Three rounds of financing secured within half a year, it has opened up a gap in Seedance's cloud business.
On-device AI has become a red-hot track in the primary market at present, with capital continuing to place heavy bets on it.
In the first half of 2026, upstream and downstream companies in sectors including on-device large models, dedicated chips, and on-device operating systems have successively secured financing at the level of billions of yuan, pushing the entire track to an unprecedented level of popularity.
However, another fact is that the industry as a whole is still in its early stage, with most projects remaining at the stage of technical demonstration and pre-installation on hardware, lacking killer products that can sustainably attract payment from a large number of users. The huge technical dividend has long failed to be translated into tangible commercial revenue.
How to break the deadlock, or what kind of new players the market needs, has suddenly become a topic of inquiry in the primary market. A few days ago, we met a young company — Silicon Token, whose ideas are bold yet pragmatic, trying to carve out a share of the cloud-based business with hardware.
At the beginning of 2026, Silicon Token was registered and established near the east gate of Tsinghua University. It is a hard-core technology startup focused on on-device AI infrastructure, and is currently promoting the R&D of its first desktop-level AI Station on-device hardware product.
Although the first product has not been launched yet, the team has already received real money-backed recognition in the primary market. The company was incubated with seed round investment from Zhen Fund. In just half a year after its establishment, it quickly completed two rounds of financing, and the new round of financing is also in progress. Its latest valuation has approached 2 billion yuan. Investors close to the project believe that its valuation is still in the left-side investment range, with sufficient room for growth.
Video Creators Have Suffered from Seedance for a Long Time
Before talking about Silicon Token, let's first look at the global AI market. There is no doubt that currently, the global AI companies that can achieve monetization mainly follow two paths: Coding and Video. Among them, the Coding track is a game for large manufacturers, where startups face slow financing, and "at most two or three players can stand out".
On the video side, cloud-based AI video generation has verified huge market demand. Massive creators of short dramas and comic dramas, video We-Media, e-commerce merchants, and marketing institutions in China and the United States have high-frequency rigid demands for content production.
At the same time, the cost of cloud computing power is also facing the dilemma of remaining high. The cloud price of Seedance is 1 yuan per second for 720P. In the actual creation process, repeated debugging and multiple regenerations are required, and a large number of wasted clips will further amplify the cost. Usually, the cost of generating 100 minutes of finished film is nearly 20,000 yuan.
Small and medium creators are under huge cost pressure for continuous production. In the early stage of the company's establishment, Bai Peng spent three months conducting research, and finally came to an intuitive conclusion: The Token cost of cloud video generation is too expensive, and video creators have suffered from Seedance for a long time!
At the beginning of this article, we mentioned that on-device AI is also facing difficulties, that is, the lack of scenarios that can realize commercial implementation. Therefore, in the view of the Silicon Token team, when on-device hardware meets AI video generation, a highly deterministic dividend track is thus opened.
In this regard, Bai Peng, Founder and CEO of Silicon Token, has a clear product goal: to develop an on-device AI hardware priced at 20,000 yuan, which uses consumer-grade graphics cards to run the world's top large models, so as to reduce the high cloud computing power cost for creators.
How to achieve this specifically?
According to Silicon Token, users can rely on AI Station to realize local video generation on the device. The calculation is completed inside the hardware, which greatly reduces the dependence on cloud computing power, avoids the high cost charged by the second, and at the same time supports offline availability, faster response, and local storage of materials.
Take MiniMax H3 as an example: the main body of the model is 33B, plus the text encoder, the full weight is 120GB. The official recommended deployment configuration is two H800s, worth 500,000 yuan. Looking at the 16GB video memory of consumer-grade graphics cards, there is a 7.5-fold gap when facing the 120GB weight.
In the past, the industry's solutions were highly consistent: quantization, distillation, resolution reduction, and duration reduction, to compress the degraded model into the hardware. Silicon Token does not compromise, and rewrites the entire inference system from scratch.
The video memory is like a warehouse, the entire model has to be moved in, otherwise it cannot run. Silicon Token transformed it into a conveyor belt: 120GB of weights flow through the 16GB video memory in a queue, and the next layer is replaced immediately after the calculation of one layer is completed. Moreover, which weights will be used in each step of the video model can be predicted. The system infers in advance the data that the GPU will need next, and delivers it before the GPU requests it, so that the data transmission time is completely hidden in the calculation process.
In this way, what determines the maximum size of the model that a machine can run is no longer the capacity of the video memory, but the scheduling capability of the system.
More importantly, on-device computing makes up for the cost shortboard of video generation, and video generation makes up for the most scarce payment scenario for on-device AI, thus forming a closed loop of business models including hardware sales, subscription, and commercial licensing.
Three Levels of Scarcity
Why Silicon Token? In fact, the track it cuts into is not a new demand. Perhaps the answer lies in the endowment of its founding team.
Bai Peng from Silicon Token has ten years of experience as a senior executive at Xiaomi Group. He has successively taken charge of many important businesses including Xiaomi TV, Internet Business, Commercial Department, and Government and Enterprise Department, reporting directly to Lei Jun. There is a landmark evaluation inside Xiaomi: he is the executive who is best at making money among product managers, and who has the best understanding of user experience among commercialization leaders.
In addition, the company currently has 21 engineers, many of whom are doctors and masters in Computer Science from Tsinghua University. Dr. Shen Youren, CTO of the company, and its core engineers all come from the Institute of High Performance Computing, Department of Computer Science, Tsinghua University. Since its establishment, the company has applied for 6 national invention patents around its core technology.
Such a team has also built the core moat of the company:
The first layer: The inference system is not an untapped field: vLLM and SGLang have pushed the inference of language models to the extreme, but they do not cover video. There are also acceleration efforts on the video side, but almost all of them are carried out in data centers — using more graphics cards to speed up generation, the premise of which is still that the model can be "loaded". There has been no system that sets the premise as "unable to load" and allows full-weight video models to run on consumer-grade hardware before — running 120GB with 16GB of video memory.
The second layer: The capability is interdisciplinary. Language models are sparse, hundreds of billions of parameters only use tens of billions at a time, which is the mainstream idea of "running large models on small graphics cards" in the past. Video models are dense, every layer and every step needs to use all parameters, no byte can be saved, and all the skills accumulated for language models are invalid, which must be developed from scratch. Researchers who understand diffusion algorithms and engineers who understand high-performance systems are originally two different groups of people.
The third layer: Time cannot be bought. The moat of system engineering does not lie in inspiration, but in accumulation — the life cycle of memory, cross-device scheduling, and every possible crash pit, must be experienced personally to be mastered. This kind of moat cannot be accelerated by financing, nor can it be taken away by poaching talents.
"Recreate All Hardware with AI"
In Silicon Token's BP, there is a page that writes "Apple in the AI Era". Although in my opinion, this statement is a bit old-fashioned. But from the underlying logic, the things the two parties do are no different.
In 1984, Apple put desktop computers in the hands of everyone. Apple's system could also run on IBM devices, but Apple chose to make its own hardware run first, to achieve the ultimate experience, and to build user awareness. The same is true for Silicon Token: its system can also run on hardware from Huawei and Lenovo, but it uses its own hardware for verification first.
Silicon Token's product path is very clear: first use one or two hardware products to verify the practicability of the platform, then open the platform to all hardware, and finally grow into the infrastructure for on-device AI. The real moat of Silicon Token lies in a self-developed inference engine base, an on-device AI platform that can run large models of text, image and video at the same time.
This on-device AI platform can be installed in all kinds of hardware: mobile phones, tablets, NAS, embodied robots, automobiles, and home appliances. "As long as it is equipped with an ARM chip or an AMD graphics card, people will think of us when they want to add AI capabilities to it." This technical architecture also determines the company's positioning — an AI infrastructure platform.
This also shows the ambition and vision of Silicon Token. The 1+N strategy refers to 1 core technology and N scenario applications.
Bai Peng put this vision into the most easily understood analogy in the Chinese business context: "Xiaomi's model is to recreate hardware with the Internet, and we are to recreate all hardware with AI." Facing the current market, he gave a very sober judgment: "American players are investing crazily regardless of ROI, while we have anchored an input-output model for the future in advance."
Up to now, the players who can survive through the cycle are never the ones who are best at telling grand stories, but those who can clearly calculate ROI. Using 20,000-yuan hardware to replace 500,000-yuan professional equipment, and helping video creators get rid of the high monthly cloud computing power bill of 20,000 to 30,000 yuan, this is a visible and calculable input-output account.
Peeling off the appearance, the core of Silicon Token is essentially a game of leverage. Using 20,000-yuan hardware to leverage enterprise-level computing power, using a lightweight hardware base to leverage a valuation of tens of billions of yuan, and using the middle-layer platform to leverage the vast market of AI-enabled hardware across all industries. The layers of leverage are interlocked: only when the technology is verified can the product have a solid foundation for implementation; only when the product is proven feasible can the platform have the soil for growth; only when the platform model is established can the 1+N large-scale expansion be realized.
This article is from WeChat official account "ChinaVenture", Author: Zan Zhu, published with authorization from 36Kr.