A flood of capital is surging into the processing-in-memory chips.

Recently, HOUMO Intelligence Technology Co., Ltd. (hereinafter referred to as "HOUMO Intelligence") announced the completion of a Series B financing of several billion RMB. The company stated that the investors in this round include upstream and downstream industrial capital, state-owned capital and international investors, many existing shareholders continued to make additional investments, and the specific list of institutions has not yet been announced.

The funds raised in this round will be used to expand the large-scale delivery of M50 chips, promote the R&D of the new generation M6X and the next-generation 3D computing-in-memory products, and improve the software toolchain and industrial ecosystem. In the past year, HOUMO Intelligence has successively completed multiple rounds of financing.

Multiple Rounds of Financing and Joint-stock System Reform Advancement

Founded in 2020, HOUMO Intelligence's founder Qiang Wu holds a doctorate in computer science from Princeton University, and has previously worked at Intel, AMD, and Facebook, and served as CTO of Horizon Robotics. In the early stage, the company laid out AI inference chips around scenarios such as intelligent driving and robotics. With the development of large model applications, it further expanded into the markets of personal computers, intelligent terminals and edge computing.

Since its establishment, HOUMO Intelligence has attracted multiple types of capital to participate. Institutions including HSG, Matrix Partners China, Qiming Venture Partners once participated in the early-stage financing, Lenovo Capital and Incubator Group has invested multiple times since the angel round; after that, Chengtong Mixed Ownership Reform Fund, funds under China Mobile and other institutions have entered one after another, and the sources of funds gradually cover market-oriented venture capital, state-owned capital and downstream industrial capital.

Along with successive financing, the company is also adjusting its organizational form. Industrial and commercial information shows that in April this year, HOUMO Intelligence completed the joint-stock system reform, and the enterprise name was changed from "Beijing Houmo Intelligence Technology Co., Ltd." to "HOUMO Intelligence Technology Co., Ltd.".

The core that supports HOUMO Intelligence's financing narrative is its continuously invested computing-in-memory technology route. Simply put, when traditional chips perform AI tasks, they need to frequently move data between the storage unit and the computing unit; the computing-in-memory technology attempts to complete part of the operations closer to the data, so as to reduce the time and energy consumption overhead brought by data movement. HOUMO develops chips based on this, and also provides boards, modules, software tools and reference solutions to help terminal manufacturers deploy models to actual products.

At present, the company's commercialization focus is the M50 released in July 2025. This chip is oriented to local large model inference, and can be used in personal computers with agent functions, intelligent terminals, robots and edge inference devices. For such customers, the purpose of purchasing chips is to enable AI applications to run continuously within the power consumption, heat dissipation and cost budget of the equipment.

According to the disclosure of HOUMO Intelligence, as of the end of August 2026, the M50 has obtained more than 150 product design wins; the company further disclosed in October that this product serves more than 100 customers. Qiang Wu said in an interview that the market demand for personal computers with agent functions and edge inference is growing rapidly, and other terminal directions are also continuing to advance.

Some cooperations have been implemented in specific terminals. In May this year, the Tiankai P7 released by Lenovo adopted the HOUMO M50 to provide computing power support for local AI applications. The all-optical network intelligent terminal launched by China Mobile in cooperation with HOUMO is oriented to scenarios such as offices and industrial parks. The funds under Lenovo and China Mobile previously participated in HOUMO's financing, and have become two cases to observe the progress of the company's industrial cooperation.

In addition to existing products, HOUMO is also advancing the R&D of the new generation M6X and deploying 3D computing-in-memory. The company previously disclosed that its first 3D CIM chip A20 has been put into production. This round of financing covers both M50 delivery and next-generation product R&D, which means the company needs to take into account existing customer onboarding, software adaptation and new product investment.

According to the data from the Finance Association Venture Capital - Zhongzhi Database, with November 2026 as the forecast benchmark time, the predicted probability of HOUMO Intelligence's subsequent financing in the next 2 years is 76.27%. This data is only a model estimation and does not represent the actual financing result.

Changes in Inference Demand Bring Continuous Large-scale Financing to Computing-in-Memory Track

In fact, not only HOUMO Intelligence, many computing-in-memory enterprises have recently announced the completion of large-scale financing. Capital is pouring intensively into this track.

On October 28, NovuMicro announced the completion of a Series C2 financing of over 1 billion RMB, led by Jian Tou Investment, with participation from Hangzhou Capital, Kaite Capital and other institutions. Combined with the previous Series C1 round, the company raised more than 2 billion RMB in two months. Its products target both the terminal and cloud-side inference markets, and the first three chips have been taped out, with plans to carry out customer testing in early 2027. The funds will support product iteration and mass production stocking.

On October 22, according to the disclosure of financial advisor YIWEI Capital, 9th Sense Semiconductor has accumulated 1 billion RMB in financing within half a year, and participating institutions include Meituan Strategic Investment, Starlight Capital under XPeng Motors, Yangtze Storage Industrial Fund, Yuanhe Puhua, OmniVision Group, Zhiwei Capital, miHoYo Strategic Investment and other institutions. Based on computing-in-memory chips, the company extends to inference boards and servers, and the financing is used for new chip R&D, tape-out and software ecosystem construction.

Behind this round of capital attention, an important change is that AI applications are increasing the continuous consumption of inference computing power.

In the past, users asked questions to the model and usually got one answer; when agents handle work tasks, they may need to repeatedly call the model, retrieve materials, use tools and check results. The longer the task chain, the more frequent the calculation and data calls, the more enterprises need to pay attention to the actual cost, response speed and energy consumption of completing a task.

At the relevant seminar of the Open Data Center Committee in May this year, researchers from the China Academy of Information and Communications Technology pointed out that the demands for state management, multiple iterations and tool calls brought by agents require the inference system to carry out collaborative optimization among computing, storage, network and scheduling. This also means that improving the peak computing power of chips is not enough to solve all problems.

For some large model inference tasks, the factor restricting the running speed also includes whether the data can be sent to the computing unit in time. Model parameters, context and intermediate results need to frequently access the storage, and moving data itself also consumes time and power. Routes such as computing-in-memory and near-memory computing have thus obtained new application opportunities: by adjusting the relationship between storage and computing, the operating efficiency of specific tasks is improved.

The terminal market further amplifies this demand. Personal computers, robots and edge devices have limited power supply, volume and heat dissipation conditions, and some tasks require low-latency response or local data processing. Chip manufacturers need to provide available AI capabilities under these constraints. When Qualcomm introduced its NPU architecture in September this year, it also listed continuous operation of agents, processing long context and reducing external memory access as optimization priorities.

From this perspective, the new opportunities for computing-in-memory come from the more specific requirements for inference cost and energy efficiency after large models gradually enter actual businesses and terminal products.

For HOUMO, what is worth observing next is how many of the existing product design wins can be converted into mass production deliveries, whether the terminals using its chips can form continuous sales, and whether software adaptation and customer support can maintain efficiency as the business expands. These progresses will determine whether the energy efficiency advantages of computing-in-memory can be further converted into revenue and repeat orders.

This article is from the WeChat Official Account "Sci-Tech Innovation Board Daily", Author: Xiao K, published with authorization from 36Kr.