GPU Technology Conference
Last reviewed
Sources
15 citations
Review status
Source-backed
Revision
v2 ยท 1,988 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
15 citations
Review status
Source-backed
Revision
v2 ยท 1,988 words
Add missing citations, update stale details, or suggest a clearer explanation.
The GPU Technology Conference (GTC) is the annual flagship technology conference hosted by Nvidia, and it is the principal stage on which Nvidia reveals new GPU architectures, data-center systems, and artificial-intelligence software. First held in 2009 as a developer event for general-purpose GPU computing, GTC has grown into one of the most closely watched gatherings in technology: its 2025 edition in San Jose drew more than 25,000 people in person and about 300,000 online. [1][2][9] The conference is anchored by a keynote from Nvidia co-founder and chief executive Jensen Huang, delivered in his trademark black leather jacket, which has become the customary venue for the company's biggest product announcements. As Nvidia's market value soared on demand for AI accelerators, press coverage came to describe GTC with superlatives such as the "Woodstock of AI" and, later, the "Super Bowl of AI." [1][2][3]
GTC is a multi-day event combining a CEO keynote, hundreds of technical sessions and tutorials, training offered through the Nvidia Deep Learning Institute, an exhibition hall of partners and customers, and developer workshops. Its subject matter spans the breadth of Nvidia's business: GPU architectures and data-center systems, high-performance computing, professional visualization, autonomous vehicles, robotics, healthcare and life sciences, and, increasingly, generative and agentic artificial intelligence. The flagship North American edition is held in San Jose, California, while online and regional editions extend the program to a global audience. [1][4]
The first GPU Technology Conference was held from September 30 to October 2, 2009, at the Fairmont San Jose hotel in California. It drew roughly 1,500 attendees and proved popular enough that Nvidia closed registration about two weeks ahead of the event. The conference grew out of Nvidia's earlier developer and visual-computing gatherings and reflected the company's push to position the GPU as an engine for general-purpose parallel computing through its CUDA programming platform. After 2009, Nvidia moved the conference to the larger San Jose Convention Center, which has served as its primary venue. [1]
Over the following decade GTC's emphasis shifted from graphics and high-performance computing toward deep learning and artificial intelligence, mirroring the broader industry's adoption of GPUs to train and run neural networks. GTC 2018 attracted more than 8,400 attendees, and the keynote eventually outgrew the convention center, with Huang's address moving to the adjacent SAP Center arena, which can hold about 17,000 people. [1][2]
Because of the COVID-19 pandemic, the 2020 and 2021 conferences were converted into digital-only events. GTC 2020's keynote, in which Huang introduced the Ampere A100 GPU, was famously recorded in his home kitchen rather than on a stage, and the digital event drew roughly 59,000 registrants, far more than any prior in-person edition. [1][5] The online format broadened GTC's reach, and Nvidia has continued to stream the keynote and many sessions free of charge even after returning to in-person gatherings.
Beyond the San Jose flagship, Nvidia has staged regional GTC editions in markets including Europe, Israel, Japan, China, and Taiwan, and in October 2025 held a Washington, D.C. edition. These regional events typically feature localized programming and, at times, a Huang keynote, extending the conference's reach to developers, enterprises, and governments outside the United States. [1][6]
The GTC keynote has become the focal point of the conference and a closely scrutinized event for the semiconductor and AI industries. Huang uses the address to lay out Nvidia's roadmap, demonstrate new systems on stage, and frame the company's view of computing trends, such as his recurring argument that "accelerated computing" is supplanting general-purpose computing. Beginning in the 2020s, Nvidia adopted an annual cadence for its data-center platforms, and the GTC keynote became the natural venue to reveal each year's new GPU architecture and rack-scale system. The keynotes have introduced or detailed a succession of architectures, including Pascal and NVLink (2014), Volta (2017), Ampere (2020), Hopper (2022), Blackwell (2024), and the Vera Rubin roadmap (2025 onward). [1][7]
The table below summarizes major architectures, systems, and software unveiled or detailed in Nvidia's GTC keynotes. Some architectures previewed at one GTC shipped in later years.
| Year | Format / venue | Selected announcements |
|---|---|---|
| 2009 | San Jose (inaugural) | Inaugural conference; emphasis on GPU computing and the forthcoming Fermi architecture |
| 2014 | San Jose | Pascal architecture preview; NVLink high-speed interconnect |
| 2017 | San Jose | Volta architecture and the Tesla V100 data-center GPU |
| 2018 | San Jose | DGX-2 server with NVSwitch; 32 GB Tesla V100 |
| 2020 | Online (digital) | Ampere architecture and the A100 GPU; DGX A100 (keynote filmed in Huang's kitchen) |
| 2021 | Online (digital) | Grace CPU; Omniverse Enterprise |
| 2022 | Hybrid / San Jose | Hopper architecture and the H100 GPU; framing of data centers as "AI factories" |
| 2024 | San Jose / SAP Center | Blackwell architecture; B200 GPU; GB200 Grace Blackwell Superchip and GB200 NVL72; NIM microservices; Project GR00T; Omniverse Cloud APIs |
| 2025 | San Jose / SAP Center | Blackwell Ultra (GB300 NVL72); Vera Rubin and Rubin Ultra roadmap with Feynman beyond; Spectrum-X and Quantum-X silicon photonics; DGX Spark and DGX Station; Isaac GR00T N1; NVIDIA Dynamo |
| 2026 | San Jose / SAP Center | Vera Rubin platform (Vera CPU and Rubin GPU) into production; Feynman architecture with Rosa CPU; robotaxi and automotive partnerships |
GTC 2024 was held in San Jose beginning March 18, 2024, with Huang delivering the keynote at the SAP Center arena before more than 12,000 in-person attendees and many more online. The centerpiece was the Blackwell platform, which Nvidia positioned as a processor designed for the generative-AI era. "We created a processor for the generative AI era," Huang said. [8] He held a Blackwell die alongside a Hopper chip to illustrate the leap in scale, and Nvidia said Blackwell delivered roughly 2.5 times Hopper's FP8 training performance per chip and up to 5 times its performance using the new FP4 format for inference. [11] The B200 GPU packs 208 billion transistors and delivers up to 20 petaflops of FP4 compute. [12] The keynote also introduced the GB200 Grace Blackwell Superchip, which pairs two B200 GPUs with a Grace CPU over a 900 GB/s NVLink chip-to-chip link, and the GB200 NVL72, a liquid-cooled rack-scale system combining 36 Grace CPUs and 72 Blackwell GPUs into what Nvidia described as a single rack delivering 1.8 exaflops of AI performance from about 600,000 parts. [11][12] Software and platform news included NIM (Nvidia Inference Microservices) for packaging and deploying AI models, the Project GR00T foundation model for humanoid robots, the Jetson Thor and DRIVE Thor computing platforms, Omniverse Cloud APIs, and the Earth-2 climate-simulation platform. [8] Bank of America analysts dubbed the gathering the "Woodstock of AI," a phrase widely repeated in coverage. [3]
GTC 2025 took place in San Jose from March 17 to 21, 2025, with Huang's keynote on March 18 at the SAP Center, which Nvidia and the press said drew more than 25,000 people in person and about 300,000 online. [2][9] Huang opened by firing T-shirts into the crowd and christened the event the "Super Bowl of AI," quipping that "the only difference is everybody wins at this Super Bowl. Everybody's a winner." [2] He said the Blackwell platform was in full production and, paired with the new Dynamo inference software, delivered up to 40 times Hopper's performance on reasoning models. [9] The keynote laid out an annual data-center roadmap: Blackwell Ultra, delivered as the GB300 NVL72 and aimed at AI reasoning and "test-time" inference scaling, for the second half of 2025; the Vera Rubin generation (a new Vera CPU paired with Rubin GPUs), with the Vera Rubin NVL144 system slated for the second half of 2026; and Rubin Ultra in an NVL576 configuration for the second half of 2027. Huang also named the architecture beyond Rubin after physicist Richard Feynman. [2][7][9][13]
Networking announcements included Spectrum-X Photonics and Quantum-X Photonics, co-packaged silicon-photonics switches that Nvidia said would improve power efficiency and network resiliency for massive AI clusters. On the systems side, Nvidia unveiled two desktop machines built on Grace Blackwell silicon: DGX Spark, a personal AI supercomputer previously known as Project DIGITS, and the larger DGX Station. In robotics and physical AI, Huang introduced Isaac GR00T N1, described as an open, customizable foundation model for humanoid reasoning, alongside the Newton open-source physics engine developed with Google DeepMind and Disney Research, the Cosmos world-foundation models, and the Halos automotive-safety system. The keynote also presented NVIDIA Dynamo, open-source software characterized as an "operating system" for AI factories, and announced a collaboration with General Motors on AI for vehicles, factories, and robots. [9][13][14]
GTC 2026 was held in San Jose from March 16 to 19, 2026, with Huang's keynote again at the SAP Center. [10][15] The headline was the move of the Vera Rubin generation toward production: Nvidia described Vera Rubin as a full-stack platform comprising seven chips, five rack-scale systems, and one supercomputer aimed at agentic AI, and detailed the Vera CPU, Rubin GPU, and BlueField-4 storage architecture. [10] Huang also previewed the successor Feynman architecture, including a Rosa CPU named after the scientist Rosalind Franklin and the Kyber rack design, and highlighted physical-AI and automotive efforts such as robotaxi programs with partners including Uber, BYD, Hyundai, Nissan, and Geely. [10] Reporting from the keynote noted that Huang said he expected purchase orders for Blackwell and Vera Rubin to reach $1 trillion through 2027. [15]
GTC has come to function as a barometer for the artificial-intelligence and semiconductor industries. Because Nvidia supplies the bulk of the accelerators used to train and run large AI models, the architectures, systems, and roadmaps revealed at the conference shape capital-expenditure plans across cloud providers, enterprises, and AI developers, and the keynote is followed closely by investors and analysts. The event's evolution, from a 1,500-person developer meeting in 2009 to an arena-scale spectacle drawing more than 25,000 people in person and hundreds of thousands online, mirrors both Nvidia's transformation into one of the world's most valuable companies and the broader rise of accelerated computing and generative AI. [1][2][9] The press shorthand of the "Woodstock of AI" and the "Super Bowl of AI" captures the conference's status as a marquee fixture of the AI era. [2][3]