100 Petaflop AI Chip and 100 Zettaflop AI Training Data Centers in 2027

By Brian Wang

XAI is using about 150 Megawatts to power its 100,000 H100 liquid cooled AI training center in a few months. Nvidia H100s have about 4 petaflops of compute power. This would be 400 exaflops of peak computing.

The next step after that will be about 300,000 Nvidia B200s with a CX8 network in the summer of 2025. The ConnectX-8 InfiniBand SuperNIC provides 800 gigabits per second (Gb/s) of data throughput. This would be about 6 Zettaflops of compute. This will use about 500 Megawatts of power.

Blackwell Ultra will push performance forward by adding 12-Hi HBM3E compatibility. This will be more memory than with the Blackwell B200.

At Computex 2024, Nvidia announced the Rubin GPU architecture. The Rubin GPU will the chip after the 20 Petaflop Blackwell B200. The Rubin GPUs will be made in 2026.