Editor’s note: This is the debut of Flashpoint, Hello China Tech’s premium quick-strike column. In under 600 words, we unpack one market-moving China tech event and why it matters–before the noise takes over. Flashpoint appears whenever the story demands it – weekdays or weekends – starting with this Saturday launch. For the full column overview, see our introduction here.
A stray line in DeepSeek’s V3.1 update was enough to add billions to Chinese semiconductor valuations. The company disclosed that its model employs the “UE8M0 FP8” precision format.
Markets quickly extrapolated: if China’s leading open-source AI developer is aligning with block-FP8 standards, then domestic chipmakers might finally compete more effectively with Nvidia.
UE8M0 is a variant of FP8 where all eight bits represent the exponent. The format cuts memory traffic by as much as 75 per cent compared with FP32 and simplifies hardware logic. For accelerators constrained by bandwidth and cost, that matters. It allows mid-tier chips to run larger models without the expense of high-end HBM memory.
