After the explosion of cloud tokens, edge computing power starts: The intelligent agent all-in-one machine welcomes six types of players.

Cloud AI witnessed an explosion in 2026, and "token" became the focus of the entire industry. However, players in the industrial chain gradually realized that AI won't only thrive in the cloud.

" The intelligent agent all - in - one machine has been quite popular. " Guo Mengming, the co - founder of Shoujie Technology, told Digital Intelligence Frontline. The industry has seen that six types of players, including almost all mainstream chip companies, traditional PC and hardware manufacturers, AI - native startups, vertical industry solution providers, general solution providers, and cross - border enterprises, have entered the game.

At the NVIDIA GTC Taipei Conference this week, Jensen Huang made it clear - "For the first time in 40 years, the PC will be completely reshaped." NVIDIA collaborated with MediaTek to develop the PC chip RTX Spark and entered the desktop computing power market, claiming that it can run a large model with 120 billion parameters locally with FP4 precision. This fall, Microsoft, Dell, and HP will also launch 40 devices.

Previously, people have been paying for cloud AI based on usage. Now, edge - side AI aims to transform cloud computing power into local infrastructure and enter consumers' bags and desks. Compared with cloud AI, edge - side AI is closer, more private, consumes fewer tokens, and is more suitable for individuals and small teams. The industry believes that the edge - side AI capabilities were at the primary or middle - school level last year and have reached the university level this year. As time goes by, the boundary between the capabilities of the edge side and the cloud will become increasingly blurred.

Computing power means revenue, and computing power means profit. "Manufacturers are doing everything possible to promote their edge - side AI hardware," said Ke Jiejing of Shenzhen Xingyi Technology. The competition has become fierce.

The intelligent agent all - in - one machine is here, and a new product category is taking shape

Behind this wave of enthusiasm is the rise of multi - agent applications. Li Kaifu, the CEO of Lingyiwanwu, introduced that a user request may be split into 20 or more agents running in parallel, and the results are aggregated to trigger the next round of collaboration. This computing model directly changes the requirements for hardware. "The hardware system must meet several conditions: local priority, edge - side processing, and a response delay of less than 100 milliseconds. In the future, ultimate token efficiency and local processing capabilities will be the key."

Guo Mengming analyzed that compared with cloud computing power, users have three driving forces for local computing power: privacy, token savings, and lowering the threshold for using agents. After the popularity of the Spring Festival "Xiaolongxia" (a certain application), the problems of high deployment thresholds and large token consumption were exposed. The all - in - one machine pre - installed with agents enables non - computer - major users to use it and significantly saves tokens.

What kind of machine is the intelligent agent all - in - one machine? There is still no unified definition in the market, and the product forms vary greatly. Ke Jiejing gave a relatively clear description: "It is a small host, a type of computer, with relatively strong local graphics card computing power." The core is to support a small and medium - sized team to deploy local models and use agents - "a small team of 3 to 5 people, or even an individual or an OPC (one - person company) can use it."

Currently, there are mainly four routes in the market, with different prices and scenarios:

The first is the Apple Mac mini. Relying on its M - series chips, it can be used as a daily computer and can also run small - parameter models and agents. After the popularity of OpenClaw, the price of the Mac Mini quickly rose from 2,900 yuan to 4,000 - 5,000 yuan.

The second is the NVIDIA series. The DGX Spark is priced at 33,000 yuan. It looks like a small box, has an ARM architecture, runs on the Linux system, and cannot be reinstalled with Windows, which is a bit of a threshold for ordinary consumers. It is more suitable for large companies to distribute to small departments. At the same time, the RTX Spark jointly launched by NVIDIA and MediaTek targets consumer - grade Windows PCs, priced from 18,000 to 25,000 yuan. In the fall of 2026, Microsoft, Dell, and HP will launch 40 models.

The third is the intelligent agent host based on AMD Ryzen AI Max+. It runs on the Windows system and has a maximum video memory of 128GB. Last year, the price of this type of machine was around 15,000 yuan, and this year it has risen to over 23,000 yuan due to the increase in memory prices.

The fourth is at the other extreme, using a low - power CPU like the Intel N97. It has 8G or 16G of memory, does not run models, and only provides an independent space for agent operation, requiring additional consumption of cloud tokens. The price is a few thousand yuan.

Except for the Mac Mini, intelligent agent all - in - one machines usually integrate an agent platform, featuring "out - of - the - box usability". NVIDIA has officially entered the edge - side AI host market, competing with AMD and Intel.

It is worth noting that the intelligent agent all - in - one machine or edge - side AI is different from the "large - model all - in - one machine" or "DeepSeek all - in - one machine" that was popular last year. Cai Youquan of CBMIC introduced the difference between the two: the large - model all - in - one machine solves the problem of "large - model deployment computing power" and is an AI server, usually with 8 graphics cards and a price in the millions. However, most of them can only do Q&A and are targeted at the production environments of governments and enterprises. The intelligent agent all - in - one machine or edge - side AI solves the problem of AI application implementation, truly helping enterprises or individuals with their work. Ke Jiejing added that in essence, it is first a personal computer with strong hardware capabilities, and secondly, it can run models and agents and provide AI empowerment.

A review of the two "return tides": What's different this time?

In fact, the AI hardware track has experienced two collective impulses and two ebb tides.

The first was in early 2025. The DeepSeek open - source model became extremely popular, and the DS all - in - one machine saw a rush to buy. However, there were subsequent idle and return situations. Although the large - model all - in - one machine is not edge - side computing power, it is also in the form of an all - in - one machine. Industry insiders believe during the review that at that time, there was computing power but no applications. "For large - model all - in - one machines, most suppliers are hardware enterprises, and the products do not have applications or only have basic Q&A. Customers blindly launched new technologies in a confused state, made large investments, and finally found them to be of no value."

The second was around the Spring Festival in 2026. The explosion of OpenClaw made the MacMini popular. This time, there were applications, but they were not put to use. "People just said that this thing seemed very popular and asked someone to install 'Xiaolongxia' (a certain application), but they didn't have specific needs for AI to help with work." Wei Yang of Beilinsi said: "Whether it's OpenClaw, various 'Xiaolongxia', or Hermes, their significance for ordinary individual users is not as great as expected. Buying a machine alone won't work; more ecological adaptation is needed." Users followed the trend to buy and finally had to return the products.

After two setbacks, what will happen this time?

Interestingly, Guo Mengming's judgment is the opposite of the timeline. He was pessimistic during the Spring Festival "Xiaolongxia" craze; now he is optimistic. "Previously, it could be seen from the long queues to install 'Xiaolongxia' that if users had obstacles in the most basic capabilities, it would be very difficult to use them." But now the situation is different. "Agents are much more capable and mature in application than last year." Many users directly look at the results when buying and truly use them to solve practical problems. This part of the market is stable. Of course, there are still users who follow the trend. "The market still needs some time to be educated." "I believe there won't be such a large return tide as last year."

Wei Yang revealed that the company won't blindly produce and purchase all - in - one machines but will allocate them according to demand. Due to the sharp increase in memory prices, everyone is waiting and watching. At the same time, many users are researching agents and AI, and it will take some time for them to be truly applied in actual work. "The overall growth may not reach a high level until next year."

Ke Jiejing admitted that it is currently a painful period. The demand is real, but after the increase in memory and CPU prices, the cost has become very high, discouraging many users. For this reason, they have made some new arrangements in the direction of localization. "Conducting localization adaptation, although the cost won't be low, can better meet users' localization needs." Ke Jiejing observed that "most ordinary users' understanding of AI is still limited to new models, using AI to check the weather, or fortune - telling, which are entertainment scenarios. "Without generating actual value, there will be so - called bubbles."

Is the new product category "neither here nor there"? Where is the real demand?

During the process of visiting customers, industry insiders felt the customers' concerns. Edge - side AI or the intelligent agent all - in - one machine seems to be in a situation of being "neither here nor there" - in terms of scenarios, the edge side currently has fewer than the cloud; in terms of computing power, it only has one graphics card and has limited capabilities; in terms of price, it is relatively high for ordinary users, especially students.

"Customers think it doesn't make economic sense," an industry insider told Digital Intelligence Frontline. "But it's different from a server. There is no software adaptation on the server, and it can't be used out of the box. Hiring a third - party software company to develop will cost tens of thousands to hundreds of thousands of yuan. While using the intelligent agent all - in - one machine, it can be used out of the box and can support a small and medium - sized team to use local models."

So, who will really pay for it? Industry insiders have identified two typical groups: small teams sensitive to data privacy and the so - called "super individuals" or OPC (one - person companies) that everyone has been talking about this year.

Wei Yang told Digital Intelligence Frontline that small teams of less than 10 people care about data privacy. "Buying a device with 128GB of memory will be more convenient." For example, research teams in fields such as biomedicine and archaeology, due to privacy and compliance requirements, often have large local databases, but searching and organizing them is extremely time - consuming. Their agents will automatically search and summarize literature, and even regularly compare research directions and actively remind when similar directions are found to avoid duplicate work. When giving advice to a friend in the water conservancy industry, he mentioned a principle: "If the data volume is large, it is recommended to deploy a set of models locally; if the data can be processed manually, there is no need to use edge - side AI."

Ke Jiejing introduced another type of customer: some enterprises don't want to expose their solution codes to cloud - based competitors. They need local deployment but don't want to build an expensive computer room all at once. The intelligent agent all - in - one machine has become an alternative. The same goes for the AI customer service in 4S stores. "The internal customer information is private, and they don't want other 4S stores to poach customers."

Another type of demand comes from super individuals. Li Kaifu proposed that in the future, there will be a large number of DRI (Directly Responsible Individuals) in AI companies. They are responsible for the business results, conduct overall planning, make key decisions, and are responsible for the final output contract. "A human DRI is at the center of the entire agent system. Surrounding him and working in collaboration are specialized clusters composed of different agents for research, execution, compliance, and monitoring." The edge - side AI device is the exclusive infrastructure for this new type of worker.

Guo Mengming added that many lawyers they have contacted are already using the all - in - one machine. "They can accept products priced at 20,000 - 30,000 yuan." Similar high - net - worth individuals in knowledge - intensive fields such as traditional Chinese medicine also have relatively high demand. These users need a local platform that can long - term precipitate personal knowledge bases, protect privacy, and continuously run agent assistants.

Cai Youquan gave a self - judgment standard for buyers: "Repeated tasks done every day are suitable to be handed over to agents. The more specific the tasks are, the more capable the agents will be." For example, parts salespeople need to send price quotes to dozens or hundreds of customer groups every day, and accountants need to check invoices for tax calculation every day. These very mechanical and repetitive tasks can be completely handed over to AI, but most people don't have this awareness yet.

"I can really understand the customers' current concerns," Guo Mengming said. The edge - side AI capabilities are going through an evolution similar to the early days of smartphones. Last year, it was at the primary or middle - school level, and this year it has reached the university level. By the end of the year, there will be more new models and useful agent frameworks, and there will be more application scenarios. "I believe that as time goes by, the boundary between the capabilities of the edge side and the cloud will become increasingly blurred." The hardware itself is already very powerful, and the current bottleneck lies in the ecosystem. "The industry is leveraging their respective capabilities to jointly improve the ecosystem."

Six types of players enter the market, and software is the dividing line

Edge - side AI computing power is becoming a battleground for all. Roughly speaking, at least six types of players have entered the game.

Traditional PC and hardware manufacturers - established enterprises such as Dell, HP, Lenovo, and ASUS, driven by the supply chain and with a focus on distribution channels; solution companies, which are mainly B - to - B service providers, deeply involved in the industry, and the integration of software and hardware is conducive to delivery; AI - native startups, such as Lingyiwanwu and Jieyuexingchen, whose core assets are models and multi - agent frameworks, and the integration of software and hardware can more easily form a business closed - loop; general - purpose companies in the industry, such as Shoujie Technology, which focus on general needs, community - based operations, and screen user feedback into standardized products; chip companies, although their main business is not all - in - one machines, are also willing to deeply integrate hardware, software, and AI capabilities with partners to jointly launch solutions for specific fields; and cross - border players such as Dreame, which are seizing the opportunity to open up new business tracks.

Currently, due to the increase in supply chain prices, ODM and OEM manufacturers have reaped the first - wave benefits, while other types of enterprises are still on the way.

Alliances and cooperation are also underway. Ke Jiejing introduced that they are collaborating with domestic chip manufacturers to adapt a type of hardware product; Guo Mengming revealed that the company is collaborating with traditional hardware manufacturers for platform pre - installation. "The laptops released by Colorful in June will be directly pre - installed with this platform."

Cai Youquan observed that this round is different from the previous large - model all - in - one machines. "Many large - model all - in - one machines were mainly hardware - based enterprises. This time, software enterprises are taking the lead in the intelligent agent all - in - one machine market." The reason is that customers are more concerned about the actual benefits generated, and the logic is closer to finding users in the C - to - B market.

Guo Mengming told Digital Intelligence Frontline: "What users are more concerned about is whether the all - in - one machine has more abundant AI capabilities, which is the most important thing when people buy hardware now." Shoujie Technology is focusing on the software level and currently solving some inconspicuous basic problems, such as the deployment environment, various agent applications, and general tools like data desensitization and basic translation. "First, improve the toolchain. After the toolchain is formed, a large number of application systems can be developed on this basis." The company also screens general needs through community operations for development. He described this approach as the "Xiaomi model."

Cai Youquan introduced that their main business is B - to - B. "We mainly promote enterprise knowledge bases. When enterprises adopt AI, the first step should be to establish a knowledge base, rather than directly deploying a large model." The value of the knowledge base is to first AI - enable the enterprise's data so that AI can understand the enterprise's data, and then the data can generate value. Putting the agent in a box is to reduce the deployment difficulty for customers.

The focus of Jiang