NVIDIA Blackwell platform arrives; B200/B300 GPUs ship with 4x H100 inference speed, 25x lower cost/energy

NVIDIA announced the Blackwell platform on June 11, 2026 at GTC, marking the arrival of its next-generation GPU architecture designed to power real-time generative AI at scale. The Blackwell portfolio includes B200 GPUs (192GB HBM3e, 8 TB/s bandwidth, native FP4 support) and B300 Blackwell Ultra (288GB HBM3e), delivering up to 4x faster LLM inference than NVIDIA H100 and 25x lower cost and energy consumption compared to its Hopper predecessor.

The flagship offering is the GB200 Grace Blackwell Superchip, pairing two B200 GPUs with NVIDIA Grace CPUs over 900GB/s NVLink, and the GB300 NVL72 rack-scale system with 72 Blackwell Ultra GPUs and 36 Grace CPUs optimized for test-time scaling and agentic AI reasoning. NVIDIA also announced project DIGITS, a personal AI supercomputer with the GB10 Grace Blackwell Superchip, bringing petaflop-scale AI performance to individual developers for prototyping and fine-tuning.

Early adoption is broad: Amazon Web Services, Google, Microsoft, Meta, OpenAI, Oracle, Tesla and xAI are among the first customers. OEM partners including Cisco, Dell, HPE, Lenovo and Supermicro are shipping RTX PRO Blackwell workstation and server variants. Cloud providers are expected to begin offering Blackwell instances within the quarter, though allocation constraints will likely persist into H2 2026.

For architects: Blackwell represents the first full-platform shift from Hopper—not just GPU but integrated CPU (Grace), networking (Quantum-X800 at 800Gb/s), and software (NVIDIA NIM inference microservices, TensorRT-LLM, new Dynamo inference-serving framework for test-time scaling). Budget infrastructure planning assuming 1-year cadence: NVIDIA announced Rubin (next generation) will arrive in 2027, and the company has formalized annual GPU release cycles, shifting from two-year cadence. Memory and power remain constraints; start procurement conversations now for H2 2026 and beyond.