ReportsMonitoringPortfolioMethodologyPricingBlogAboutStart for free
Back to blog

Analysis

How Does Nvidia Make Money? By the Numbers, Not the Hype

How does Nvidia make money? Not gaming cards anymore. Segment revenue, margin, customer concentration and the price of export controls, straight from the 10-K.

9 min read
See a sample report

This is a method, not a recommendation. Nothing here, or anywhere else on Taufolio, is investment advice. Treat every example as a starting point for your own research.

How does Nvidia make money? Nine dollars out of ten, the answer is data center chips that train and run AI models. (Most of the tenth is still gaming.) In the fiscal year ended January 25, 2026, the Data Center segment brought in $193.7bn, which is 89.7% of the company's total revenue. Gaming, the product most people still associate with Nvidia, is down to 7.4%.

This piece breaks Nvidia's business model down into numbers from its own annual report, not an investor slide deck. I use the same method on every company; the version built on Costco's one-page business model shows it on a completely different money engine.

The money engine: data centers, not graphics cards

Nvidia's total revenue in fiscal 2026 was $215.9bn, up from $130.5bn a year earlier, a 65.5% increase. That growth is not spread evenly across the business. The Data Center segment grew 68.2% year over year, from $115.2bn to $193.7bn, and on its own accounts for most of the company's entire revenue increase.

Gaming, by comparison, brought in $16.0bn, up from $11.4bn, a 41.3% increase. Fast growth, on a shrinking slice of the pie: from 8.7% to 7.4% of revenue in one year. Professional Visualization added $3.2bn (up 70%) and Automotive $2.3bn (up 39%); together, the two still make up under 3% of the total.

segment FY2026 FY2025 change
Data Center $193.7bn (89.7%) $115.2bn (88.3%) +68.2%
Gaming $16.0bn (7.4%) $11.4bn (8.7%) +41.3%
Professional Visualization $3.2bn $1.9bn +70%
Automotive $2.3bn $1.7bn +39%

Read nominally, every segment is growing. Read as a share of the pie, only one really is.

The mechanism is easy to name and hard to copy: companies training large language models need thousands of specialized chips at once, and Nvidia has spent years building a full platform around them, not a single product. Chips, the CUDA software layer, and the networking that links thousands of chips into one machine. A competitor selling just the chip, without that platform, is selling less than half the offer.

A decade ago the same revenue split ran the other way: in fiscal 2016 gaming brought in $2.8bn of $5.0bn and the data center business $339m. Gaming pulled the company, and server rooms were the side business. The swap took a few years and one event in the background, the boom in large language models, but the 10-K numbers show it's no longer a passing trend. It's the company's new center of gravity.

Where the demand comes from, and what could stop it

Data center chip demand is driven by companies building their own AI infrastructure: the big cloud providers, model labs, increasingly governments and telecoms too. That's exactly the type of customer that explains the concentration numbers below. Only a handful of companies on earth can place multi-billion-dollar annual orders, so when two of them account for more than a third of Nvidia's revenue, that isn't an anomaly. It's what you'd expect from who actually buys at that scale.

What could slow this engine down? A fair counterargument: some of Nvidia's biggest customers are designing their own specialized chips, tuned to their own models, specifically to reduce dependence on one supplier. That's a real risk, and Nvidia's 10-K doesn't publish a figure showing how fast in-house silicon is eating into market share, so I won't invent one where the company hasn't. This is exactly the kind of question where the honest answer is "I don't know, and here's why it matters": the more of Nvidia's biggest customers design their own chips, the less even the best software platform counts for.

Margin that fell while revenue kept climbing

Here's where it gets interesting. Nvidia's gross margin fell from 75.0% to 71.1% between fiscal 2025 and fiscal 2026, even as revenue grew by two-thirds. That's not a demand signal. The company points to two things at once: the move from selling Hopper HGX systems to full-scale Blackwell data center solutions, and a one-time $4.5bn charge for H20 inventory and purchase obligations.

It's worth reading that number against the year before it, too. Margin actually rose from 72.7% in fiscal 2024 to 75.0% in fiscal 2025, before falling in fiscal 2026. So the three-year shape isn't a straight decline, it's a peak followed by a drop, and only the next filing will say whether fiscal 2026 was the bottom of a temporary dip or the start of a new, lower normal for this architecture generation.

That charge has a specific cause, and it's political, not technical. In April 2025 the US government required a license to export H20 chips to China, citing national-security technology controls. Demand for H20 in China collapsed, and Nvidia wrote down $4.5bn of inventory and purchase commitments in the first quarter of fiscal 2026. Licenses were granted in August 2025, with a catch: Nvidia recognized about $60m of revenue from those sales, a fraction of the charge itself. On top of that, US officials have said they expect the government to receive 15% or more of the revenue from licensed sales. A separate H200 license followed in February 2026, with its own catch: every H200 must pass an inspection in the United States before it ships, and that import carries a 25% tariff.

Two customers, more than a third of revenue

The bigger the Data Center segment gets, the fewer customers actually drive it. In fiscal 2026, one direct Nvidia customer accounted for 22% of total revenue, another for 14%, 36% combined from two customers, both attributed mainly to the Compute & Networking segment, the reporting segment that holds the data center business. A year earlier it was 12% from one customer and 11% each from two others.

Concentration is rising alongside the segment driving it, not shrinking with scale.

It's a number worth tracking quarter over quarter, because it says more about risk than the headline growth rate does: the fewer customers decide the outcome, the more that outcome depends on a handful of purchasing decisions, not a broad market.

What protects this engine from being copied

Nvidia's moat isn't really in the chip. It's in the platform built around it. CUDA, the software layer customers have built their models on for years, means switching to a competing chip requires rewriting code, not just swapping a card. It's the same kind of switching cost I've written about in companies that build an ecosystem around their product, just applied to AI infrastructure instead of office software.

Competitors have a real shot at the chip itself. Building an entire platform, network and developer community around a rival standard in a few years is a much harder task. That's the evidence I'd look for in future filings: not whether someone ships a faster chip, but whether anyone actually convinces Nvidia's customers to rewrite their code.

Selling a rival chip without that platform is a bit like opening a bakery next to a place that also rents out the ovens, sells the only compatible flour, and trained half the town's bakers for free. (Somewhere, a hardware engineer just nodded very slowly.) The oven itself might even be better. Getting anyone to switch kitchens is the actual fight.

My own view is that the switching cost, not the raw chip advantage, is the more durable part of this moat today. A chip loses its edge faster than an entire developer ecosystem built over a decade does.

Real competitors do exist on the chip side, from AMD's own accelerator line to the custom silicon that cloud providers design in-house for their own data centers. Naming them here is illustrative, not a forecast of who wins: the point isn't whether a faster chip ships next year, it's whether it ships with a software ecosystem deep enough that a customer would actually migrate a production model onto it.

Three questions for the next filing

Three questions for the next annual report, each with a threshold, not a hunch.

  1. Does gross margin move back toward 75% as Blackwell production matures, or does it settle permanently lower as the new normal for this architecture?
  2. Does the concentration from the top two customers rise past 36%, or start spreading across more buyers as more companies build their own AI data centers?
  3. Does China export policy shift again, and by how much of Data Center revenue?

None of these questions has an answer today that comes with a target price. They have answers that come with a number and the date of the next 10-K, which is exactly what an investment thesis should be built from, at Nvidia or anywhere else.

At a company where the Data Center segment moves this fast quarter to quarter, an annual filing alone is too rare a check-in. Every Nvidia earnings call adds one more point to the same line: Data Center share, margin, customer concentration. One reading proves nothing on its own, but four in a row show whether the trend in this piece is accelerating or reversing. That's also where management's own language earns a second look: a strong quarter described in confident, specific numbers reads differently than the same quarter wrapped in "robust momentum" and "exciting demand signals."

What you just read is a slice of what a Full report does systematically for any company: break revenue into segments, check margin against prior quarters, and list customer concentration wherever a company discloses it. You can see the full version in Taufolio's research.

Frequently asked questions

Mostly from the Data Center segment, the chips that train and run AI models. In fiscal 2026 it brought in $193.7bn, which is 89.7% of the company's total revenue. Gaming, the product most people associate with Nvidia, is down to 7.4%.
89.7% in fiscal 2026, up from 88.3% a year earlier. The segment grew 68.2% year over year to $193.7bn, per the company's 10-K for the year ended January 25, 2026.
Gross margin dropped from 75.0% to 71.1% for two reasons at once, per the 10-K: the move from selling Hopper HGX systems to full-scale Blackwell data center solutions, and a one-time $4.5bn charge for H20 inventory and purchase obligations after the US government restricted exports of that chip to China.
In fiscal 2026 one direct customer accounted for 22% of revenue and another for 14%, 36% combined from just two customers. A year earlier it was 12% from one customer and 11% each from two others - concentration is rising alongside the Data Center segment itself.
A $4.5bn one-time charge in the first quarter of fiscal 2026, after the US government required a license to export H20 chips. Once licenses were granted in August 2025, the company recognized about $60m of H20 revenue - a fraction of the charge.
Not the chip itself, but the platform around it: the CUDA software layer that customers have built their models on for years, plus the networking that links thousands of chips into one computer. Switching to a competing chip means rewriting code, not just swapping a card, which is a far more expensive decision than the price of the chip.
  • research
  • fundamentals
  • moat
Share:XLinkedIn

Posts are produced with AI tools and go through editorial review by the Taufolio team before publishing. How we write and correct posts

Cookies and analytics

Essential cookies keep you signed in. Analytics runs only if you agree.