Hard-Kr Exclusive Coverage at IFA: AI Hardware Can Do Far More Than Just Answer Questions

If we say that at the Consumer Electronics Shows over the past two years, all products were in a rush to prove they are "AI-enabled", then at IFA 2026, a shift is underway: AI is no longer just an add-on stuffed into TVs, refrigerators and computers, but has begun to become the underlying logic for hardware operation.

From September 4 to 8, IFA 2026 was held in Berlin, Germany. According to official data, this edition of the exhibition gathered more than 1,900 brands, with robots, artificial intelligence, smart home and even flying vehicles becoming the key highlights on display.

However, what deserves more attention than the number of new products is that the competition landscape of AI hardware has changed.

In the past, manufacturers competed to see who could connect to larger models and generate more fluent text; now the industry is starting to answer more specific questions: Can AI run independently locally? Can it perceive the physical world and execute tasks? Can it save users time, reduce costs, and even enhance human physical capabilities in real scenarios?

From agentic PCs on the desk to exoskeletons worn on the body, AI is stepping out of the chat box and into the real world.

AI goes local: computers are evolving from tools to "agents"

In this round of AI wave, personal computers once found themselves in an awkward position.

On one hand, PC manufacturers have intensively rolled out AI PCs with continuously improved AI computing power of processors; on the other hand, when users actually use large models, most of the computing still occurs in the cloud. Computers are more like an entry point connecting to AI services, rather than devices that independently host AI.

But the cloud is not the optimal solution for all scenarios. For ordinary consumers, uploading data to the cloud may only raise privacy concerns; for industries such as government affairs, healthcare, finance and scientific research, whether data can leave local premises is often a clear red line for compliance. In addition, when agents need to run for a long time, frequently call models to complete tasks and process internal enterprise data which consumes more tokens, the latency and cost of cloud inference will be further amplified.

Therefore, a clear direction at this IFA is to keep more AI tasks local and complete them through end-side models with larger parameters and higher intelligence. GMK globally debuted its desktop-level AI supercomputing workstation EVO-X5 Pro at the exhibition, equipped with AMD Ryzen AI Max+ PRO 495 processor, the world's first agentic PC that can fully run 300B parameter-level large language models (LLM) offline, and it is defined as an Agentic PC oriented to the agent era.

This product is equipped with AMD Ryzen AI Max+ PRO 495 processor, adopts a collaborative architecture of CPU, GPU and NPU, and is equipped with 192GB of unified shared memory, supporting a maximum allocation of 160GB as dedicated video memory. According to GMK, the EVO-X5 Pro can run 300B parameter-level large language models completely offline, and it is mainly oriented to scenarios such as government affairs, biomedicine, financial investment, classified scientific research and enterprise commercial use. It also supports multi-device interconnection to build a distributed computing cluster, and is integrated with out-of-band remote management and an independent security chip to meet enterprises' requirements for continuous operation and data security.

What really deserves attention is not just the number "300B", but the change in the role of PCs: it is no longer only responsible for invoking AI, but begins to host models, data and agent tasks locally. If AI PC solves the problem of "whether a computer can run AI", the Agentic PC tries to answer the question of "whether a computer can continuously complete tasks on behalf of people".

Of course, local large models still face problems such as power consumption, heat dissipation, software ecosystem and actual deployment costs. The parameter scale does not equal the real experience. But when the model operation location moves back from the cloud to the desktop, a new round of competition centered on data, computing power and agent entry points has begun.

AI integrates with human body: robots do not have to look like humans

Another path for AI to enter the physical world is robots.

This year, IFA lists robots as one of the core future topics. Exhibited products include humanoid robots that can dance, perform tai chi and complete difficult movements, as well as smart devices for home use, rescue and assisted living. The IFA official summarizes this trend as Physical AI: AI no longer only processes information, but perceives the environment, responds to changes and completes real-world tasks.

However, embodied intelligence does not only have one solution of "building a complete human being". Compared with humanoid robots that pursue general capabilities, exoskeletons choose a path that is closer to humans: not replacing humans, but identifying human movement intentions and then amplifying human physical capabilities.

In the embodied intelligence track, most of the spotlight falls on humanoid robots that try to replace humans, but some enterprises choose another path: making intelligent technology collaborate with the human body to enhance human mobility.

Yuanshan Zhixing is one of them. Centered on human full-life-cycle services, the company builds an EAI full-scenario smart ecosystem, and has now launched exoskeleton robots, quadruped robots and a cloud technology platform, trying to apply intelligent technology to scenarios such as mobility assistance, companionship, safety and health management.

At this IFA, Yuanshan Zhixing brought the "Tashan S1" hip mobility assistance exoskeleton and the "Banshan M1" light-duty handling exoskeleton, among which the Tashan S1 makes its overseas debut. It is understood that different from some products on the current market that pursue single parameters such as weight, battery life or assistance intensity, the Tashan S1 is mainly oriented to active elderly people, and is specially tuned based on gait data of more than 10,000 elderly people, focusing more on safety and comfort in daily use. The product is integrated with functions such as fall-proof airbag accessories, multi-layer mechanical locking structure and remote footprint sharing. At present, Yuanshan Zhixing has entered hundreds of experience points in more than 50 cities across China, has served more than 10,000 elderly users in total, and promotes commercialization through paths such as cultural tourism leasing, elderly care communities and C-end sales.

The exoskeleton is not a new concept, but in the past it was mostly applied in professional fields such as rehabilitation medicine and industrial handling. Now, with the continuous maturity of sensors, motors, algorithms and batteries, exoskeletons are trying to evolve from medical devices or industrial equipment to consumer-grade products that ordinary users can purchase, rent and use continuously.

This also means that the value of robots is not limited to imitating humans. Compared with completing backflips on the stage, whether a machine can stably and safely help the elderly walk one more kilometer may be closer to the real starting point of the commercialization of embodied intelligence.

Chinese AI hardware goes global: entering the "scenario competition" era

From GMK to Yuanshan Zhixing, the two companies have completely different product forms: one puts larger models into desktop devices, the other wears algorithms and motors on the human body. But behind them points to the same trend — AI hardware is shifting from demonstrating capabilities to solving practical tasks.

This also means that the way Chinese technology enterprises compete at European exhibitions is changing.

In the past, the most prominent advantage of Chinese enterprises' going global was supply chain efficiency: they could make products faster, with higher configurations and lower prices. Today, these capabilities are still important, but they are no longer sufficient to build long-term barriers. Especially when AI becomes the standard configuration of all products, simply connecting to large models and adding voice interaction can be easily replicated by competitors.

The real differentiation in the next stage will come from the understanding of scenarios.

A local AI workstation can not only prove how large a model it can run, but also answer questions such as which industries are willing to deploy it, how to ensure data security, and whether it can integrate into the original workflow; a mobility assistance exoskeleton can not only let people experience the freshness of "being pushed to walk" on the exhibition booth, but also answer questions such as whether the elderly dare to use it, whether it is comfortable to wear for a long time, how to protect users when accidents occur, and who is willing to pay for it.

This may also be the signal released by IFA 2026: when AI starts to move from the cloud to the desktop, from the screen to the human body, the end point of competition for hardware companies is no longer to launch a new product with AI, but to find a real scenario that is specific enough, frequent enough, and sufficient to support a closed commercial loop.

In the first half of the AI hardware industry, everyone rushed to prove that machines can understand humans.

In the second half, what really determines the outcome is whether machines can start doing things for humans, or help humans do more.