Houmo AI Raises Several Billion Yuan on a Bet That AI Chips Should Compute Where They Store

Houmo AI, a four-year-old Chinese chip startup built on compute-in-memory architecture, said on November 4 that it has raised a Series B round of several billion yuan — the Chinese scale for "tens of" billions — with participation from upstream and downstream industrial capital, state-owned investors and international funds, plus continued support from existing backers. The investor list was not disclosed; earlier rounds included Sequoia China, Matrix Partners China, Qiming Venture Partners, Lenovo Capital and funds affiliated with China Mobile.

The company was founded in 2020 by Wu Qiang, who holds a PhD in computer science from Princeton and previously worked at Intel, AMD and Facebook before serving as CTO of Horizon Robotics. In April 2026 it completed a share reform and converted into a joint-stock company, a step Chinese firms typically take ahead of a domestic listing.

Houmo's commercial flagship is the M50, an edge inference chip launched in July 2025 that keeps computation inside the memory array instead of shuttling data between separate storage and logic. The company cites 160 TOPS of INT8 compute, 150 GB/s of memory bandwidth and roughly 10 watts of typical power — figures that are vendor-reported and not independently verified. Since entering production, it says the chip has won more than 150 product design wins as of late August 2026 and serves more than 100 customers across agent PCs, smart devices, robots and edge inference.

Deployments are the more telling signal. Lenovo's palm-sized KaiTian P7 AI host, which weighs about 300 grams and draws at most 30 watts, uses the M50 to run models of up to 122 billion parameters locally according to the companies; a Great Wall laptop runs 35-billion-parameter models on-device, and accessory maker Ugreen builds the chip into NAS devices with local AI features. The Series B money goes toward scaling M50 delivery, developing the next M6X generation and a 3D compute-in-memory successor, and expanding the software toolchain — including an open-source stack for quantization, compilation and deployment that the company says will roll out before January 2027.

The funding lands amid a shift in where the industry thinks inference will run. IDC estimates edge and terminal AI chip shipments grew 45 percent year over year in the first quarter of 2026 while cloud AI chip growth was 12 percent, and DRAM contract prices jumped 90 to 95 percent in the same quarter — squeezing exactly the data-movement costs compute-in-memory is designed to avoid. NVIDIA's reported plan to invest in d-Matrix, a digital compute-in-memory inference startup, suggests the mainstream has noticed.

For Houmo, the pitch is that edge inference will be judged on tokens delivered per yuan and per watt rather than peak TOPS. The hard part is now delivery: turning 150 design wins into shipped products, and a novel architecture into a software ecosystem developers do not have to fight.