NVIDIA

NVIDIA is an American semiconductor company based in Santa Clara, California, and the largest supplier of the accelerators used to train and run large AI models. It was founded in 1993 to build graphics chips for personal computers and now sells complete rack-scale systems, networking equipment and software 1.

The company's position is unusual in the history of computing. It does not fabricate its own chips — those are made by TSMC — and it does not sell the models that made its hardware valuable. It sells the machinery, the programming model and the interconnect, and it has captured a large share of the spending that the AI boom has generated 2.

In 2025 NVIDIA became the first company whose market capitalisation passed four trillion dollars, and then five. The figures reflect expectations about future demand for computing more than they reflect any single product, and they have made the company the subject of both admiration and a sustained argument about whether the spending it depends on is sustainable 3.

NVIDIA was founded in April 1993 by Jensen Huang, Chris Malachowsky and Curtis Priem, three engineers who had worked at LSI Logic and Sun Microsystems. Their thesis was that personal computers would need dedicated hardware for three-dimensional graphics, and that a company focused only on that problem could out-engineer the larger chipmakers 1.

The early years were precarious. The first product failed commercially, the company nearly ran out of money in 1996, and its survival depended on a contract with Sega for a games console chip that was never released. The turnaround came with the RIVA 128 in 1997, a graphics chip that shipped in volume, and with the decision to design a new architecture every year rather than every few years 1.

NVIDIA listed on Nasdaq in January 1999 and released the GeForce 256 that year, marketing it as the first graphics processing unit — a chip that offloaded transformation and lighting from the CPU. The name stuck, and the GPU became the company's core product 1.

Huang has led the company since its founding, one of the longest tenures of any chief executive in the semiconductor industry, and its culture reflects his preference for flat teams and public technical detail 2.

For its first fifteen years NVIDIA sold chips to people who wanted to render images faster. The property that made them useful for other work was accidental: a GPU performs the same operation on many pieces of data at once, which is exactly what neural networks need for matrix multiplication 1.

The turning point was AlexNet in 2012. Two researchers trained a convolutional network on two GeForce cards and won the ImageNet competition by a wide margin, demonstrating that consumer graphics hardware could do serious machine learning. NVIDIA responded by building hardware and software for that market rather than treating it as a curiosity 4.

The data-centre roadmap followed: the Kepler generation, then Pascal and the DGX-1 in 2016, a server with eight GPUs designed for deep learning and given to OpenAI as its first unit. Volta, in 2017, added tensor cores — specialised units for the matrix operations that neural networks use — and established the pattern of a new architecture every two years 5.

Ampere, in 2020, and Hopper, in 2022, became the standard for large-model training. The H100, with 80 gigabytes of high-bandwidth memory and a transformer engine that accelerates attention, was the defining chip of the 2023–25 boom. Blackwell, announced in 2024, moved from a chip to a rack-scale design, and the GB200 NVL72 packages 72 GPUs and 36 CPUs into a single liquid-cooled unit 6. The generational cadence has become a commitment: NVIDIA has said it will release a new architecture every year, with Blackwell Ultra and Rubin following in succession, a pace that requires simultaneous work on design, packaging, memory supply and cooling 6.

CUDA, released in 2006, is the reason NVIDIA's hardware advantage has persisted. It is a programming model that lets developers write general-purpose code for the company's GPUs using an extension of C, together with a set of libraries for linear algebra, deep learning, signal processing and communication 4.

The strategic effect is a moat built from habit and accumulated code. Decades of scientific software, thousands of research projects and most machine-learning frameworks target CUDA first, because that is where the libraries and the optimised kernels are. A competitor's chip may be faster on paper, but porting a codebase and re-tuning it is expensive, and the ecosystem around CUDA is deeper than any rival's 4.

NVIDIA has defended the position by investing in the layer above the driver. cuDNN provides optimised primitives for neural networks; TensorRT compiles models for deployment; NCCL handles communication between GPUs; and the company contributes to PyTorch and other frameworks so that new features work on its hardware first 1.

The moat is not absolute. Compilers such as Triton and MLIR-based toolchains let developers target different accelerators from a common source, and cloud providers have built their own software stacks for their own silicon. But the switching cost remains high, and it is highest for the largest and most performance-sensitive workloads 4.

NVIDIA's most important strategic decision of the last decade was to sell systems rather than components. A modern cluster is limited less by the arithmetic throughput of individual chips than by how fast they can exchange data, and the company has built the interconnect, the switches and the reference designs that determine how a data centre is laid out 6.

NVLink connects GPUs within a server at high bandwidth; InfiniBand, which came with the 2020 acquisition of Mellanox for about 6.9 billion dollars, connects servers across a cluster; Spectrum-X provides an Ethernet-based alternative for customers who prefer open networking. The GB200 NVL72 extends the NVLink domain to a whole rack, so that 72 GPUs behave like one large accelerator 6.

The result is that NVIDIA sells AI factories rather than chips. A customer buying a cluster buys GPUs, CPUs, switches, cables, software, reference architecture and, often, the company's engineers to help install it. That bundling raises the average selling price and makes the company's products harder to substitute piecemeal 1.

The business has scaled accordingly. In the fiscal year ended January 2025, NVIDIA reported revenue of 130.5 billion dollars, of which 115.2 billion came from the data centre segment, with gross margins around 75 per cent. No semiconductor company of comparable size has earned margins of that kind 2.

NVIDIA is a public company with no controlling shareholder. Its shares are held largely by institutional investors, and its founder holds a stake worth tens of billions of dollars but a small fraction of the total. The company has few employees relative to its revenue — around 30,000 at the end of the 2025 fiscal year — because it designs chips and software and outsources fabrication, packaging and assembly 2.

Its financial position is exceptional. Revenue grew by more than 100 per cent in the 2025 fiscal year, and the company generated enough cash to fund a large share buyback while increasing capital expenditure on its own infrastructure. It carries little debt relative to earnings, and its supply commitments to TSMC and memory makers have grown to tens of billions of dollars 2.

The market capitalisation has become a cultural phenomenon. NVIDIA passed one trillion dollars in 2023, four trillion in July 2025, and five trillion in October 2025, making it briefly the most valuable company in the world. Those valuations rest on the assumption that demand for AI computing will continue to grow faster than supply for several years 3.

The company's customer base is concentrated. A handful of hyperscalers and AI laboratories account for a large share of data-centre revenue, and several of them are designing their own accelerators. That concentration is both a source of visibility and a structural risk 2.

China was once NVIDIA's largest market for data-centre chips, and US export controls have progressively closed it. Restrictions imposed in 2022 and tightened in 2023 barred the sale of the most capable accelerators, and the company responded by designing reduced-capability parts — the A800 and H800, and later the H20 — that would fit within the rules 7.

In April 2025 the United States required licences for H20 exports as well. NVIDIA disclosed a 4.5 billion dollar charge for excess inventory and purchase obligations and said it could not ship a further 2.5 billion dollars of revenue in the quarter. The H20 had been the company's main compliant product for China, and the restriction effectively ended that business for a period 7.

The policy environment has been unstable in both directions. Licences have been granted and withheld, and in 2025 the US government agreed to allow some China sales in exchange for a share of the revenue — an arrangement reported at 15 per cent, and one that raised questions about the legal basis for export taxation. NVIDIA has said it will comply with whatever rules apply while warning that restrictions accelerate Chinese alternatives 7.

Those alternatives are real. Huawei's Ascend chips, Cambricon and a cluster of domestic start-ups have taken share in the Chinese market, supported by state procurement and by the difficulty of buying American hardware. The longer the controls last, the more capable those substitutes are likely to become 8.

China's competition regulator opened an investigation into NVIDIA in December 2024, citing suspected violations of the Anti-Monopoly Law and of the conditions attached to the company's 2020 acquisition of Mellanox. The conditions had required NVIDIA to license interconnect technology to competitors and not to bundle products; the investigation followed the imposition of US chip controls and was widely read as a reciprocal move 8.

The export controls themselves are the largest dispute in which the company is a party rather than a target. NVIDIA argues that restrictions on sales to China harm American competitiveness, strengthen Huawei and reduce the revenue available for research. Critics of the policy argue that the chips are a strategic asset and that permitting sales undermines the controls. The company has lobbied intensively on both sides of the argument, with limited success 7.

The company has also been criticised for the concentration of its customer base and for its allocation practices. During the shortage of 2023 and 2024, access to H100s depended on relationships as much as on orders, and smaller laboratories and start-ups complained that they could not obtain capacity while large customers received preferential treatment. NVIDIA has said it allocates fairly and that supply was genuinely constrained 2.

A further line of criticism concerns the circularity of the AI economy. NVIDIA has invested in AI companies that buy its chips, and some of those companies have raised capital partly on the strength of their access to NVIDIA hardware. Analysts have asked whether the resulting revenue is as independent as it appears 3.

Whether the demand for AI computing is durable is the largest open question in the industry, and NVIDIA is the company most exposed to the answer. The optimistic case is that inference, agents and robotics will require far more computing than training, that the largest models are still improving with scale, and that the build-out is in its early stages. The pessimistic case is that the spending is concentrated in a few companies whose returns have not yet materialised, and that a slowdown in their capital budgets would cascade through the supply chain 2.

The durability of the CUDA moat is also contested. NVIDIA's software ecosystem is deep, but customers have a strong incentive to escape a single supplier, and the largest of them — Google, Amazon, Microsoft and Meta — already design their own chips for some workloads. NVIDIA's answer is to keep moving the frontier faster than competitors can follow, which requires it to remain a generation ahead in both hardware and software 4.

A third disagreement concerns valuation. At five trillion dollars, NVIDIA is priced as a company that will grow into its multiple, not as one that has already earned it. Supporters point to the earnings and the visibility of demand; sceptics point to the history of semiconductor cycles, in which shortages become gluts within a year or two 3.

NVIDIA is the closest thing the AI industry has to a chokepoint. Almost every frontier model was trained on its hardware, almost every inference service runs on it, and its roadmap sets the pace for the entire supply chain, from TSMC's advanced packaging to the memory makers' high-bandwidth products 1.

Its dependence on TSMC is the most obvious vulnerability, and the one least within its control. Taiwan's foundries produce the most advanced chips in the world, and a disruption there would affect NVIDIA more than any competitor. The company has diversified some packaging and assembly, but the leading-edge logic remains concentrated 2.

For the encyclopedia, NVIDIA is the company that turned a games chip into the infrastructure of a new kind of computing, and then into one of the most valuable businesses in the world. Its story is a reminder that in a technology boom the suppliers of the tools can earn more than the people using them 3.