AI agent all-in-one machines are taking center stage, prompting a collective shift across the industry chain.

Cloud AI experienced explosive growth in 2026, with 'tokens' becoming the focal point of the entire industry. However, players across the industry chain are gradually realizing that AI will not remain confined to the cloud.

'AI agent all-in-one machines are already generating significant buzz,' Guo Mengming, co-founder of Shoujie Technology, told DigitFront. Nearly all mainstream chip companies, traditional PC and hardware manufacturers, AI-native startups, vertical industry solution providers, general-purpose solution vendors, and even cross-industry players have entered the market.

At NVIDIA's GTC Taipei event this week, Jensen Huang made it clear: 'For the first time in 40 years, the PC is about to be completely reshaped.' NVIDIA, in collaboration with MediaTek, has developed the RTX Spark PC chip, entering the desktop computing market with claims that it can run a 120 billion-parameter large model locally using FP4 precision. This fall, Microsoft, Dell, and HP will launch 40 new devices.

Previously, everyone was paying for cloud AI on a usage basis. Now, edge AI aims to transform cloud computing power into local infrastructure, making its way into consumers' bags and onto office desks. Compared to the cloud, edge AI is closer, more private, consumes fewer tokens, and is better suited for individuals and small teams. The industry believes that edge AI capabilities have evolved from elementary school or middle school level last year to college level this year. Over time, the distinction between edge and cloud capabilities will continue to blur.

Computing power equals revenue, and computing power equals profit. 'Vendors are doing everything they can to promote their AI edge hardware,' said Ke Jiejing of Shenzhen Xingyi Technology, noting that competition is heating up.

01 The Arrival of AI Agent All-in-One Machines: A New Category Takes Shape

Behind this surge is the rise of multi-agent applications. Li Kaifu, CEO of 01.AI, explained that a single user request might be split into 20 or more agents running in parallel, with results aggregated to trigger the next round of collaboration. This computational model directly changes hardware requirements. 'The hardware system must meet several conditions: local-first processing, edge-side execution, and response delays under 100 milliseconds. In the future, extreme token efficiency and localized processing capabilities will be key.'

Guo Mengming analyzed that compared to cloud computing power, users have three key drivers for local computing power: privacy, token savings, and lower barriers to using agents. After the crayfish craze during the Spring Festival, issues like high deployment thresholds and significant token consumption came to the forefront. Pre-installed agent all-in-one machines allow non-computer science users to get started easily while substantially reducing token usage.

What exactly is an AI agent all-in-one machine? The market has yet to agree on a unified definition, and product forms vary widely. Ke Jiejing provided a relatively clear description: 'It's a small host, a type of computer with relatively strong local graphics card computing power.' The core is supporting small to medium-sized teams in deploying local models and using agents—'for teams of 3-5 people, or even individuals and one-person companies (OPCs).'

Currently, there are four main approaches in the market, each with different price points and use cases:

The first is the Apple Mac mini, which relies on its M-series chips. It can function as a daily computer while also running small-parameter models and agents. After the OpenClaw craze, Mac Mini prices quickly rose from 2,900 yuan to 4,000-5,000 yuan.

The second is NVIDIA's lineup, including the DGX Spark priced at 33,000 yuan. It comes in a small box with an ARM architecture running Linux, unable to reinstall Windows, posing a barrier for ordinary consumers. It's more suitable for large companies distributing to small departments. Meanwhile, NVIDIA's RTX Spark, developed with MediaTek, targets consumer-grade Windows PCs, priced between 18,000-25,000 yuan, with 40 models set to be released by Microsoft, Dell, and HP in fall 2026.

The third approach involves agent hosts based on AMD's Ryzen AI Max+. Running Windows with up to 128GB of memory, these machines cost around 15,000 yuan last year but have risen to over 23,000 yuan this year due to memory price hikes.

The fourth approach represents the other extreme, using low-power CPUs like Intel's N97 with 8GB or 16GB of memory. These don't run models but provide an independent space for agent operation, requiring additional cloud token consumption and priced in the thousands of yuan.

Except for the Mac Mini, AI agent all-in-one machines typically integrate agent platforms, emphasizing 'out-of-the-box usability.' NVIDIA's formal entry into the edge AI host market puts it in competition with AMD and Intel.

It's worth noting that AI agent all-in-one machines or edge AI are different from last year's popular 'large model all-in-one machines' or 'DeepSeek all-in-one machines.' Cai Youquan from China National Building Material Information explained the distinction: large model all-in-one machines address 'large model deployment computing power' issues, functioning as AI servers typically equipped with eight graphics cards and priced in the millions of yuan. However, most can only handle Q&A and are aimed at government and enterprise production environments. AI agent all-in-one machines or edge AI, on the other hand, focus on AI application implementation, truly helping businesses or individuals get work done. Ke Jiejing added that they are essentially powerful personal computers capable of running models and agents for AI empowerment.

02 Looking Back at Two 'Return Waves': What's Different This Time?

In reality, the AI hardware sector has already experienced two waves of collective enthusiasm followed by two retreats.

The first occurred in early 2025 when the DeepSeek open-source model went viral, leading to a rush for DS all-in-one machines, which were later idle and returned. Although large model all-in-one machines weren't edge computing, their form factor was similar. Industry insiders reviewing the situation believe there was computing power but no applications. 'Large model all-in-one machines were mostly supplied by hardware companies, with products lacking applications or only featuring basic Q&A. Customers, blinded by new technology, went online without understanding its value, investing heavily only to find it useless.'

The second wave came around the 2026 Spring Festival when OpenClaw went viral, boosting Mac Mini sales. This time, there were applications, but they weren't effectively utilized. 'Everyone just thought it seemed popular and had someone install crayfish software, but they didn't have specific needs for AI to help with work,' said Wei Yang from Belins. 'Whether it's OpenClaw, various crayfish software, or Hermes, their significance for ordinary individual users isn't as great as imagined. Buying the machine alone isn't enough; ecosystem adaptation is needed.' Users followed the trend and ultimately had to return their purchases.

After two failures, what will happen this time?

Interestingly, Guo Mengming's judgment runs counter to the timeline. He was pessimistic during the crayfish craze but is now optimistic. 'Previously, seeing everyone line up to install crayfish software showed that if users struggled with basic capabilities, it would be hard to use effectively.' But now, things are different. 'Agents are much more capable and mature than last year. Many users buy based on results, using them to solve real problems, creating a stable market.' Of course, trend-following users still exist, 'and the market still needs time to educate them.' 'I believe we won't see as large a return wave as last year.'

Wei Yang revealed that his company won't blindly produce or purchase all-in-one machines but will allocate based on demand. Due to sharp memory price increases, everyone is waiting to see what happens. Meanwhile, many users are studying agents and AI, but actual implementation in the workplace will take time to ferment. 'Overall, significant growth might not happen until next year.'

Ke Jiejing admitted that the current period is painful. Demand is real, but rising memory and CPU costs have made prices prohibitive for many users. In response, they're making new layout (strategic moves) in localization. 'Adapting to local needs won't necessarily lower costs but can better meet users' localization requirements.' He observed that 'most ordinary users still see AI as just new models for asking about the weather or fortune-telling—entertainment scenarios. Without real value, there will be foam (bubbles).'

03 The New Category's Dilemma: Where Is the Real Demand?

While visiting clients, industry insiders sense their concerns. Edge AI or AI agent all-in-one machines seem stuck in a 'neither high nor low' position—in terms of scenarios, edge currently offers fewer than the cloud; in computing power, it's limited to a single card; and in price, it's relatively high for ordinary users, especially students.

'Clients feel the math doesn't add up,' one industry insider told DigitFront. 'But it's not the same as servers. Servers lack software adaptation and can't be used out-of-the-box. Developing custom software requires tens to hundreds of thousands in investment. With agent all-in-one machines, you can support small to medium-sized teams in using local models right away.'

So, who will actually pay for them? Industry insiders identify two typical groups: small teams sensitive to data privacy and the 'super individuals' or OPCs (one-person companies) everyone's talking about this year.

Wei Yang told DigitFront that small teams of under 10 people care about data privacy. 'Buying a 128GB memory device is convenient.' For example, scientific research teams in biomedicine or archaeology often have vast local databases due to privacy compliance requirements, but searching and organizing them is extremely time-consuming. Their agents can automatically retrieve and summarize literature, even periodically compare research directions and issue alerts to avoid duplicate work. When advising a friend in the water conservancy industry, he suggested a principle: 'If there's a large volume of data, recommend local model deployment; if a person can handle it, edge AI isn't necessary.'

Ke Jiejing described another type of client: companies that don't want to expose their solution codes to cloud competitors. They prefer local deployment without building expensive server rooms, making agent all-in-one machines an alternative. Similarly, 4S store AI customer service 'has private internal client data and doesn't want other 4S stores poaching customers.'

Another demand comes from super individuals. Li Kaifu proposed that future AI companies will have many DRIs (Directly Responsible Individuals) who take responsibility for business outcomes, overall orchestration, key decisions, and final output contracts. 'A human DRI sits at the center of the entire agent system, surrounded by specialized clusters of research, execution, compliance, and monitoring agents.' Edge AI devices provide the Exclusive infrastructure (dedicated infrastructure) for these new types of workers.

Guo Mengming added that they've seen many lawyers adopting all-in-one machines. 'They can accept products priced at 20,000-30,000 yuan.' Similarly, knowledge-intensive, high-net-worth groups like traditional Chinese medicine practitioners have significant demand. These users need a local platform that can accumulate personal knowledge bases over the long term, protect privacy, and continuously run agent assistants.

Cai Youquan suggested a self-assessment standard for buyers: 'Any repetitive daily task is suitable for agents. The more specific the task, the better agents can handle it.' For example, parts salespeople who send quotes to dozens or hundreds of client groups daily, or finance staff who review invoices and calculate taxes, could fully automate these mechanical tasks with AI, but most people haven't realized it yet.

'I fully understand clients' concerns now,' Guo Mengming said. Edge AI capabilities are evolving like smartphones in their early days—last year, they were at elementary or middle school level; this year, they've reached college level. More new models and user-friendly agent frameworks will emerge by year-end, expanding application scenarios. 'I believe that over time, the distinction between edge and cloud capabilities will continue to blur.' The hardware is already powerful; the current bottleneck lies in the ecosystem. 'The industry is leveraging its strengths to jointly improve the ecosystem.'

04 Six Key Players Enter the Fray: Software Becomes the Differentiator

Edge AI computing power is becoming a battleground, with at least six types of players entering the market.

Traditional PC and hardware manufacturers—Dell, HP, Lenovo, ASUS, and other established companies driven by supply chains and channel dominance; solution providers, B2B-focused vendors that delve into industries, where integrated hardware and software aid delivery; AI-native startups like 01.AI and Jieyue Xingchen, whose core assets are models and multi-agent frameworks, making integrated hardware and software easier to form commercial closures; general-purpose industry players like Shoujie Technology, which address universal needs, operate community-driven development, and refine user feedback into standardized products; chip companies, though not primarily in the all-in-one machine business, are happy to deeply integrate hardware, software, and AI capabilities with partners to launch domain-specific solutions; and cross-industry players like Dreame, which leverage trends to open new business avenues.

Currently, due to supply chain price hikes, ODM and OEM manufacturers have reaped the first wave of benefits, while other types of companies are still catching up.

Alliances and collaborations are also forming. Ke Jiejing mentioned their cooperation with domestic chip companies to adapt hardware products. Guo Mengming revealed that his company is partnering with traditional hardware manufacturers for platform pre-installation, with Colorful set to release notebooks pre-installed with their platform in June.

Cai Youquan observed that this round differs from the previous large model all-in-one machine trend. 'Many large model all-in-one machines were led by hardware companies, but this time, software companies are taking the lead.' The reason is that clients care more about tangible benefits, making the logic more akin to B2C user acquisition.

Guo Mengming told DigitFront: 'Users care most about whether the all-in-one machine offers richer AI capabilities. That's the top consideration when buying hardware now.' Shoujie Technology focuses on software, currently addressing basic but overlooked issues like deployment environments, various agent applications, and general-purpose tools like data desensitization and translation. 'First, we need a relatively complete toolchain. Once that's in place, we can develop numerous application systems on top of it.' The company also uses community operations to screen universal needs for development, a strategy Guo describes as 'the Xiaomi approach.'

Cai Youquan explained that their main business is B2B. 'We promote enterprise knowledge bases. The first step for companies adopting AI should be a knowledge base, not directly deploying large models.' The value of a knowledge base lies in AI-ifying enterprise data so AI can understand it, enabling the data to generate value. Packaging agents into a box eliminates deployment difficulties for clients.

Belins in Jiangsu is also focusing on evolving its self-developed software in the second half of the year. 'With intense market competition and many waiting to see what happens, if your software is smarter and more user-friendly, clients are more likely to choose your product,' said Wei Yang, highlighting their software's self-learning mechanism: 'It gets smarter the more you use it.'

Shenzhen Xingyi Technology's software has iteration (iterated) to its fifth version. Version 5.0 incorporates OpenClaw capabilities. Ke Jiejing explained with a comparison: 'Doubao on phones can access phone data to complete tasks... OpenClaw is built on computers, which are more work-oriented, while phones are more for entertainment.' To lower barriers, the company pre-installs industry-validated agents, allowing finance, HR, and other roles to use them out-of-the-box.

In 2026, cloud-based AI will achieve its breakthrough moment, while edge AI will just take its first step. It is driving the collective transformation of the industrial chain and may change the 40-year narrative of the computing industry.