Nvidia

Nvidia is a fabless semiconductor company: it designs graphics processors and the systems around them, and pays TSMC in Taiwan to manufacture them on the most advanced process available. For its first fifteen years the customer was a gamer who wanted more frames per second. Today the customer is a handful of companies — Microsoft, Amazon, Google, Meta, Oracle and the model labs they fund — who buy accelerators by the hundred thousand to train and serve artificial-intelligence models. Data centre revenue reached about $115 billion in the year to January 2025, from $15 billion two years earlier. Gaming, the original business, is now under a tenth of the total.

The product is not really the chip. It is the chip plus CUDA, the programming platform Nvidia released in 2006 that lets developers write general-purpose code for a graphics processor, plus the libraries, compilers and networking that have accreted around it for nearly two decades. CUDA is free. The margin it enables is not: Nvidia's gross margin was 75% in FY2025, a figure that belongs to a software company and is being earned on hardware. A hyperscaler pays around $30,000 for an accelerator that costs a few thousand dollars to build, because the alternative is rewriting its entire software stack for a chip nobody's engineers know.

The constraint is not demand. It is TSMC's capacity on its advanced nodes and on CoWoS packaging, and the supply of high-bandwidth memory from SK Hynix, Samsung and Micron. Nvidia's revenue for the past three years has been, in effect, whatever those suppliers could make.

The moat is CUDA, and it is worth being precise about what CUDA is. It is not a patent or a chip design. It is a programming model, released in 2006 and given away since, that almost every researcher who has worked in machine learning has learnt to write in. PyTorch and TensorFlow were built on it. The optimised libraries for matrix multiplication, attention, and inference all assume it. A company that switches to a rival accelerator does not just buy different hardware; it rewrites, retests and re-tunes years of code, and it does so with engineers who trained on Nvidia. Those are switching costs, and they compound with each year of software written on top.

Around the switching cost sits a network effect: more developers on CUDA attract more libraries and tools, which attract more developers, and Nvidia funds the whole thing out of hardware margin. Scale finishes it. Nvidia spent about $12.9 billion on research and development in FY2025 and can afford to design a new architecture every year while pre-buying TSMC and memory capacity that smaller rivals cannot secure at any price.

Where the moat is thinner: the largest customers are all designing their own chips — Google's TPU, Amazon's Trainium, Microsoft's Maia, Meta's MTIA — and inference, where the software dependency is lighter than in training, is the fastest-growing workload. AMD's accelerators are credible hardware. And a model as efficient as DeepSeek's, which trained a frontier system on a fraction of the expected compute in January 2025, took roughly $590 billion off Nvidia's market value in a single day by suggesting that demand for compute might not be infinite. The moat protects the price of the chip. It does not protect the quantity anyone needs.

Nvidia's defining allocation decision was made in 2006 and looked like a mistake for six years. Jensen Huang committed the company to making every GPU it sold programmable for general-purpose computing, which cost die area, cost margin, and served a market — scientific computing, later machine learning — that barely existed. Gaming customers paid for a feature they did not use; investors saw the gross margin dilution and the stock fell. Nvidia kept funding CUDA through the 2008 financial crisis, through a defective-chip scandal that cost hundreds of millions, and through years of research revenue too small to report separately. In 2012 a neural network trained on two GeForce cards won the ImageNet competition by a margin nobody had seen before, and the market Nvidia had built a platform for arrived.

Since then the record has been a mix of reinvestment and return. Research spending has stayed near 10% of revenue at every scale, which at today's size is a $13 billion annual budget nobody else in accelerators can match. The one large acquisition, Mellanox for $6.9 billion in 2020, bought the networking that turns thousands of chips into one computer and now generates well over $10 billion a year; the attempted $40 billion purchase of Arm was blocked by regulators in 2022 and would have been the most consequential deal in the industry. Nvidia returned about $34 billion through buybacks in FY2025 and pays a token dividend.

The newer pattern deserves scrutiny. Nvidia has begun investing in its own customers — cloud providers, model labs, most notably a staged commitment of up to $100 billion in OpenAI — and in suppliers such as Intel. Vendor financing is how the telecoms equipment boom ended in 2001: suppliers lent customers the money to buy the equipment, and when the customers failed, the revenue turned out to have been the supplier's own capital coming back. Nvidia's version is equity rather than debt and the customers are better capitalised. But the shape is the same, and a student should know what shape it is.