OpenAI cut standard API prices for GPT-5.6 Luna by 80 percent and Terra by 20 percent on July 30, 2026, while leaving Sol unchanged. The company credited lower serving costs and better token-generation efficiency; Chinese vendors are also expanding lower-cost and open-weight alternatives.
NVIDIA's August 6, 2026, configuration guide separates AI training and inference systems and says hardware choices vary by workload. A buyer still needs a controlled run of its own model, software and traffic because peak chip specifications do not measure end-to-end service or three-year cost.
NVIDIA’s DGX H100 design guide rates one system at a maximum of 10.2 kW and a four-system rack at 40.8 kW. Facility teams need to plan from peak equipment demand, then verify the complete power, cooling and network design under load.
