NVIDIA

NVIDIA Corporation is an American multinational technology company founded in 1993 by Jensen Huang, Chris Malachowsky, and Curtis Priem. Originally established to address the 3D graphics processing demands of the PC gaming market, NVIDIA pivoted over the 2010s to become the dominant supplier of hardware and software infrastructure for artificial intelligence. By 2025, NVIDIA held approximately 92% of the discrete GPU market for AI workloads and had become the most valuable publicly listed company in the world by market capitalisation, driven by extraordinary demand for its data centre AI accelerators.

History

NVIDIA was founded in April 1993 in Santa Clara, California. The company launched its first GPU, the GeForce 256, in 1999, coining the term "GPU" (Graphics Processing Unit) to distinguish graphics accelerators from general processors. Gaming revenue sustained the company through its early decades.

The pivotal technological bet came in 2006 with the release of CUDA (Compute Unified Device Architecture), a parallel computing platform and API that allowed developers to write general-purpose programs that run on NVIDIA GPUs. CUDA transformed NVIDIA GPUs from graphics accelerators into programmable scientific computing engines. When deep learning researchers, led by Andrew Ng's group at Stanford and Geoffrey Hinton's group at the University of Toronto, discovered that GPU-accelerated training dramatically reduced the time required to train neural networks, NVIDIA's trajectory changed permanently.

AlexNet's GPU-trained victory in the ImageNet competition in 2012 demonstrated deep learning's potential at scale and validated CUDA-accelerated training. NVIDIA's data centre GPU revenue, negligible before 2012, began growing rapidly. The company released the Volta architecture with Tensor Cores in 2017, adding hardware specifically optimised for the matrix multiplication operations that dominate neural network training.

Key Products

Data Centre GPUs

NVIDIA's data centre GPU line has defined the performance trajectory of AI training:

| GPU | Architecture | Release | Key Use | |---|---|---|---| | V100 | Volta | 2017 | First-generation deep learning training | | A100 | Ampere | 2020 | Dominant training GPU 2020-2022 | | H100 | Hopper | 2022 | Large language model training | | H200 | Hopper | 2024 | LLM training with HBM3e memory | | B200 / GB200 | Blackwell | 2024-2025 | Next-generation training and inference | | Vera Rubin | Rubin | 2026 | Announced next-generation platform |

The H100 became the defining AI chip of the early large language model era. Clusters of thousands of H100 GPUs were used to train GPT-4, Gemini, Llama 3, and most major models released between 2022 and 2024. The Blackwell architecture (B200 and the NVL72 Grace Blackwell Superchip) was introduced in 2024 and 2025, offering substantially higher FP8 training throughput and a larger NVLink interconnect bandwidth.

CUDA and Software Ecosystem

CUDA is as important to NVIDIA's dominance as its hardware. The CUDA ecosystem includes cuDNN (deep learning primitives), cuBLAS (linear algebra), NCCL (multi-GPU communication), TensorRT (inference optimisation), and Triton Inference Server. Because deep learning frameworks including PyTorch, TensorFlow, and JAX are all built on CUDA, the switching cost from NVIDIA hardware is extremely high — rewriting training and inference stacks for alternative hardware requires substantial engineering effort.

DGX Systems

NVIDIA's DGX product line provides integrated AI computing appliances — from the DGX H100 (8 GPUs in a single server) to DGX SuperPOD (hundreds of nodes interconnected with NVLink and InfiniBand) — enabling organisations to deploy AI compute infrastructure without assembling their own hardware. In 2025, NVIDIA repositioned its data centre business around the concept of the AI Factory: a modular, scalable computing facility where intelligence is manufactured as an industrial output.

Market Position and Competition

NVIDIA's GPU market dominance has attracted substantial competition. AMD's Instinct MI300 series (2024) and MI350 (2025) offer competitive performance and have attracted customers including Microsoft and Meta seeking supply diversification. Google's Tensor Processing Units (TPUs) power internal training workloads and are available to Google Cloud customers. Amazon's Trainium and Intel's Gaudi accelerators also compete in training markets, though at smaller scale.

Despite competition, NVIDIA's CUDA ecosystem lock-in and its continuous product cadence have maintained its dominant position. CEO Jensen Huang announced in 2025 that the company had secured over 00 billion in orders for Blackwell and upcoming Rubin GPUs, underscoring the scale of demand.

NVIDIA in the Inference Era

As large language models transition from training to widespread deployment, inference — running trained models to serve user requests — has become an increasingly important revenue driver. NVIDIA's H100 and B200 GPUs excel at inference as well as training, and its TensorRT-LLM software optimises LLM inference throughput. The company's NIM (NVIDIA Inference Microservices) platform, launched in 2024, packages optimised model inference into containerised microservices, simplifying enterprise deployment.

References

1. Carbon Credits. (2025). NVIDIA Controls 92% of the GPU Market in 2025 and Reveals Next Gen AI Supercomputer. carboncredits.com.
2. Network World. (2025). Top 10 NVIDIA stories of 2025: From data center to AI factory. Network World.
3. The Edge Malaysia. (2025). YTL Power completes first Nvidia-powered AI data centre in Johor, YTL AI Cloud now operational. theedgemalaysia.com.
4. NVIDIA Investor Relations. (2025). Industry Leaders Transform Enterprise Data Centers for the AI Era With NVIDIA RTX PRO Servers. investor.nvidia.com.
5. Enkiai. (2025). NVIDIA's AI Strategy: Powering the 2025 T Data Center. enkiai.com.
6. Data Center Dynamics. (2024). Nvidia and YTL Power partner for .3bn AI data centres in Malaysia. datacenterdynamics.com.