OpenAI engineer Clive Chan has criticized the hardware recommendations in the V4 technical report, describing them as "surprisingly mediocre and even erroneous." Chan highlighted several issues, contrasting the V4 report with its predecessor, V3, which had been highly regarded for its detailed guidance at the ISCA academic conference. Chan specifically objected to the report's advice on power consumption, arguing that the recommendation for chipmakers to allocate additional power headroom is counterproductive. He explained that increasing power headroom could reduce operating frequency and computational performance. Additionally, Chan challenged the report's suggestion to use a pull-based data transfer method between GPUs, stating it is slower than a push-based approach. He also disagreed with the recommendation to replace SwiGLU activation functions, asserting that optimal performance can still be achieved with them. Chan speculated that the section might have been intentionally weakened by DeepSeek.