Source of this article: Tiger Sniff Author: Fang Xiaonan

Since the beginning of 2025 years, the technology upstarts represented by DeepSeek and Unitree Technology are reshaping the competitive landscape of the AI circle, surging and fierce battles, and for those AI players who have broken the waves in the previous era, how to squeeze into the new wave is particularly urgent.

4月10日下午，上海西岸，一艘又一艘的船駛過黃埔江，風起時，沿著滾滾江水急速順流而下。與此同時，位於岸邊的西岸大劇院內，上一代老牌AI選手代表——商湯科技，正在舉辦“商湯2025技術交流日”活動。活動現場，商湯發佈了新版本“商湯日日新 SenseNova V6”多模態融合大模型，這是一款具備最長64K思維鏈、數理分析、多模態深度推理、全域記憶等能力的多模態大模型，而多模態正是2024年以來商湯主要的發力方向之一。同時，商湯還推出了“商湯大裝置 SenseCore 2.0”，這是自2021年大裝置發佈以來首次全新升級。（虎嗅注：商湯大裝置，即提供包括底層計算基礎設施服務IaaS、深度學習平台服務PaaS，以及模型部署及推理服務MaaS在內的三大服務。簡單來說就是它通過整合算力、演演算法、數據，提供一個基礎設施解決方案，以降低模型開發和應用的門檻。）

From the point of view of time, this is the first technical exchange after SenseTime lost weight and lost weight, transformed and reorganized.

Looking back on the past nearly a year, SenseTime has been very busy.

As an AI-native company, SenseTime has struggled in this new round of AI competition in the last wave of AI with the help of computer vision recognition. Under the new narrative mode, how to get the ticket of the new era, SenseTime has also been thinking, exploring, and making a lot of attempts.

In the first 2023 years, SenseTime began to shift its business focus to generative AI, and restructured its business into three sectors: generative AI, traditional AI and smart cars.

去年10月，商湯科技正式發佈“大型設備-大模型-應用”三位一體戰略，並同步構建更為集中和高效的組織結構。

A month later, Xu Li, chairman and CEO of SenseTime, lost weight and slimmed down again, reiterating the importance of generative AI, establishing a "1+X" architecture, and restructuring its business into generative AI, smart cars, and vision AI.

而商湯科技的生成式AI相關業務，包括提供算力的“大型設備”、AI基礎模型和應用。其中，商湯的大裝置是在2021年底投入使用，這也是它區別於其他AI公司的一大重要特點，其實在過去很長一段時間里，AI公司是否需要自建基礎設施都是一個非共識，而商湯也因此走出了一條獨特的商業道路。

According to Yang Fan, co-founder of SenseTime and president of the large device business group, relying on SenseTime's large device, the cost of SenseTime's multimodal inference application has been greatly reduced. There is no doubt that in the new wave of AI, the technical iteration of large models and the application of scenarios are inseparable from the support of a good AI Infra. (Tiger sniffing note: AI Infra will infrastructure the three elements of AI, "computing power, algorithms, and data", and provide a full set of standardized tool chains for the entire production cycle of models, so as to reduce the cost of model development and application and improve efficiency.) ）

Tiger Sniff communicated with Yang Fan, and he shared his thoughts on the path of SenseTime's large-scale device technology development and iteration. The following is a transcript of the exchange, with some of the contents slightly abridged:

Tiger Sniff: At present, in the area of computing infrastructure, some large manufacturers are also investing hundreds of billions of yuan, so what does SenseTime want to do in such a competitive landscape?

Yang Fan:The advantage of SenseTime is that we understand the model and the customer. In recent years, a very strong feeling is that we began to put forward the proportion of light and heavy assets from the second half of 3 to the beginning of 0 years, and we want to increase the operation of light assets, the reason for this is obvious, that is, the current chip hardware is about 0 years of iteration, and the downstream user's model is about three months of iteration, front-line users must be looking forward to using newer hardware resources to iterate the latest technology.

From this point of view, it can be seen that the lightweight of assets is actually unburdened to customers, so we are trying to increase the proportion of light assets.

The change on the supply side is that the technology and application iteration speed of the downstream demand-side model is very fast, and the upstream supply chain, whether it is NVIDIA or domestic chips, is also in a high-speed change from production capacity to cost to market supply level.

When the market, especially the upstream and downstream, is changing at the same time, maintain the flexibility of the system, and make clearer portrait descriptions for different types of customers, and provide customers with a solution that is more in line with their needs.

I think this is more important than the size of the assets as a whole.

Tiger Sniff: SenseTime's big device has been done since 2021 years, and there have been some upgrades in the middle, but why is such a large upgrade being carried out at this point in time this year?

Yang Fan:This is indeed because we have made a relatively large upgrade in the system architecture.

We started doing it in 2021 years, but the real scale of the market was after the generation of AI. At that time, the state was that there were no particularly good and mature products and solutions in the entire market, including SenseTime itself, although some kilocalorie experiments were done very early, there would still be many problems in the service process, because the chip was also iterating, and there would be many unexpected problems when it really ran in practical scenarios. Therefore, when we do it at that point, it is more about iterating product functions according to the needs of users.

Today, the situation is very different, and the whole market trend has changed a lot in the past two years.

Tiger Sniff: What are the specific changes? In what ways?

Yang Fan:I think there are two main points, the first is that open source will become more and more mainstream, or more and more influential. This kind of open source is not only the open source of the model, but also the emergence of open source components of various middleware and tools.

So when we see this becoming a trend, or we saw this trend a year ago before the market, an important decision we made was to make the framework of the whole system have a better converged system, which is very important for the compatibility of the whole open source system.

The second is that more and more AI training and pre-production links have entered the application link, which has introduced many new challenges for Infra, such as some elasticity, rapid deployment, and performance optimization. In fact, we have done this kind of thing before, but it will always be like this, technology iteration only happens when you have enough users, that is, when his business scale and number of users rise, you will find that he will put forward more requirements for you, and these requirements will guide your core technology to iterate.

In fact, our SenseCore 0.0 was started more than half a year ago, and we saw these two trends at that point in time, and thought that the entire market in the next three years would be very different from the past three years, so our entire product system would have to make major changes to this trend.

Tiger Sniff: In fact, for a long time, there was a non-consensus on whether AI companies should build their own computing infrastructure, and SenseTime was one of the more typical cases, so do you think this is still a non-consensus? Or what kind of consensus is reached?

Yang Fan:The consensus is that you shouldn't build your own (laughs), and only big enough vendors will build your own. Or the more important question is how to define an AI company, for example, SenseTime was established for 10 years, and in the past, everyone had a lot of things that did well and a lot of things that did not do well.

Tiger Sniff: How do you think AI companies should be defined, or what should the next generation of healthier AI companies look like?

Yang Fan:I think it should be more relevant to a certain industry and a certain scenario, like some of our customers, because only in this way can his business model be easier to determine, so as to solve the problem of profitability to a greater extent, which is a challenge for most companies.

AI is a general-purpose technology, so if the definition of AI companies is from the perspective of commercialization in the long run, it is healthier to have more, smaller, and more focused on a certain segment to provide products and services with AI as a key capability, and there are a large number of such companies, which is what the entire industry should look like in five or ten years.

And for these people, there is no doubt that they do not have the need or strong demand to build their own infrastructure, because their more important thing is to provide a better product and service to their customers.

AI may be a key capability in this process, but it is not necessarily his starting point, and for the technical iteration of AI itself, such as the iteration of models and the accumulation of data, their demand for resources is constantly growing, so the more this is the case, the more he should use third-party services to reduce costs, because this is not his core.

His core lies in how to embed this little bit of innovation or superior capabilities of AI into his products and solutions, which will be the mainstream of the market in the future.

Of course, for some leading manufacturers, including some leading innovation companies, they will indeed be more inclined to build themselves, because the scale is large enough to bring enough (revenue) in this matter.

From the perspective of China's future economic development, the overall need to transform to high-quality development, the core grasp lies in scientific and technological innovation, if the most important and effective element in the middle is AI, then in five or ten years, there should be 99 real AI companies in 0 companies in the market, then our goal is to serve these companies.

Tiger Sniff: Last year, SenseTime proposed the trinity strategy of "large equipment, large model, and application", so it can be explained in combination with this product update and upgrade, how are they linked to each other?

Yang Fan:First of all, they are mutually reinforcing.

For example, from the perspective of large devices, if I want to provide a good AI Infra, then I need to know what the needs of users are, and in SenseTime's own model research iteration, as mentioned earlier, it will put forward various requirements for you, and in this kind of demand, if many things are standardized, they will become market-oriented things.

The same is true for applications, models may be more involved in model training, R&D, and production, and the use stage of the model is actually closely related to the application.

For example, when we explore a new subdivision scenario, Wensheng Diagram, we will first use SenseTime's own Wensheng Diagram product to see how to provide customers with a better service based on their data and changes in call volume.

In turn, better AI Infra will definitely support (our own) model technology research and development, measurement and application scenario, etc.

Tiger Sniff: From the perspective of SenseTime's overall AGI goals, what is the strategic significance of the large installation?

Yang Fan:First of all, our cognition and know-how of models and applications can actually help large devices form better differentiation advantages, and the performance of this differentiated advantage in the market is to make more money, because we all know that the investment in model research and development is very huge, and the better the profit of the large device, the more it can support SenseTime's overall continuous technology development and iteration, which is certain.

On the other hand, from a purely technical point of view, the iteration of the model also needs a good Infra cooperation and support. In the process of model training, we sometimes need some targeted, even customized Infra improvements, and only when you have a larger platform will respond better to such improvements.

Nowadays, the application of large models, I think many scenarios do not have the life and death line of cost performance, so you will see a lot of applications, one of which is to spend money to pull traffic, because it may be C-side-oriented, and it can afford to spend money, and does not pursue cost performance and computing power.

But if you go back to the B-end service market, you will find that many of today's applications are still in the early iteration stage, and the technology may reach 80 or 0 points, but the cost performance has not yet achieved a relatively good state.

In fact, a very important value for us to make a big device is that we can provide customers with good cost performance, which of course can also provide good cost-effective support for SenseTime's own model application, which will have a great guarantee for its competitiveness in the B-end market in the future.