Core Summary
Houmo Intelligence is a company that develops “integrated storage and computing” chips. By using a groundbreaking technology, it enables small devices the size of a palm (such as AI hosts or robots) to run large models with billions of parameters at a low power consumption of just 10 watts, solving the problem of excessive time and energy consumption associated with data transfer in AI computations. These chips transform local devices into “free Token factories” – once the hardware is purchased, users can utilize AI indefinitely without having to pay per usage, unlike in the cloud. The surge in large models has turned this technology from an academic concept into a commercial reality, with clients including Lenovo and Great Wall, among others. In the future, these chips could become key players in the shift of AI computing power from the cloud to local devices.
Detailed Analysis
1. Integrated Storage and Computing: Solving the “Traffic Jam” in AI Computation
Traditional chips operate like luxury cars transporting goods; data must first be retrieved from storage (like picking up cargo at a warehouse) before being processed by the computing unit (like driving on a racetrack). When the amount of data is massive, it’s akin to hundreds of thousands of vehicles stuck in traffic every second, resulting in extremely low efficiency. NVIDIA GPUs have expanded the “highways” (by increasing bandwidth), but data still needs to be transferred. Integrated storage and computing, on the other hand, “builds the warehouse next to the racetrack,” allowing computations to occur directly within the storage, eliminating the need for data transfer and saving time and energy.
For example, Houmo’s chips can run large models with billions of parameters at 10 watts of power, enabling devices to function as if they had a “superbrain” using just the power required to charge a smartphone, whereas conventional solutions would require much higher power and be more energy-intensive. This is a critical requirement for local devices (such as PCs and robots); after all, we don’t want home AI devices to consume as much power as servers.
2. Large Models: Turning Integrated Storage and Computing from a Niche Technology into a Hotcommodity
Previously, integrated storage and computing technology was only suitable for low-computing tasks (such as simple AI in earbuds) or for niche applications in computer vision (CV), where established players held a dominant position. However, the explosion of large models has changed the landscape:
- Exponential Data Growth: Large models process much more data than CV systems, highlighting the advantage of not having to transfer data.
- New Market Opportunities: There is now a demand for running large models on local devices (e.g., AI hosts or robots with privacy-sensitive applications), where there are no established players.
- Increasing Customer Interest: After 2023, companies like Lenovo and Great Wall sought devices capable of running large models, and Houmo’s chips met their needs, leading to orders exceeding expectations. Large models have provided an opportunity for integrated storage and computing technology to move from the laboratory to the market.
3. Local “Free Token Factories”: Changing the AI Pricing Model
Cloud-based AI services are priced on a pay-per-use basis (e.g., paying per token processed), similar to mobile data usage. In contrast, local integrated storage and computing chips offer a one-time purchase that allows unlimited use of AI without additional costs. This brings two benefits:
- Cost Savings: For example, with cloud-based services for companion robots, the more users use them, the higher the manufacturer’s costs (due to fees to the cloud). With local chips, manufacturers only need to sell the hardware, and users don’t incur extra expenses.
- Privacy Security: Data is processed locally, preventing privacy breaches from cameras or voice recordings (e.g., in home AI devices, no one wants their personal data to be transmitted to servers).
In the future, we may see a combination of local and cloud services: simple tasks will be handled locally (for free), while complex ones will be processed in the cloud (for a fee), with local solutions becoming the mainstream choice.
4. Houmo’s Advantages: A “Curver-Catcher” without Historical Burdens
Compared to giants like NVIDIA, Houmo has two key advantages:
- Starting from the Same Groundline: Integrated storage and computing is still an emerging technology, while NVIDIA relies on traditional GPUs for revenue and is reluctant to abandon its existing business models. Houmo started from scratch with a fully integrated approach, free from historical constraints.
- Cost-Effectiveness: With domestically produced chips and open-source models, although their absolute performance may not match the best overseas offerings, the lower cost makes them attractive to overseas customers in regions with poor internet connectivity or high privacy concerns (e.g., rural areas where users can benefit from local Token factories without relying on expensive data plans).
However, challenges exist: As the company grows, organizational management will become more complex, and new startups may emerge with even more disruptive technologies.
5. When Will These Chips Become Common in Every Home?
Two conditions are needed for widespread adoption:
- Cost Reduction: As production scales up, hardware costs will decrease (similar to how smartphone prices have dropped from thousands to hundreds of yuan).
- Innovative Killer Applications: There needs to be a compelling reason for consumers to buy these chips, such as a “home AI center” that can connect all smart devices, provide real-time translations, assist with learning, and control household appliances without the need for internet connectivity or additional fees.
Once these conditions are met, local AI solutions will become as common as routers, with every home equipped with its own “local AI brain.”
Conclusion
Houmo Intelligence’s story represents the shift of AI computing power from being centralized in the cloud to being distributed locally. By using integrated storage and computing technology, they have addressed the challenges associated with running large models on local devices, making AI more affordable, private, and user-friendly. In the future, when “free Token factories” become ubiquitous, our lives could be profoundly transformed by local AI services that are always available without unnecessary expenses. This is not only a technological victory but also a revolutionary change in business models.