Investors are backing chips that run large models on PCs, phones and robots, and Houmo's 150 design wins give device makers a domestic option for local AI.
Houmo AI, a Chinese designer of compute-in-memory AI chips, has completed a Series B round worth several billion yuan, Wenhui Daily reported on November 4. The investors include industrial companies from the chip supply chain and its customers, state-owned capital and international investors, and several existing shareholders added to their stakes. Neither the investors’ names nor the exact amount were disclosed, though the mix of upstream and downstream industrial capital points to suppliers and device-maker customers among the backers.
The round follows several earlier raises in the past year, according to Wenhui Daily. Securities Times reported the same funding details, and Jiemian, citing a person familiar with the matter, said Houmo will use the money to scale up deliveries of its M50 chip, develop the next-generation M6X and a new 3D compute-in-memory product, and improve its software tools and partner ecosystem.
Compute-in-memory chips carry out calculations inside or right next to memory cells, rather than shuttling data back and forth between separate memory and processor blocks. That cuts the energy spent on moving data, a large share of the power needed to run AI models on small battery- or desk-powered devices.
The M50 is selling into PCs, phones and robots
Houmo AI’s M50 delivers 160 TOPS of INT8 compute with 150 GB/s of memory bandwidth at a typical power draw of 10 watts. Wenhui Daily reported the specifications. TOPS, or trillions of operations per second, is the standard measure of an AI chip’s raw speed, and INT8 refers to the lower-precision arithmetic commonly used to run trained models.
The M50 entered mass production less than a year ago and has since won more than 150 product designs and more than 100 customers, figures the company disclosed at its ecosystem partner conference, Jiemian reported. The design wins span four areas: AI agent PCs, smart mobile devices, embodied robots and edge inference servers.
One visible example is Lenovo’s AI host P7, which Lenovo launched in May and which was shown again at a developer conference organized by CSDN. The palm-sized device weighs 300 grams and uses the M50 for local inference at 10 watts, while the whole unit draws up to 30 watts. It can run models with up to 122 billion parameters on the device and is designed for continuous operation with data kept locally, according to Wenhui Daily and Lenovo’s launch materials.
Houmo is betting on 3D stacking for its next chips
Houmo’s roadmap moves toward three-dimensional designs that stack memory and logic. Its first 3D compute-in-memory chip, the A20, is already in production, and the M6X, the next generation of its M series, is under development, according to Wenhui Daily. Neither report gave specifications or a launch date for the M6X.
The company is also opening its software. Houmo plans to open-source its AI ecosystem software, covering model adaptation, compiler optimization and developer guidance, with related projects going online in stages before January, Wenhui Daily reported. For a chip startup, a working software stack is often what decides whether device makers can move models onto new hardware quickly.
Houmo was founded in 2020. Its chief executive, Wu Qiang, holds a doctorate from Princeton University, previously worked at AMD and Meta, and was chief technology officer of Horizon Robotics, the autonomous driving chip company. EastFrontier has tracked other Chinese startups building chips for on-device AI, including Transcputing’s latest raise for edge AI chips.
Shipments from the design wins will test the business
A design win means a customer has chosen a chip for a product, often one still in development. Lenovo’s P7 is already on sale, so the next sign of how much revenue Houmo’s 150 design wins will produce is shipment or sales figures for the P7, or a second major device maker bringing an M50-based product to market.
Two further markers are worth watching. The first is whether the open-source software projects appear on schedule before January, which would make it easier for smaller device makers to adopt the M50. The second is any disclosure of investor names, valuation or listing plans, which would show how the market is pricing China’s edge AI chip designers.