Chip AI startup Fangqing hits $1.5B valuation in third raise in six months
What's the deal? Shanghai Fangqing Technology, a Chinese chip startup, has reached a $1.5 billion valuation with a Series A round. The company is betting on disaggregated AI inference — yet it has not completed a chip tape-out or named a manufacturing partner.
Why now? The round is Fangqing's third in roughly six months, following its Pre-A and Pre-A+ deals. The pace reflects investor appetite for hardware that could cut the rising cost of running large language models at scale.
What's the endgame? Fangqing wants to split the transformer workload across dedicated hardware modules. One unit would handle attention, which is limited by memory bandwidth; the other would run the dense matrix maths of feedforward networks.
The pitch: a standard GPU must run both phases through one pipeline, forcing a trade-off between bandwidth, compute density, and power. Separating them could raise throughput and scalability — but the company has published no benchmarks on performance, efficiency, or power draw.
Who's behind it? Investor confidence rests partly on chief executive officer Liang Jun. He joined HuaweiDealroom has a profile for this one. Try Dealroom → in 2000 as lead architect of the HiSiliconDealroom has a profile for this one. Try Dealroom → Kirin SoCs used in its flagship phones, then served as CambriconDealroom has a profile for this one. Try Dealroom →'s chief technology officer from 2017, overseeing the Siyuan 290, 370, and 590 accelerators, before taking over Fangqing in 2024.
What could go wrong? The central risk is the "communication wall." Splitting the model means intermediate activations must cross an interconnect with high bandwidth and low latency; if that transfer costs more time or energy than specialisation saves, the system loses its edge over an integrated accelerator. Details on memory, packaging, and the production process remain undisclosed.
The signal: Capital is funding an architecture still unproven in silicon, but the market backdrop is expanding fast. Cited projections put growth for custom AI ASICs at 22%, against 19% for general-purpose GPUs, with the segment potentially reaching $118 billion by 2033. To turn valuation into an industrial advantage, Fangqing must show disaggregation genuinely improves performance per watt and inference costs.
Read more: Tom's Hardware
Image credit: brewbooks
