NVIDIA Blackwell
Blackwell is NVIDIA’s data-center GPU architecture, announced at its GTC conference on March 18, 2024 as the successor to Hopper and named after the mathematician David Harold Blackwell.1 NVIDIA says every Blackwell GPU packs 208 billion transistors on a custom TSMC 4NP process, as two reticle-limited dies joined by a 10 TB/s chip-to-chip link that act as a single GPU.2 The architecture added a second-generation Transformer Engine with 4-bit floating-point (FP4) support.1 It ships in two main forms covered on MetaPrompt: the GB200 Grace Blackwell Superchip, which pairs two Blackwell GPUs with a Grace CPU and fills the liquid-cooled, 72-GPU GB200 NVL72 rack,3 and the eight-GPU HGX B200 server board.1 The later Blackwell Ultra GPUs, in the GB300 NVL72 and HGX B300, are a separate part.35
NVIDIA’s datasheet gives each Blackwell GPU in the GB200 NVL72 186 GB of HBM3E at 8 TB/s and a configurable maximum of 1,200 W, against 180 GB at 7.7 TB/s and up to 1,000 W in the HGX B200. It rates the GB200 NVL72 GPU at 10 PFLOPS of FP8 with sparsity, and states dense throughput as one-half of that, 5 PFLOPS.2 Those memory figures put Blackwell at #10 and #11 in AI Accelerators With the Most HBM per Chip (October 2026). NVIDIA said on February 4, 2025 that CoreWeave had become the first cloud provider to make Blackwell generally available, with GB200 NVL72-based instances,4 and its HGX page says HGX B200 is shipping now.5
Blackwell also fills some of the largest AI campuses. Crusoe says Oracle began delivering the first GB200 racks to the Stargate campus in Abilene, Texas, in June 2025, a site covered in Stargate’s Sites: Operating vs Under Construction vs Planned.6 Microsoft describes its Fairwater datacenters as built on GB200 NVL72 racks that can scale to hundreds of thousands of Blackwell GPUs,7 and SpaceX’s S-1 filing says the first cluster at xAI’s Colossus II, about 110,000 GB200 processors and about 210 megawatts of compute power, came online in 91 days.8 Those sites appear in Largest AI Data Centers by Estimated Power Draw (October 2026).
Key facts
| Maker | NVIDIA1 |
|---|---|
| Announced | March 18, 2024 (GTC)1 |
| Transistors and process | 208 billion per GPU, two dies, custom TSMC 4NP Vendor-stated2 |
| Memory per GPU, GB200 NVL72 | 186 GB HBM3E, 8 TB/s Vendor-stated2 |
| Memory per GPU, HGX B200 | 180 GB HBM3E, 7.7 TB/s Vendor-stated2 |
| FP8 per GPU, GB200 NVL72 | 10 PFLOPS sparse; 5 PFLOPS dense Vendor-stated2 |
| Maximum TDP per GPU | Up to 1,200 W (GB200 NVL72); up to 1,000 W (HGX B200) Vendor-stated2 |
| First cloud general availability | February 4, 2025 (CoreWeave, GB200 NVL72)4 |
As of . NVIDIA’s datasheet states Tensor Core throughput with sparsity; dense is one-half. How figures are labeled.
Appears in
Sources
- NVIDIA Newsroom, “NVIDIA Blackwell Platform Arrives to Power a New Era of Computing” (March 18, 2024), retrieved 2026-10-13.
- NVIDIA, NVIDIA Blackwell Datasheet (PDF), retrieved 2026-10-13.
- NVIDIA, NVIDIA GB200 NVL72, retrieved 2026-10-13.
- NVIDIA Blog, “NVIDIA Blackwell Now Generally Available in the Cloud” (February 4, 2025), retrieved 2026-10-13.
- NVIDIA, NVIDIA HGX Platform (HGX Rubin NVL8, HGX B300 and HGX B200 specifications), retrieved 2026-10-13.
- Crusoe, “Crusoe Announces Flagship Abilene Data Center is Live” (September 30, 2025), retrieved 2026-10-13.
- Microsoft, “From Wisconsin to Atlanta: Microsoft connects datacenters to build its first AI superfactory”, retrieved 2026-10-13.
- Space Exploration Technologies Corp., Form S-1 registration statement (SEC EDGAR), retrieved 2026-10-13.