NVIDIA Corporation NVDA

Key data

The business

NVIDIA sells the engines of AI computing: data-center GPUs like Blackwell, the NVLink and networking gear that lashes thousands of them into one machine, and the CUDA software programmers have built on for fifteen years. That software is the moat. The chips are fast, but the real switching cost is the millions of lines of code written to run only on them. The buyers are cloud service providers, AI model makers, and a growing list of governments building "sovereign" AI. Data center is now essentially the whole company, and gaming, the original business, is a rounding error. The thing to hold in your head: five direct customers each buy more than a tenth of what NVIDIA ships, so this is a business with staggering scale and a dangerously short list of people who can say no.

The numbers

The trajectory is the whole story.

Revenue near-doubled year over year and gross margin recovered to 75.0% from 72.4%. One nuance: Q1 FY27 net income of $58.3B actually exceeded operating income of $53.5B, flattered by non-operating gains, so the modest 3% step to Q2 understates the real momentum, where operating income rose 19% in the quarter. Non-GAAP results came in above the estimate in each of the last four quarters.

Half of FY27 already matches nearly all of FY26's net income. Consensus sees revenue reaching $904B by FY29 (36 analysts) and crossing $1T in FY30. On the FY29 estimate of $21.16 in EPS, today's price is under 11 times earnings, against roughly 30 times trailing. That gap between the two multiples is the bet itself: the market only half-believes its own consensus, and what settles it is whether sequential data-center growth holds double digits through FY27.

The newest and least-discussed line is on the balance sheet.

In one quarter NVIDIA raised about $25B of debt, took $2.8B of customer prepayments, tripled raw materials, and separately guaranteed 4.25 GW of data-center power and transmission. It also extends payment terms of 90 days to a year on large builds. It is starting to finance its own demand.

Management

Insiders only sell. Zero purchases in twelve months against 55 sales worth $992M, and director Mark Stevens alone sold $329M across three tranches, the latest on Aug 31, plan status not disclosed. At a company compounding this fast that is unremarkable, but there is no insider buying signal to lean on. Capital return is real: $40.1B repurchased in FY26 after $33.7B in FY25, though buying at nearly 30 times earnings is a very different act than the buybacks made when the stock was cheap. Return on equity runs above 100% and return on invested capital near 60%, so the reinvestment case is intact. The tell to watch is not the selling. It is whether buybacks keep accelerating into a rising multiple.

How it fails or surprises you

Customer concentration cracks. Five direct customers each take more than 10% of sales, and the biggest, the hyperscalers, are all building their own AI silicon (Google's TPU, Amazon's Trainium, Microsoft's Maia). If one shifts volume in-house or trims 2027 capex, the order book bends fast. First sign: sequential data-center growth slipping to single digits, or a named customer cutting its capex guide.

NVIDIA is financing its own demand. This is the fact the clean-growth story explains least. A $25B debt raise, $2.8B of customer advances, inventory up 48% to $31.6B, and a guarantee on 4.25 GW of power all landed in one quarter. If AI returns disappoint, this circular funding unwinds violently. Watch inventory provisions, already $784M this quarter, and days-sales-outstanding, now 76 days.

Inference and sovereign demand overshoot (right tail). The market pays about 30 times trailing but only 11 times the FY29 revenue consensus of $904B. If inference workloads and government buildouts scale the installed base past those estimates, and robotics and automotive finally break out, the earnings land and today's price is cheap. The divergence flagged last week between data center and those newer lines has not appeared, and the Q3 print in November is the next test.

Closing thoughts

Both tails are fat, which is rare at a $5.6T market value. The market has quietly de-rated the stock to about 30 times trailing earnings, below its own 33-to-55 band of the last decade and well under the roughly 54 times its rivals carry, so it is pricing hypergrowth and doubting its durability at the same time. The November earnings call settles it: if data-center revenue grows double digits sequentially and no major customer cuts its capex guide, the hypergrowth is real and the de-rating was wrong. If concentration breaks or the self-funding unwinds, the loss is permanent and large, because a company this size cannot grow its way out quickly. If inference demand is as deep as the bulls claim, the FY29 numbers make today's price look modest. The left tail is capital impairment. The right tail is a re-rating on earnings that already sit in the run rate.

The bet is still that the giant cloud companies keep spending more every year on AI data centers, and keep buying NVIDIA's chips to fill them. It breaks the day two numbers move apart, sequential data-center growth stalling into single digits while a top customer's capex guide comes down. Watch those two together on the November call. If data-center revenue grows double digits again and no major buyer blinks, the bet is intact. The day either one turns, you are no longer making the same bet.