NVIDIA was founded on 5 April 1993 by Jensen Huang, Chris Malachowsky and Curtis Priem, with the aim of bringing 3D graphics to the gaming and multimedia markets. [1]

Its first product, the NV1, was introduced in May 1995. It was ambitious: a single chip that combined 2D and 3D graphics, wavetable audio, full-motion video acceleration and a digital joystick interface. Diamond Multimedia sold it on the Diamond Edge 3D board, and Sega of America signed an exclusive licensing agreement with NVIDIA to bring Saturn games to the PC. [2]

The NV1 rendered 3D scenes using quadratic texture mapping rather than the triangle-based polygon rendering that later became the industry norm, and before Microsoft introduced DirectX NVIDIA had expected game developers to write to its own SDK. NVIDIA, Diamond and SGS could not get game developers to invest in the quadratic approach, the expensive board failed to win the market, and when Sega dropped NVIDIA after commissioning the follow-on NV2 chip for its Dreamcast console, the setback almost killed the company, according to the graphics analyst Jon Peddie. [2]

NVIDIA responded by abandoning its own rendering method and building a conventional triangle-based design. The result, the RIVA 128, arrived about two years after the NV1 and became one of the most successful graphics chips of its time, putting the company on a sound footing. [2]

In 1999 NVIDIA launched the GeForce 256, which it marketed as the world’s first graphics processing unit, or GPU. The defining feature was hardware transform and lighting: the geometric calculations needed to position and light 3D objects, previously done on the computer’s CPU, were moved onto the graphics chip itself. NVIDIA’s corporate timeline still lists 1999 as the year it invented the GPU. [1][3]

The first GPU claim is partly a marketing definition, since earlier chips also accelerated 3D graphics. What mattered more was the direction it set: each generation moved more work off the CPU and onto a specialised, increasingly programmable parallel processor.

The decisive step toward today’s NVIDIA came in November 2006 with the G80 chip, first sold as the GeForce 8800 and also used in the Tesla C870 compute card. G80 replaced separate vertex and pixel pipelines with a unified array of processors that could run graphics or general computing programs, and it was the first GPU that developers could program in C, without having to express their problems as graphics operations. [4]

The programming model that made this possible was CUDA, which NVIDIA describes as opening the parallel processing capabilities of GPUs to science and research. G80 also introduced the single-instruction, multiple-thread (SIMT) execution model, plus shared memory and barrier synchronization for communication between threads. Subsequent generations added double-precision arithmetic with the GT200 in June 2008 and error-correcting memory with the Fermi architecture, features aimed squarely at scientific computing. [1][4]

CUDA is arguably the most important decision in NVIDIA’s history. It created a software ecosystem, libraries and a generation of programmers trained on NVIDIA hardware, long before there was an obvious commercial market for it. That ecosystem, more than any single chip, explains the company’s position in AI today.

In 2012 Alex Krizhevsky, Ilya Sutskever and Geoffrey Hinton entered a deep convolutional neural network, later known as AlexNet, in the ImageNet image-recognition competition. The network had 60 million parameters and was trained for five to six days on two NVIDIA GTX 580 graphics cards with 3 GB of memory each. A variant of the model won with a top-5 test error of 15.3 per cent, against 26.2 per cent for the second-best entry. [5][1]

NVIDIA’s own timeline credits AlexNet with sparking the era of modern AI. The result showed that consumer GPUs running CUDA could train neural networks far larger than had been practical on CPUs, and researchers quickly adopted the approach. [1][5]

NVIDIA began building systems specifically for this market. In August 2016 Huang personally delivered a DGX-1, a server with eight Pascal-generation P100 GPUs rated at 170 teraflops of half-precision performance and listed at 129,000 dollars, to OpenAI in San Francisco. In May 2017 the Volta-based Tesla V100 arrived with 640 Tensor Cores, units designed specifically to accelerate the matrix arithmetic at the heart of deep learning. [6][7]

NVIDIA broadened from GPUs into networking. In April 2020 it completed its 7 billion dollar acquisition of the networking company Mellanox, giving it end-to-end technology from computing to networking for large clusters. In May 2020 it also introduced the Ampere-based A100, with 54.2 billion transistors and the ability to be split into as many as seven isolated GPU instances. [8][9]

The Hopper-based H100, announced in March 2022, added a Transformer Engine that dynamically chooses between FP8 and 16-bit calculations to accelerate the transformer models behind large language models. In 2018 the company had also introduced RTX, which it describes as the first GPU capable of real-time ray tracing, keeping its gaming business at the leading edge while data-center demand grew. [10][1]

The March 2024 Blackwell platform reflected how the product had changed. Its GPU joined two dies containing 208 billion transistors, and NVIDIA also offered it as the GB200 NVL72, a liquid-cooled rack of 72 GPUs and 36 Grace CPUs linked by NVLink. On January 5, 2026 NVIDIA announced the Rubin platform, six new chips including the Vera CPU, Rubin GPU and NVLink 6 Switch, saying Rubin was in full production and that partner products would be available in the second half of 2026. [11][12]

The shift shows clearly in NVIDIA’s results. In fiscal 2023, which ended on 29 January 2023, total revenue was 26.97 billion dollars, of which data center contributed a then-record 15.01 billion and gaming 9.07 billion. By fiscal 2026, which ended on 25 January 2026, revenue had reached 215.9 billion dollars, with data center accounting for 193.7 billion. [13][14]

Growth continued into fiscal 2027. For the quarter ended 26 July 2026, NVIDIA reported revenue of 96.2 billion dollars, up 106 per cent from a year earlier, including 89.0 billion from data center. The company said Vera Rubin was ramping into full production, with racks running at partners including CoreWeave, Google Cloud, Microsoft Azure, Oracle Cloud Infrastructure and Nebius, and guided to about 108 billion dollars of revenue for the following quarter, assuming no data-center compute revenue from China. [15]

As of September 2026, NVIDIA is primarily a data-center company whose graphics heritage is now a minority of its revenue. The same parallel-processing design that began with GeForce, and the CUDA software built around it, underpins the business.

NVIDIA’s cadence of a new data-center architecture every couple of years, with refreshes in between, shapes the whole GPU rental market. Each launch resets the performance frontier, while older generations such as A100 and H100 remain widely available and often become better value. On Kovara you can compare hourly prices for each NVIDIA generation across cloud providers, look up the specifications of a particular GPU, or ask Kova to weigh a newer part against an older one for your workload.