Nvidia's real moat was written in 2006, and it is software
Nvidia reported $215.9 billion of revenue in fiscal 2026, almost all of it from data centres. The chips get the headlines, but the advantage that rivals struggle to copy is CUDA, the programming platform Nvidia began giving developers two decades ago.

Key takeaways
- Nvidia's revenue for fiscal 2026 was a record $215.9 billion, up 65%, and fourth-quarter data centre revenue alone was $62.3 billion.
- The company launched CUDA in 2006, letting programmers use graphics chips for general computing long before there was a large market for it.
- Years of free tools, libraries and university adoption mean that most AI software is written for CUDA first, which raises the cost of switching to a rival chip.
- The moat is strong but has edges: customers are building their own chips, and export rules limit sales to China.
Nvidia's results for the year ended January 2026 read like a record of a company at the centre of a gold rush. Revenue was $215.9 billion, up 65% from the year before. In the fourth quarter alone the company booked $68.1 billion, and its data centre business brought in $62.3 billion of that. Gaming, the business Nvidia was built on, now makes up only a small slice of sales.
It is tempting to explain this with the chips. Nvidia's processors are fast, and demand for the computing power needed to train and run AI models has outrun supply. But fast chips can be matched. Several companies make capable accelerators. What has proved much harder to match is something Nvidia started giving away in 2006.
A bet made before the market existed
Nvidia was founded in 1993 by Jensen Huang, Chris Malachowsky and Curtis Priem to make graphics chips for PCs. Graphics processors are built to do many small calculations at the same time, which is what drawing millions of pixels requires. By the mid-2000s some researchers had noticed that this design was also good for scientific work, but using it meant disguising their maths as graphics code.
In 2006 Nvidia introduced CUDA, a platform that let programmers write ordinary code for its graphics chips. There was no large market for this at the time. The company was spending money on tools for a small group of scientists and engineers, and it kept doing so for years: compilers, debuggers, maths libraries, training material and support for universities.
The payoff came in 2012, when a team at the University of Toronto trained an image recognition network called AlexNet on two Nvidia graphics cards and won a major competition by a wide margin. Deep learning took off, and the researchers who built it were already writing CUDA code.
How the moat works
A moat is anything that makes it costly for a customer to leave. Nvidia's has three layers.
The first is code. Most popular AI frameworks were built and tuned on CUDA. A company that wants to move to another chip has to make sure its models, tools and custom code run just as well on the new hardware. That work takes engineers and time, and any performance gap shows up directly in cost.
The second is libraries. Over two decades Nvidia has built specialised software for the most common operations in AI and scientific computing. These libraries squeeze extra speed out of its hardware. A rival needs not only a good chip but equivalent software for each of those tasks.
The third is people. Universities have taught CUDA for years. Engineers hired into AI teams usually already know it. Choosing another platform means retraining staff or hiring for a rarer skill.
Each layer on its own could be crossed. Together they mean that a customer considering a switch has to win on price and performance by a large enough margin to pay for the move, and keep winning for long enough to make it worthwhile.
Where the edges are
No moat is permanent. Nvidia's largest customers, the big cloud companies, spend so much on its chips that building their own has become attractive. Several have designed in-house accelerators for their own workloads, and they control the software stacks that run on them. Open-source projects are also working to make AI code less dependent on any single vendor's platform.
Politics is another edge. US export rules restrict the sale of advanced AI chips to China, once a significant market. In its fiscal 2026 results Nvidia said its outlook for the first quarter of fiscal 2027 did not assume any data centre compute revenue from China.
Concentration cuts both ways too. A business that earns most of its revenue from a handful of very large customers is exposed if those customers slow their spending or succeed in replacing it.
What founders can take from it
The most useful lesson from Nvidia is about timing. The company invested in a developer platform years before it could show a return, and kept investing when the market was small. When demand arrived, the tools were already in thousands of hands.
For founders, the question is what the equivalent investment looks like in their market. It might be an API that partners build on, a format that becomes a standard, or training that makes your product the one people learn first. These things rarely pay off in the quarter they are built. They pay off when a market shifts and your product is the default.
The verdict
Nvidia's numbers are extraordinary, and the chips deserve credit. But the reason customers keep buying them, even at high prices and with alternatives on offer, goes back to a software bet made almost twenty years ago. The company did not just sell hardware. It taught a generation of programmers to build on its platform, and that habit is the hardest thing for a rival to buy.
Sources and filings
Every figure in this story comes from the documents below. If you spot an error, write to corrections@themarketanatomy.com.
This article is for information and education. It is not investment, legal or financial advice, and The Market Anatomy has no financial relationship with the companies named. See our editorial policy and disclaimer.


