first-half profit which did not come from operations should not be capitalised.

21 — AppendixFinancial summary, fiscal 2026 and fiscal 2027

Figures as reported. Items marked — were not separately disclosed. Q2 FY2027 non-GAAP net income is derived from non-GAAP diluted EPS and share count. Fiscal years end in late January.

22 — SourcesPrimary and secondary references

Built from primary disclosure wherever possible; every chart figure traces to a document below.

• NVIDIA, Q2 FY2027 results, press release, 26 Aug 2026.
• NVIDIA, Q1 FY2027 results, press release and CFO commentary, 20 May 2026.
• NVIDIA, Q4 and full-year FY2026 results, press release, 25 Feb 2026.
• NVIDIA, Form 8-K, $150 billion repurchase authorisation, 28 Sep 2026.
• NVIDIA, agreement to acquire Hugging Face, 3 September 2026.
• NVIDIA, AI compute financing platforms with Apollo, BlackRock, Blackstone, Brookfield, Goldman and KKR, 12 Aug 2026.
• Lawrence Berkeley National Laboratory, 2024 United States Data Center Energy Usage Report.
• Reported coverage of the Super Micro export case and the October 2026 plea.
• Reported coverage of the SpaceX debt package for Rubin accelerators, October 2026.
• Reported coverage of the 8 October 2026 session decline after revised customer revenue disclosure.
• Third-party market-share compilations, 2026; market data and consensus as of 12 October 2026.

23 — DisclosureBasis of preparation and conflicts

Read next:  ·  Microsoft — The Franchise and the Furnace →  ·  Amazon — The Cost of Capacity →