Why Nvidia Dominates AI Chips: CUDA, Systems and Scarcity

Nvidia dominates AI chips because it sells a complete computing platform, not a chip: GPUs, memory, the links that join them into racks, networking and the CUDA software that most AI code is written for. It built that platform for twenty years before demand arrived, so when the boom came in 2023 it was the only supplier whose systems were known to work at scale. Control of scarce supply then let it price on the value of the work rather than the cost of the silicon, at gross margins of about 70–75%.

The turn came fast. On 24 May 2023 Nvidia guided to about $11 billion of revenue for its next quarter, about 50% above analysts' expectations. Six months earlier it had been writing down gaming inventory and had just been barred from selling its best data-center chips to China. It reported $13.5 billion for that quarter, and on 30 May 2023 its market value passed $1 trillion. Revenue went from about $27.0 billion in fiscal 2023 to about $215.9 billion in fiscal 2026 (company), and it became the first company worth $4 trillion in July 2025 and $5 trillion in October 2025.

What did Nvidia bet on in 2006?

On 8 November 2006 Nvidia announced the G80 graphics architecture together with CUDA, a C-like programming model that let developers use a graphics chip's many simple cores for general calculation. The software was free. The important choice was to put this capability into every consumer graphics card, not a specialist product. Every gamer who bought a GeForce card added another CUDA-capable machine to the world, and every student had a parallel computer on the desk at gaming prices.

The bet was expensive. Making each consumer chip programmable added die area and power that gamers did not use, and Nvidia carried a large software team with no direct revenue. In fiscal 2009, through the financial crisis and a charge of about $196 million for defective laptop chips, the company lost money and its shares fell sharply. Jensen Huang, co-founder and chief executive since 1993, kept funding CUDA anyway (reported).

It paid off in 2012. Alex Krizhevsky trained AlexNet in CUDA on two Nvidia GTX 580 gaming cards and won the ImageNet competition by a wide margin. The deep-learning community standardized on Nvidia. The cuDNN library for neural networks followed in 2014, and the frameworks that came after, from TensorFlow to PyTorch, ran best on Nvidia from their first versions. By 2023 CUDA had about fifteen years of code, libraries and trained engineers behind it, and Nvidia claimed 5–6 million developers by 2025 (company claim).

What does Nvidia actually sell?

A lab building a training cluster does not buy GPUs one at a time. It buys a system:

• the GPU and the high-bandwidth memory packaged beside it;
• NVLink, which joins 72 GPUs in a GB200 NVL72 rack into a single memory domain;
• InfiniBand or Ethernet networking between racks, from the roughly $6.9 billion Mellanox acquisition that closed in April 2020;
• CUDA and the libraries on top of it.

Networking alone brought in about $11 billion in the fourth quarter of fiscal 2026 (company). Buyers choose Nvidia because the whole system is known to work at scale, the engineers they hire already know CUDA, and the models they want to run were tuned for Nvidia first. That is why an H100 could sell for a reported $25,000–40,000 in 2023 and a GB200 NVL72 rack for about $3 million (reported; Nvidia publishes no list prices).

Two further habits reinforce the position. During the 2023 shortage Nvidia decided who got scarce H100s, and it spread supply across hyperscalers and new GPU clouds such as CoreWeave and Lambda, creating buyers who would never design chips of their own. And from 2024 it moved to a one-year product rhythm, Hopper to Blackwell to Blackwell Ultra to Vera Rubin, so rivals aim at a moving target and customers always have a reason to buy again.

How can a chip get more expensive and still win customers?

The book works through a stylized example that explains how Nvidia raised prices while keeping buyers loyal. The numbers are illustrative and do not describe any real product.

Generation B costs half as much again per chip, yet cuts the cost of each unit of work by about 62%, because throughput rose fourfold. A vendor in that position can raise its price per chip and still lower its customers' cost per token. That is how Nvidia's revenue and its customers' loyalty rose together. The rule for anyone reading a chip business is simple: a rising chip price is sustainable only while each generation lowers the buyer's cost per unit of work faster.

Who else earns money from Nvidia's boom?

A buyer of an Nvidia system pays several margins at once. Each layer earns according to how scarce it is.

The lesson is that value follows scarcity, and scarcity moves. In 2023 the bottleneck was TSMC's CoWoS advanced packaging, which mounts the GPU beside its memory. In 2024–25 it was high-bandwidth memory, where Counterpoint estimated shares of about 50% for SK Hynix, 33% for Samsung and 18% for Micron in the second quarter of 2026 (reported). By 2025–26 it was power and data-center shells. Whoever owns the scarcest input in a given year earns the excess. The semiconductors and AI chips study tracks each shift.

What threatens Nvidia's dominance?

Nvidia's largest risk is that its biggest customers are also its best-funded competitors. If a merchant GPU carries a ~75% gross margin, about three-quarters of the price is the vendor's profit, research budget and moat. A buyer spending tens of billions of dollars a year can afford its own chip team and keep part of that margin.

Custom chips from the buyers

Google has done this longest. Its TPU has been co-designed with Broadcom since 2015, and in October 2025 Anthropic agreed to use up to one million TPUs. Amazon built Trainium and Project Rainier, a cluster of about 500,000 Trainium2 chips for Anthropic, and said in October 2026 that most inference on its Bedrock service already ran on Trainium. On 13 October 2025 OpenAI agreed with Broadcom to deploy 10 gigawatts of custom accelerators from late 2026 through 2029. Broadcom's AI semiconductor revenue was about $20 billion in its fiscal year to November 2025 (company).

AMD, the merchant challenger

AMD's MI300X found buyers at Microsoft, Meta and Oracle. On 6 October 2025 OpenAI agreed to deploy 6 gigawatts of AMD GPUs and received warrants for up to 160 million AMD shares, about a tenth of the company, vesting with deployment and share-price milestones; Meta signed similar terms in February 2026. A challenger that pays for its reference customer with its own equity says a good deal about how hard the incumbent is to displace.

Software that runs anywhere

Compilers and serving tools such as OpenAI's Triton, JAX, PyTorch's compiler, vLLM and SGLang let developers write once and target several chips. Inference, a few model architectures served at enormous scale, spreads the cost of porting over far more use than training does. The CUDA moat is therefore deep for the long tail of researchers and enterprises and for frontier training at most labs, but shallow for the handful of buyers who spend tens of billions a year.

Export controls

Politics has been as costly as competition. US export controls from 2022 onward, and Chinese counter-measures, took Nvidia's China data-center business from most of the market to close to nothing. A licence requirement on the H20, a chip designed for China, led to a charge of about $4.5 billion in 2025. The geopolitics and export controls study covers that history, and Chapter 21 of the book explains why almost every leading accelerator is still made and packaged in Taiwan.

How has Nvidia responded?

Nvidia's answers have been to keep CUDA free while charging for layers on top, to keep shipping a new architecture every year, and to buy or license what threatens it. It took a licence-and-hire of the inference-chip start-up Groq on 24 December 2025, reported at about $20 billion. It has also become a financier of its own demand: it agreed to buy capacity CoreWeave could not sell through April 2032 at an initial value of $6.3 billion, bought $2 billion of CoreWeave stock in January 2026, and made an equity investment of about $30 billion in OpenAI's 2026 round (reported). Those deals support demand, but they also put contingent risk on Nvidia's books and blur how much end demand is independent of Nvidia's own money.

The DeepSeek shock showed how sensitive the market is to that question. On 27 January 2025 Nvidia lost about $589 billion of market value in one day, then the largest one-day loss for any company, after a Chinese lab showed near-frontier results at low cost. The fear passed within weeks as the big platforms raised their capex plans, and Nvidia became the first $4 trillion company that July.

Which factor of the intelligence equation explains Nvidia?

The Intelligence Equation argues that an AI company creates durable value only when four factors hold at once: capability that survives real work, unit economics that hold at real usage, a position the next layer cannot route around, and the capital and permission to last the cycle. Nvidia's story turned on the third. CUDA, the full system and the power to allocate supply in a shortage gave it a position no buyer could route around in 2023–25, and the first two factors followed, since customers could verify the capability and each generation cut their cost per unit of work.

The threat is on the same factor. The few buyers large enough to build their own chips are trying to route around Nvidia, and how much margin they take back will decide how much of the boom's value it keeps. For the wider picture of how value flows down the AI stack, see our guide to how AI companies make money and the Nvidia company profile.

The takeaway

Nvidia dominates AI chips because it spent two decades building a platform, CUDA plus complete systems, that was ready when demand exploded, and because scarcity let it price on the value of the work. Its margins are a target: Google, Amazon, OpenAI and AMD are all trying to take part of them, and export controls have already closed China. The useful test for any chip business is whether each new generation lowers the buyer's cost per token faster than its price rises, and how many customers could build their own. Chapter 5 of the book, Chips, sets out the full case.