Technology
GPUs
Graphics Processing Units enabling parallel computing for AI, gaming, scientific simulation, and cryptocurrency mining.
19 companies
Explainer
GPUs explained
A GPU, or graphics processing unit, is a chip with thousands of small cores that can run the same operation on a lot of data at the same time. It was first built to draw images in games. The same design suits any job that applies simple maths to large arrays of numbers.
Training and running neural networks is mostly matrix multiplication, which fits this design well. Modern data center GPUs add dedicated matrix units (NVIDIA calls them Tensor Cores), fast stacked HBM memory and high-speed links to other GPUs. Software layers such as CUDA from NVIDIA and ROCm from AMD let programmers use them.
How it differs from the usual approach
A CPU is built to finish one varied task quickly. A GPU is built to finish a huge number of similar tasks at once, which is called throughput. That makes GPUs far faster than CPUs on parallel work such as graphics and neural networks, and less suited to code with many branches. Compared with fixed-function AI chips, GPUs stay programmable, and that flexibility is a large part of their market position. So is their software ecosystem, which took years to build.
Where it is used
- Training large AI models in data centers.
- Running AI inference for chatbots, search and image generation.
- Rendering graphics in games, film and design software.
- Scientific simulation on supercomputers.
- Video encoding and analytics.
What to watch
- NVIDIA introduced CUDA in 2007 and still leads AI training. AMD and Intel are the other large GPU makers, and cloud providers also build their own accelerators.
- The US has restricted exports of the most advanced AI chips to some countries since October 2022, which shapes where and how they are sold.
- Supply of HBM memory and advanced packaging is a recurring constraint on how many GPUs can be built.
- Software portability layers aim to let models run on non-NVIDIA hardware without rewriting code.
Frequently asked questions
Why are GPUs used for AI?
Neural networks are mostly large matrix multiplications, and GPUs do many such operations in parallel. They also have fast memory and mature software, so they are practical to use.
What is the difference between a GPU and a TPU?
A GPU is a programmable parallel processor that began in graphics. A TPU is Google's chip designed specifically for the matrix maths in machine learning. TPUs are less flexible but can be efficient on the workloads they target.
Why is NVIDIA so dominant?
Its hardware is strong, and it also built the CUDA software ecosystem over many years. Most AI code and tools are written for it, which makes switching costly.
- OLOxmiq Labs Oxmiq Labs licenses graphics processor designs and software rather than selling its own chips. Its OxCore architecture is a GPU built on the open RISC-V instruction set, OxQuilt combines compute and memory chiplets in one package, and OXPython aims to run existing CUDA and PyTorch code on other hardware without changes. The company was founded by Raja Koduri, a former chief architect at Intel. Its first outside investor was MediaTek, which took part in a seed round announced in 2025. GPUs · Custom Silicon · AI Accelerators $20M seed (Aug 2025, MediaTek participated)
- ICIluvatar CoreX Iluvatar CoreX designs general-purpose GPUs for AI training and inference, along with the software stack and systems that run on them. Its Tiangai series targets training and its Zhikai series targets inference, and it also sells edge chips. General-purpose here means the chips are programmable like Nvidia GPUs, not fixed to one task. The company is based in Shanghai and listed on the Hong Kong Stock Exchange on 8 January 2026 under the ticker 9903. Wikipedia dates its founding to December 2015. GPUs · AI Accelerators HKEX IPO Jan 2026 raised ~$474M
- Intel Arc Intel Arc is the brand for Intel's discrete graphics processors, used for gaming, content creation and some AI work. It first appeared with the Alchemist generation, launched in March 2022 on laptops, and Battlemage cards such as the Arc B580 followed on 13 December 2024. Arc chips are designed by Intel and manufactured under contract by TSMC, which sets them apart from Intel's integrated graphics built into CPUs. Intel Corporation is headquartered in Santa Clara, California. GPUs —
- CoreWeave CoreWeave rents out GPU computing capacity in the cloud, mainly to AI developers and large enterprises that need to train and run models. It builds and operates its own data centres filled with Nvidia GPUs and sells access by the hour or by contract, rather than offering the broad range of services found at general-purpose clouds. The company was founded in 2017 by Michael Intrator, Brian Venturo and Brannin McBee, originally as Atlantic Crypto. Its shares trade on the Nasdaq under the symbol CRWV after a March 2025 IPO, and it is based in Livingston, New Jersey. GPUs · AI Accelerators IPO Mar 2025 raised $1.5B at $23B; $2B NVIDIA investment Jan 2026; ~$28B equity+debt over 12 months
- BTBiren Technology Biren Technology designs general-purpose GPUs and computing systems for AI and high-performance computing in China. Its BR100 chip, unveiled in 2022, was presented as a domestic alternative to Nvidia processors, and the BR104 is a single-die version of it. The company was founded in 2019 in Shanghai and was added to the US Entity List in October 2023, which restricts its access to some technology. It listed on the Hong Kong stock exchange on 2 January 2026. GPUs · AI Accelerators HK$4.85B (~$624M) Hong Kong IPO Jan 2026
- VeriSilicon VeriSilicon runs what it calls a silicon platform as a service model, which lets customers outsource much of the work of making a custom chip. It offers chip design services, volume production services that manage manufacturing, and licensable IP such as processor and graphics blocks. This makes it closer to a design partner than to a chip vendor, because customers own the finished product. The company is based in Shanghai and is listed on the Shanghai Stock Exchange under code 688521. Custom Silicon · AI Accelerators · GPUs —
- LTLisuan Technology Lisuan Technology designs graphics processors (GPUs) in Shanghai. It was founded in 2021 by three former S3 Graphics employees. Its in-house TrueGPU architecture is built on TSMC's 6 nanometer process and powers the LX 7G100 consumer card with 12 GB of GDDR6 memory, alongside a professional line with 24 GB aimed at multi-user and remote graphics work. The company markets the design as handling both rendering and AI inference in one GPU. GPUs —
- DTDenglin Technology Chinese GPGPU developer founded in 2017 whose GPU+ architecture targets CUDA and OpenCL workloads; its Goldwasser accelerator was described as China's first mass-produced GPGPU AI accelerator. Backers include Qualcomm Ventures and the China Internet Investment Fund, and a March 2026 report said it plans a Hong Kong IPO to raise US$300 to 400 million. GPUs · AI Accelerators Series C 2023 (China Internet Investment Fund); later round with Qualcomm Ventures
- Imagination Technologies Imagination Technologies is a British company that designs graphics processor architectures, sold under the PowerVR name, and licenses them to chip makers who build them into their own processors. It is an intellectual property supplier, so it does not sell chips of its own. Its newer GPU designs include the IMG DXD architecture, released in 2023. The company was founded in 1985 as VideoLogic, is based in Kings Langley, England, and was bought in 2017 by Canyon Bridge Capital Partners after leaving the London Stock Exchange. GPUs · AI Accelerators · CPUs & RISC-V —
- CDCurtiss-Wright Defense Curtiss-Wright Defense Solutions is the defense electronics arm of Curtiss-Wright Corporation, and it builds rugged embedded computing boards and systems for aircraft, vehicles and ships. Much of its range follows the VPX standard, a card format for ruggedized computers that plug into a shared backplane. It also offers GPU modules aligned with SOSA, a US military open architecture that lets parts from different vendors work together. The parent company was formed in 1929 and is headquartered in Davidson, North Carolina. GPUs · AI Accelerators —
- BGBolt Graphics Sunnyvale startup developing the Zeus GPU, a RISC-V based design for path tracing and FP64 HPC with up to 384 GB of LPDDR5X and DDR5 memory and 400 or 800 Gbps networking. It showed Zeus at CES 2026, announced tape-out of a TSMC 12 nm test chip in 2026, and reported test-chip physics results in October 2026; performance claims are the company's own and volume delivery is not expected before 2027 or 2028. GPUs · HPC & AI Systems —
- Moore Threads Moore Threads designs graphics processing units, or GPUs, which are chips built to run many calculations at the same time for AI, graphics and scientific computing. Its MUSA architecture is the company's own hardware and software platform, which it says is compatible with mainstream GPU ecosystems. It was founded in October 2020 by former Nvidia staff, and chief executive Zhang Jianzhong spent 14 years at Nvidia, including as China general manager. The company listed on Shanghai's STAR Market in December 2025. GPUs · AI Accelerators · HPC & AI Systems IPO Dec 2025 ($1.13B)
- JMJingjia Micro Changsha-based GPU designer founded in 2006 and listed on the Shenzhen Stock Exchange since 2016, with the JM5400, JM7200 and JM9 graphics chip series. It was added to the US Entity List in December 2021. On 17 January 2025 it disclosed that its JM11 series GPU met design targets in interim testing after tape-out. GPUs —
- INInnosilicon Wuhan-based chip and interface IP company that also designs GPUs. Its Fenghua No. 1 and No. 2 graphics cards used Imagination PowerVR IP; in September 2025 it presented Fenghua No. 3, a RISC-V based GPU with HBM, with CUDA compatibility claimed by the company and not independently verified. It also licenses HBM4/3E, GDDR7 and LPDDR6 PHY and controller IP and offers custom ASIC services. GPUs · Custom Silicon —
- NVIDIA NVIDIA designs GPUs, CPUs, DPUs and networking products such as NVLink, InfiniBand and Ethernet, and develops the CUDA software platform that runs on its GPUs. Its Data Center products are widely used for AI training and inference, and it also sells GeForce graphics cards for gaming and RTX workstation GPUs. NVIDIA is fabless. Foundries such as TSMC make its wafers, memory comes from suppliers including SK hynix, Micron and Samsung, and partners handle assembly and packaging. The company was incorporated in 1993 and is based in Santa Clara, California. GPUs · AI Accelerators · HPC & AI Systems —
- Glenfly (Granfi) Shanghai graphics chip company, the GPU arm of Zhaoxin, which holds 27.44% as largest shareholder. Its Arise-GT10C0 is a 28 nm discrete GPU, and the company also makes integrated graphics and AMOLED display driver chips. It completed A-share listing counselling registration in February 2025. GPUs —
- Lambda Labs Lambda, widely known as Lambda Labs, runs a cloud service that rents GPU servers for training and running AI models. It also began as a maker of GPU workstations and servers, and that hardware business is now a smaller part of its revenue. Unlike general-purpose clouds, it focuses only on AI workloads. The company was founded in 2012 by brothers Stephen and Michael Balaban. Most sources place its headquarters in San Jose, California, while the record lists San Francisco. In November 2025 it announced a multi-year agreement with Microsoft to deploy Nvidia GPUs. GPUs · AI Accelerators · HPC & AI Systems $1.5B Series E (Nov 2025, TWG Global-led)
- AMD AMD designs processors and accelerators across several product lines. These include EPYC server CPUs, Instinct data center GPUs for AI and high performance computing, Ryzen client processors, Radeon graphics cards, and FPGAs and adaptive SoCs from its Embedded business. Its ROCm software supports Instinct GPUs. AMD does not own fabs. It relies mainly on TSMC for wafers, with packaging and test handled by partners in Asia. The company was incorporated in 1969 and is headquartered in Santa Clara, California. GPUs · AI Accelerators · HPC & AI Systems —
- STShanghai Tianshu Zhixin Shanghai Tianshu Zhixin trades as Iluvatar CoreX and designs general-purpose GPUs, known as GPGPUs, for AI training and inference. A GPGPU is a graphics-style processor used for general parallel computing rather than for drawing images. The company was founded in Shanghai at the end of 2015 and lists Centurium Capital and HongShan among its investors. Its shares began trading on the Hong Kong Stock Exchange on 8 January 2026 under the code 9903. AI Accelerators · GPUs HKEX IPO (Jan 2026)