DeepSeek V4 and the FP4 Inflection Point

Much has been written about DeepSeek, but we believe one technical detail deserves closer attention from a hardware investment perspective: DeepSeek V4 is among the first frontier models trained using a mixture of FP8 and FP4 precision — a meaningful step down in numerical precision that roughly halves memory requirements compared to FP8 alone.

The hardware implication is significant. Among Chinese chipmakers, Huawei's Ascend 950 series (specifically the 950PR) is currently the only domestic chip to have publicly announced FP4 hardware acceleration support, delivering up to 2 PFLOPS at FP4 precision.

On the NVIDIA side, FP4 hardware acceleration remains exclusive to the Blackwell architecture within the data center lineup. As FP4 training gains mainstream traction — with DeepSeek V4 serving as a high-profile proof of concept — we believe demand for Blackwell chips stands to rise accordingly.