In early November, Houmo.AI announced the completion of its Series B funding round, with a single-round amount reaching several billion RMB. Over the past year, the company has completed multiple funding rounds. The simultaneous appearance of industrial capital, state-owned enterprises, and international investors this time appears more like a natural continuation following the accumulation of prior technical validation and commercial signals.
| A Track Selected Against Conventional Wisdom
Compute-in-Memory (CIM) is not a new concept; academic research in this field has spanned decades. Its fundamental logic is to break the data movement overhead caused by the "separation of computation and storage" under the von Neumann architecture—on the edge side, the energy consumption for moving data is often far higher than that for computing itself.
The industry's trajectory is interesting.
Over the past few years, most companies entering the CIM space chose low-compute, ultra-low-power scenarios: voice wake-up, wearables, and sensor endpoints. This route is commercially safe but has a visible ceiling. As large models became the protagonists of AI, the meaning of "edge AI" changed; only chips capable of running inference have pricing power.
When founded in 2020, Houmo.AI chose a different direction: using SRAM-CIM as the root technology to directly target high compute and edge-side large model inference. At the time, this was a contrarian choice—abandoning the "easy-to-implement small market" to tackle the "difficult but large market." According to company disclosures, adopting SRAM-CIM compressed their core computing unit area to approximately 30% of traditional architectures and reduced power consumption to about 10%. Whether these figures will be fully realized under real-world loads requires time to verify; however, the directional judgment is gaining acceptance in the industry. At the ecosystem partner conference held last month, less than a year after mass production began, the M50 chip had secured over 150 product design-ins and served more than 100 customers—a deployment speed rarely seen among startups at the same stage.
| Two Paths, One Metric
Edge workloads have varying requirements for capacity and bandwidth. Therefore, Houmo.AI diverged into two paths based on the same root technology: 2D CIM combines SRAM with memory configuration to solve the "capacity fit" issue; 3D CIM introduces SRAM-CIM into stacked logic computing layers, integrating with advanced packaging to solve the "speed" issue while addressing thermal challenges brought by stacking.
Both paths point to the same metric: Token/$ and Token/W—Token output per unit cost and per unit power. This framing is noteworthy. Most chip companies talk about TOPS, bandwidth, and process nodes, but Houmo.AI uses "how much intelligence can be produced per dollar spent and per watt consumed" as its North Star metric. This is a typical "buyer perspective": for equipment manufacturers, TOPS is a procurement parameter, whereas Token cost is an operational parameter.
Wu Qiang positions the company as the compute foundation for an "edge free Token factory." "Free" is certainly rhetorical—the true implication is that when edge Token costs are driven sufficiently low, they approach zero from the user's perception, allowing AI to embed into terminals like water or electricity. Whether this logic holds depends on whether the M50, the first 3D CIM chip A20, and the in-development M6X can continuously deliver energy efficiency advantages. Currently, the M50's 160 TOPS@INT8 and 10W typical power consumption enable it to cover localized execution needs for models ranging from 8B parameters to nearly 1000B parameters (the latter requiring multi-chip systems), making the product matrix steps clear.
| Software: The Underrated Second Battlefield
For CIM companies, the real test often lies not in the chip but in the software. New architectures imply new compilers, new operator libraries, and new adaptation costs—developers have no reason to pay extra learning costs for a non-mainstream path. This is the common cause behind the failure of various new architecture chips historically.
Houmo.AI's choice is to open-source the "Houmo.AI Ecosystem Software," covering the toolchain from quantization, compilation, and inference deployment to application development, with a plan to launch sequentially before January 2027. A pragmatic detail is that this toolchain largely stems from common issues accumulated across over 150 design-in projects—in other words, it is not a forward-looking plan but a product refined by real customer needs. Open-sourcing and standardizing adaptation experience essentially lowers partners' switching costs, which is also the most effective step for new architecture chip companies building ecosystems.
| Several Observations
Objectively speaking, Houmo.AI's narrative still has several aspects that require time to validate:
First, mass production scale and yield. Over 150 design-ins are order signals, but realizing stable large-scale delivery still requires passing supply chain and capacity tests, which is precisely the main use of funds from this round.
Second, the engineering complexity of 3D CIM. A20 entering production is a good signal, but heat dissipation, yield, and cost control in 3D stacking are industry-level challenges, and mass production ramp-up remains uncertain.
Third, the window for ecosystem competition. Major GPU players and large model vendors are extending towards the edge. Houmo.AI's energy efficiency advantage needs to maintain a sufficient gap to offset the inertial attraction of general-purpose ecosystems.
| Conclusion
Intensive capital injection indicates that the market has begun to believe one thing: the endgame of edge AI competition is not in the compute arms race, but in the cost-efficiency race. Houmo.AI took six years to bring a niche path to mass production with over 100 customers, which itself constitutes a sample significance for Chinese chip innovation. However, whether the "Token factory" vision can succeed is ultimately determined not by funding, but by next-generation products—the mass production performance of M6X and 3D CIM will be the most critical next page of this ledger.
Editor: Shuo Xin Yu-W²