Nvidia

Where we land

Nvidia's processors for artificial-intelligence data centres have turned a large chip company into one of the most profitable businesses anywhere. Its yearly operating profit, what is left after paying for the products and for running the company, rose from 96.0bn to 197.6bn in a single year. Yet the shares trade at 30 times earnings, the cheapest of the chip companies we compare them with. The market is pricing this boom as if it will fade soon. That could prove right, and there are serious reasons to think so, but the gap between what the price expects and what the business has recently delivered is unusually wide.

We award Nvidia a silver medal. Analysts rank it among the very best; what keeps it from gold is that its profits are not backed by incoming cash as fully as we look for.

What the business is

Nvidia reports all of its sales as a single line, Multimedia Semiconductors. Inside it sit the processors and complete computing systems that train and run AI models in data centres, alongside gaming chips and computing platforms for robots and cars. The company does not split out each contribution, though from what it sells, data-centre AI makes up most of it.

What stands out is how much of each sale the company keeps. Its gross margin, the share of each sale left after the direct cost of making the product, is 75%. After the costs of running the whole company, the operating margin is 65%. These are figures more associated with software than hardware, and they show customers paying for something they cannot easily get elsewhere.

That pricing power is not new, only stronger. Over five years the gross margin has had a middle value of 70%, typically moving about 6 points either side. Today's reading sits towards the top of that band rather than outside it: a business running at full stretch, not one whose economics have changed character.

Why now

Demand for AI computing has grown so strong that some buyers are now borrowing to pay for it. Multiple reports on 6 and 7 October say SpaceX is in talks with banks and investors, led by the investment firm Apollo, to raise $40 billion specifically to buy Nvidia chips for Elon Musk's AI expansion. Commentators have begun to describe money, not appetite, as the limit on how many chips get bought.

The shares have reached record highs as the company approaches a value of around $6tn. The climb, though, is less forceful than a year ago: the stock has been recovering ground rather than breaking away. The market is enthusiastic about the business but not carried away by it.

The tension at the heart of the case is that Nvidia's margins are far richer than its chip peers', yet it trades at a far lower multiple of earnings. Either the market knows something unsettling about how long these earnings will last, or it is being unusually cautious about a business that keeps outrunning expectations.

The case for

The core of the argument is the pace at which profit has grown. Twelve-month operating profit stood at 10.8bn in July 2023 and reached 197.6bn by July 2026. Sales grew 83.4% in the latest year, on top of a three-year compound growth rate of 110%: not one lucky year but a sustained run.

Those who follow the company most closely have been moving in one direction. Over the past 45 days, earnings forecasts were raised 83 times and cut only twice. The consensus expects earnings per share to rise from 9.28 to 15.77, roughly 70%. Revisions this consistent usually mean the business keeps reporting better than its modellers assumed.

The SpaceX financing shows where future sales could come from: a single buyer arranging that much money for one supplier's products points to a large additional order, and to customers willing to stretch their balance sheets to secure supply. With the shares at 30 times earnings, investors are not paying for perfection, which is the opening this case relies on.

Where the growth could come from

The engine today is data-centre AI: the processors and full systems that large technology companies and AI builders buy to train and run their models. What has changed is how that demand is funded. If debt markets become a regular way to pay for chips, demand is no longer capped by what customers earn from their own operations, which supports both the expected 70% earnings growth and the stream of upward revisions.

The catch is the cost of borrowing in the United States. The Federal Reserve raised its policy rate to 4.00% in September, reversing a cut the previous December, and the ten-year US government bond yield, a benchmark for long-term borrowing costs, has risen to 5.31%. Money borrowed to buy chips is getting dearer just as more buyers reach for it, and purchases funded by debt are more fragile than those funded from cash flow.

A further avenue is inference, the work of running AI models day to day once trained. Nvidia has struck a $20 billion deal with Groq, a company known for chips that run trained models quickly. This is where cloud companies' own custom chips compete most directly with Nvidia, and as AI spending shifts from training towards running models, holding ground here will decide whether growth and margins last. Its contribution to sales is not separated out. Success would show in Nvidia launching products built on Groq technology and winning named inference customers.

Further out lies an option in robots and self-driving cars, where Nvidia sells computing platforms and software to carry its AI lead into machines that act in the physical world. This is most likely a small part of sales today, as it is not reported separately. If it grew into a meaningful business, it would give Nvidia a second growth path and make earnings less dependent on what a handful of data-centre buyers can spend or borrow. Production programmes from carmakers or robot makers built on Nvidia platforms, and the company eventually reporting this as its own line, would confirm it.

The case against

The low earnings multiple is the bull case's most persuasive point, and also where the case against begins. Nvidia's gross margin of 75% compares with 50% at AMD and 42% at MediaTek; its 65% operating margin dwarfs AMD's 16% and MediaTek's 15%. Nvidia trades at 30 times earnings, AMD at 167 and MediaTek at 81. Nvidia looks cheap on earnings precisely because its margins are exceptional, and exceptional margins are what competition and customer pushback tend to erode.

On measures that do not depend on margins, the shares are among the most expensive in the market. Enterprise value, the value of the whole company including debt and net of cash, stands at 18.8 times sales. The share price is 25.2 times book value, the net assets on the balance sheet. On both, Nvidia sits near the bottom of the companies we cover for value. If margins slipped even towards peer levels, the earnings multiple would no longer look modest and that premium would be fully exposed.

The quality of demand deserves scrutiny too. Part of the order flow now rests on customers' access to credit as borrowing costs rise. While most coverage celebrates new highs, at least one analyst has downgraded the stock, arguing that rising inventory signals weakening demand, and others have published long lists of concerns. Neither point proves the boom is ending, but together they describe a situation in which reported sales could flatter how much computing is genuinely being used.

There is also a governance question. Shareholders are suing over the Groq transaction, one of the company's largest uses of capital, amid wider criticism of deals structured mainly to acquire a company's people. For a business generating profit on this scale, how that money is deployed matters a great deal.

What the price assumes

One way to judge a share price is to ask what growth it needs to be fair. Suppose the price stayed flat for five years and the shares then traded at 18.5 times earnings, the middle value across every profitable listed company we cover. Nvidia's earnings would need to compound at about 10% a year, putting earnings per share at 12.94 in five years.

Over the past three years, operating profit has grown at about 102% a year. If that trend simply continued for five more years, the company's own past carried forward rather than a forecast, earnings per share would reach 266.01, or 2055% of what the price requires. Nobody should expect growth of that kind to persist, but the comparison shows how much slowing the price has already absorbed.

The analysts' view makes the point more modestly and more tellingly. Their forecast for next year, 15.77, is already above the 12.94 the price needs in five years' time: if it were met and earnings then went nowhere for four years, the requirement would still be covered. The shares have risen 29% over the past year and 31% over the past six months, yet the bar they set remains low relative to the earnings now expected.

The other side of the medal

The risk that matters most is that today's demand is partly paid for with borrowed money while supply may be running ahead of what customers actually use. If buyers pause to absorb what they have already bought, the hit would land on nearly all of Nvidia's sales, because AI-related products sit within a single line that is all of its revenue. With sales up 83% in a year, growth could slow sharply and the generous forecasts would be cut rather than raised. There is no other large business to cushion the blow.

This risk explains the price better than any other. It is why a company with margins this high can trade at a lower earnings multiple than its peers while looking dearer than most companies we rate on assets and sales. Investors are accepting today's earnings but declining to pay much for their durability.

What to watch

The most useful evidence is the relationship between inventory and sales. If reported inventory and supply commitments grow more slowly than sales, demand is being met rather than stockpiled. If they climb faster than revenue, chips are piling up ahead of real use, which is the heart of the downgrade case.

Next is funding: whether the SpaceX financing closes at its full size and whether other AI builders announce similar debt-backed purchases, or instead struggle to fund them as borrowing costs rise. For Groq, the test is the lawsuit being dismissed or settled without unwinding the deal.

Finally, our reading would be wrong if the gross margin broke below its long-run band of around 70% while analysts started cutting next year's earnings forecasts. Any one of these signals would suggest the market's caution was well placed; all of them together would mean the boom the price discounts was indeed fading.

The numbers

The same industry, for comparison

Research for information, not personal investment advice or a recommendation to buy, sell or hold any security: see Important information. Data as at 7 October 2026.