NVIDIA Blackwell Platform
The data-center GPU platform most frontier AI models are trained and served on.
Nobody buys a Blackwell chip at retail — entry means a seven-figure rack or renting by the GPU-hour. NVIDIA's platform spans B200 GPUs, GB200 superchips pairing two GPUs with a Grace CPU, and the GB200 NVL72: 72 GPUs and 36 Grace CPUs fused into one liquid-cooled rack that behaves like a single accelerator. The Blackwell Ultra refresh (GB300, with 288GB of HBM3e per GPU) now carries volume shipments, with the Vera Rubin generation queued behind it.
Best for
Hyperscalers, frontier labs, and enterprises provisioning training or high-throughput inference clusters — and, realistically, anyone renting Blackwell capacity through CoreWeave, AWS, Azure, or Google Cloud rather than buying racks.
Against AMD's Instinct MI300/MI350 line and Google's TPUs, Blackwell's real moat is NVLink rack-scale coherence plus CUDA: the NVL72 makes 72 GPUs address one memory fabric, and two decades of CUDA software mean every framework is tuned for it first. AMD now competes credibly on memory capacity and price per GPU; nobody yet matches the full rack-as-one-computer package.
Core features
AI compute
B200 and Blackwell Ultra GB300 GPUs built for frontier training and low-precision inference, with up to 288GB of HBM3e per GPU on the Ultra refresh.
Platform ecosystem
One stack from silicon to software: Grace CPUs, NVLink, InfiniBand and Spectrum-X networking, and the CUDA libraries every major AI framework targets.
Data-center scale
GB200 and GB300 NVL72 racks fuse 72 GPUs into a single coherent domain — the unit hyperscalers now plan AI factories around.
Pros
- + NVL72 rack-scale design makes 72 GPUs behave as one accelerator — no rival ships an equivalent turnkey unit
- + CUDA plus NVLink and InfiniBand networking mean every major framework and model is optimized for Blackwell first
- + The Blackwell Ultra (GB300) refresh extends the platform to 288GB of HBM3e per GPU without a rack redesign
Watchouts
- - Entry costs run $2–3 million per NVL72 rack, plus liquid cooling and power densities most existing data centers can't supply
- - The announced Vera Rubin successor generation puts a depreciation clock on every Blackwell purchase
Power List links
People behind NVIDIA Blackwell Platform
Alternatives
Frequently asked
What is NVIDIA Blackwell Platform?
The data-center GPU platform most frontier AI models are trained and served on.
Who builds NVIDIA Blackwell Platform?
NVIDIA Blackwell Platform is built by NVIDIA, a publicly traded company (NVDA).
How much does NVIDIA Blackwell Platform cost?
NVIDIA Blackwell Platform uses enterprise. Current tiers: No list price — B200 cited at $30K–$40K; NVL72 racks ~$2M–$3M. Check the official site before buying — AI pricing changes often.