Every System · Tracked

The DGX GB200 NVL72 is NVIDIA's rack-scale AI system: 72 Blackwell GPUs and 36 Grace CPUs in one liquid-cooled rack, linked by NVLink. The dataset lists 720 PFLOPS of FP8 compute and 120 kW of rack power. Known deployments include Microsoft Azure, Oracle Cloud, CoreWeave and xAI.

Total cost of ownership per PFLOPS per year puts different AI systems on one scale. BenchGecko adds the hardware cost amortized over 3 years to the power cost at $0.05/kWh, then divides by the system's FP8 compute in PFLOPS. Lower is better.

Current AI accelerators draw up to 1,200 W each in the dataset, and the DGX GB300 NVL72 rack draws 140 kW. Air cooling cannot remove that much heat efficiently. Liquid cooling (direct-to-chip or immersion) allows denser racks and a lower PUE: 6 of the 10 tracked systems are liquid cooled.

They are different architectures. The TPU v5p Pod links 8,960 TPU chips in one domain over Google's ICI fabric, while NVIDIA's NVL72 is a 72-GPU rack. TPU pods are only available on Google Cloud; DGX systems are sold by many OEMs for on-premise deployment.

The Cerebras CS-3 is built around a single wafer-scale chip, the WSE-3, which uses a whole silicon wafer. Instead of HBM stacks it has 44 GB of on-chip SRAM, which removes the memory bandwidth bottleneck of conventional GPU systems. The tradeoff is a higher per-system cost and a different programming model.

Google (TPU), AWS (Trainium), Microsoft (Maia) and Meta (MTIA) build their own AI chips to reduce dependence on NVIDIA, tune for their own workloads and lower cost at scale. Custom silicon also diversifies supply. Systems data as of Apr 14, 2026.