Bufetico Join the waitlist
Cheap or expensiveCase study

The P/E kept you away, and you never knew what it meant: Nvidia's own teams can't get enough Nvidia chips

Bryan Catanzaro runs applied AI research at Nvidia, and his teams can't get enough Nvidia chips. What does the P/E ratio actually mean when the profit under it grows that fast? I divide one share price by two different years of profit, with Nvidia's own numbers.
26 September 2026 · · 5 min read

A big P/E made you nervous enough to stay away, and you never knew what that number was telling you. The man who helped point Nvidia toward artificial intelligence now has to ask for Nvidia chips for his own teams, and he doesn’t always get them.

Bryan Catanzaro leads applied deep learning research at Nvidia. His teams use AI for graphics, speech recognition and simulation. Back in the early to mid 2010s, he was one of the first people to notice researchers buying Nvidia’s chips to train AI models. Fortune says that signal helped push Jensen Huang, the chief executive, to double down on AI.

At a conference in San Francisco in the spring of 2026, Catanzaro told Fortune’s Sharon Goldman that the main thing his people complain about is that they don’t have enough computing power. They want more GPUs, the chips that train and run AI models, and Fortune put them at over $30,000 apiece. He said one of his main jobs now is trying to get more of them for his teams. And when he asks the boss, the answer can be short: “Jensen will say, ‘I’m sorry, Bryan, but those are sold.’”

Catanzaro also leads the team that builds Nemotron, a family of AI models anyone can download for free. Nvidia makes them so developers keep building on its hardware and software. He said that work had only started getting more attention inside Nvidia in the six months before the interview.

One price, two years of profit

Nvidia’s second-quarter report covers the period up to July 26, 2026. In its 10-Q, the quarterly report it files with the SEC, net income for the first six months of the fiscal year was $118,010 million. For the whole fiscal year that ended in January 2026, it was $120,067 million. So in six months, Nvidia made almost as much as in the whole year before.

A P/E is the price of one share divided by one year of profit for each share. It tells you what the market pays today for each dollar of profit from that year. The price already has the market’s guess about the years after that built in, but the ratio doesn’t tell you what that guess is.

On September 21, an Nvidia officer sold shares for about $223 each, under a plan set up ahead of time. That comes from a Form 4 filed with the SEC two days later. Nvidia’s annual report for the year to January, the 10-K, gives diluted earnings per share of $4.90. That’s the profit spread over every share, counting the ones promised to employees. $223 divided by $4.90 is about 46. Now do it with the twelve months to July. Start with the $4.90 for the year to January, drop the $1.84 from that year’s first six months, and add the $4.85 from the first six months of the next one. That gives you $7.91. And $223 divided by $7.91 is about 28.

It’s the same share at the same price, and the ratio goes from 46 to 28 just because you divided by a different twelve months. And part of that newer E, the profit, came from outside the chip business. In the six months, Nvidia’s income before tax was $141,410 million, and $24,140 million of it was other income, mostly paper gains on stakes it holds in other companies.

Nvidia’s diluted earnings per share were $1.19 for the fiscal year that ended in January 2024. Two years later they were $4.90. When profit moves that fast, the P/E you read today is divided by a year that’s already over.

When a company’s profit barely changes from year to year, the year you divide by looks a lot like the next one, so the ratio holds still. Nvidia’s profit doesn’t behave like that. Fortune says that across the industry, getting those chips is still a bottleneck.

Inside Nvidia, the shortage shapes Catanzaro’s Nemotron work. He told Fortune it’s part of why those models are built to get by on fewer chips. And he said that when something gets more efficient, people find all sorts of new ways to use it.

When I measure how fast a company grows, I take the median of its year-to-year changes, the one in the middle once you line them up. That way one wild year ends up at the edge of the list and doesn’t set the pace by itself. I also put a cap on how much growth I’ll count.

Where this breaks

My twelve-month figure stitches two fiscal years together by hand, and each per-share number rests on a different average count of shares. Nvidia doesn’t report a twelve-month number like that. So this one is mine, and it’s only close.

It’s also one day’s price, from one insider’s sales. A few days earlier, on September 18, a director sold at about $220. With that price, every ratio here would come out a little lower.

I wrote about what a P/E ratio says on a business whose profit swings with its industry, and about what earnings per share is made of. The course has a lesson on how to read an income statement.

Who Catanzaro’s chips were sold to

Further down the 10-Q, well below the profit, there’s a note that shows how few buyers there are, without naming them. In the quarter through July 26, 2026, one direct customer, a company that buys straight from Nvidia, was 16 percent of total revenue. Nvidia also estimates that one AI research company, which it doesn’t name either, brought in a meaningful amount of revenue by buying cloud services from Nvidia’s customers.

At the end of July 2026, five direct customers accounted for 22, 14, 13, 11 and 10 percent of the money customers owed Nvidia. So five companies owed it 70 percent of that money. Nvidia counts a sale, and the profit on it, when the products pass to the customer. The 10-Q says payment is generally due shortly after delivery. But Nvidia has given some buyers from 90 days up to a year to pay. They’re investment-grade buyers, the ones with strong credit, building large data centers. Nvidia may keep doing that, and the profit on those sales is already counted in the E.

With Nvidia, I’d work out the P/E myself after each quarterly report. I’d divide the share price by the last twelve months of earnings per share, stitched together the way I did above. Then I’d go to the note on customers, to see how much of Nvidia’s revenue comes from a single direct customer and how much of what customers owe Nvidia sits with five of them.

To see whether Nvidia looks cheap or expensive today, search for it in Bufetico.

The P/E that kept you away from Nvidia changes a lot depending on which twelve months of profit you divide by.

Key factsThe numbers and where they come from · 26 September 2026
  • In the first six months of its fiscal year, through July 26, 2026, Nvidia’s net income was $118,010 million, against $120,067 million for the whole fiscal year that ended in January 2026 (Nvidia’s 10-Q).
  • A P/E is the price of one share over one year of profit per share. Nvidia’s 10-K gives diluted earnings per share of $4.90 for the fiscal year that ended in January 2026, up from $1.19 two years earlier.
  • A Form 4 filed with the SEC shows a price of about $223 a share on September 21, 2026. That puts Nvidia’s P/E at about 46 on the year to January and about 28 on the twelve months to July.
  • One direct customer was 16 percent of Nvidia’s total revenue in the quarter through July 26, 2026, the 10-Q shows. At the end of July, five direct customers accounted for 70 percent of what customers owed it.

Bufetico puts every company on more than 30 exchanges through the same six layers. Prices update after every market close, and company financials once a month. What passes is decided by thresholds you set yourself. See it.

← All articles

Sold too early, or held on too long?

Join the waitlist and I'll email you when Bufetico opens, plus a new article or tool now and then.

No more than one email a week. You can unsubscribe anytime, and the Privacy Policy explains how we handle your name and email.

Can't tell a cheap stock from a dying one?

The Bufetico Filter checks more than 30,000 companies against six layers and shows you only the ones that pass. You decide how strict each layer is. Bufetico opens soon.

Join the waitlist »