Nvidia’s data-center GPU roadmap has run through four generations since 2022: Hopper (H100 and H200), Blackwell (B200 and GB200), Blackwell Ultra (B300 and GB300) and Vera Rubin, which Nvidia launched in January 2026 and says partners will offer from the second half of 2026. Each step has added memory per GPU, from 80 GB on the H100 SXM to 288 GB of HBM4 on a Rubin GPU, and Nvidia’s flagship system has grown from an eight-GPU server to a 72-GPU rack. This guide follows the roadmap as Nvidia itself announced it.

On this page

The roadmap at a glance

Dates are those of Nvidia’s own announcements. Memory figures are from Nvidia’s product pages; for some systems Nvidia gives a total rather than a per-GPU figure, and the table keeps Nvidia’s form.

Generation Products Announced by Nvidia Memory (as Nvidia lists it)
Hopper H100 March 22, 2022 H100 SXM: 80 GB HBM3, 3.35 TB/s; H100 NVL: 94 GB, 3.9 TB/s
Hopper H200 November 13, 2023 141 GB HBM3e, 4.8 TB/s
Blackwell B200, GB200 NVL72 March 18, 2024 HGX B200: 1.4 TB total (8 GPUs); GB200 NVL72: 13.4 TB HBM3e, 576 TB/s
Blackwell Ultra B300, GB300 NVL72 March 18, 2025 HGX B300: 2.1 TB total (8 GPUs); GB300 NVL72: 20 TB, up to 576 TB/s
Vera Rubin Rubin GPU, Vera Rubin NVL72 January 5, 2026 Rubin GPU: 288 GB HBM4, 19.2 TB/s; NVL72: 20.7 TB HBM4, 1,400 TB/s

Specs for each of these and for competing accelerators are in our AI chip catalog.

Hopper: H100 and H200

Nvidia announced the Hopper architecture and the H100 on March 22, 2022. It described the H100 as “packed with 80 billion transistors” on TSMC’s 4N process, and as the first GPU to use HBM3 memory. Nvidia said the H100 would be available from the third quarter of 2022. Its product page today lists the H100 SXM with 80 GB of memory at 3.35 TB/s and the H100 NVL with 94 GB at 3.9 TB/s.

The H200, announced on November 13, 2023, kept the Hopper architecture and raised memory to 141 GB of HBM3e at 4.8 TB/s. Nvidia said H200 systems were “expected to begin shipping in the second quarter of 2024.” Clouds still list both Hopper GPUs for rent; current H100 and H200 prices per GPU-hour are on our GPU price tracker.

Blackwell: B200 and GB200 NVL72

Nvidia announced Blackwell on March 18, 2024. Blackwell GPUs are “packed with 208 billion transistors” and made on “a custom-built 4NP TSMC process,” according to Nvidia. The generation added 4-bit floating point (FP4) inference through a second-generation Transformer Engine; Nvidia said Blackwell “will support double the compute and model sizes with new 4-bit floating point AI inference capabilities.”

Blackwell also changed the flagship product. Alongside eight-GPU HGX B200 servers, which Nvidia lists with 1.4 TB of total GPU memory, it introduced the GB200 NVL72: a rack of 36 Grace Blackwell Superchips, each with two Blackwell GPUs and one Grace CPU, for 72 GPUs and 36 CPUs acting as one system. Nvidia lists the rack with 13.4 TB of HBM3e and 576 TB/s of memory bandwidth.

Blackwell Ultra: B300 and GB300 NVL72

Blackwell Ultra followed on March 18, 2025. The GB300 NVL72 connects “72 Blackwell Ultra GPUs and 36 Arm Neoverse-based NVIDIA Grace CPUs in a rack-scale design,” and Nvidia said Blackwell Ultra products were “expected to be available from partners starting from the second half of 2025.” Memory rose again: Nvidia lists the GB300 NVL72 with 20 TB of GPU memory and the eight-GPU HGX B300 with 2.1 TB, against 1.4 TB for HGX B200. HGX B300 also doubles networking bandwidth to 1.6 TB/s, while NVLink bandwidth between GPUs stays at 1.8 TB/s per GPU, according to Nvidia.

Vera Rubin

Nvidia launched the Rubin platform on January 5, 2026 as “six new chips”: the Vera CPU, Rubin GPU, NVLink 6 Switch, ConnectX-9 SuperNIC, BlueField-4 DPU and Spectrum-6 Ethernet switch. The Rubin GPU delivers “50 petaflops of NVFP4 compute for AI inference,” according to Nvidia; its product page lists 35 petaflops of NVFP4 for training and 288 GB of HBM4 at 19.2 TB/s per GPU.

The flagship system, Vera Rubin NVL72, integrates “72 NVIDIA Rubin GPUs, 36 NVIDIA Vera CPUs” connected by NVLink 6. Nvidia lists it at 3,600 petaflops of NVFP4 inference, 2,520 petaflops of NVFP4 training and 20.7 TB of HBM4 with 1,400 TB/s of bandwidth. Nvidia said in January that “Rubin-based products will be available from partners the second half of 2026,” with AWS, Google Cloud, Microsoft and OCI among the clouds deploying Vera Rubin instances.

At GTC on March 16, 2026, Nvidia expanded the platform to seven chips in full production and added racks around the GPUs: a Vera CPU rack with 256 Vera CPUs, the Groq 3 LPX inference rack with 256 LPU processors and 128 GB of on-chip SRAM, BlueField-4 STX storage racks and Spectrum-6 SPX Ethernet racks. Nvidia’s claims for the platform against Blackwell are training with “one-fourth the number of GPUs” and “up to 10x higher inference throughput per watt at one-tenth the cost per token.” Those are Nvidia’s own comparisons. Nvidia said Groq 3 LPX entered full production on August 24, 2026.

What changes from generation to generation

Three trends run through Nvidia’s announcements:

  • Memory per GPU. 80 GB (H100 SXM), 141 GB (H200), 288 GB (Rubin). More memory lets one GPU hold a larger share of a model and its working data.
  • Lower-precision formats. Blackwell’s second-generation Transformer Engine added FP4 inference, and Nvidia quotes Rubin’s headline performance in NVFP4. Peak numbers quoted at different precisions cannot be compared directly.
  • The rack as the product. From GB200 NVL72 onward, Nvidia’s flagship is a 72-GPU rack rather than an eight-GPU server, and Vera Rubin adds CPU, inference, storage and networking racks designed around it.

Rack-scale systems of this kind fill the AI campuses listed in our AI data center tracker, where each project’s announced chips are recorded.

Sources