CES 2026 heralds the dawn of a new era: the AI industry is poised to enter the authentic 'Era of Action Intelligence.'

For the B2B market, this will mark a cycle characterized by a slower pace of progress but greater certainty. AI will not abruptly revolutionize all industries overnight; rather, it will subtly yet profoundly integrate into the core business processes of enterprises and the operational logic of the physical world. Those capable of leading the way in transitioning from 'selling technology' to 'selling results' are more likely to secure a stable foothold in the industrial landscape of the next decade.

In the first half of AI's evolution, imagination reigned supreme; in the second half, the true challenge lies in who can effectively integrate intelligence into the real world and sustain its operation steadily and consistently over the long haul.

Author | Dou Dou

Editor | Pi Ye

Produced by | Industry Insight

In January, Las Vegas was ablaze with a grand tech spectacle.

CES transcends being a mere carnival for consumer electronics enthusiasts; it serves as a centralized 'reconciliation day' for chipmakers, device manufacturers, automakers, internet giants, robotics firms, and B2B suppliers across diverse industries.

This year's CES has generated even more buzz than the previous two years.

An intuitive observation is that the term 'Consumer Electronics Show' no longer fully encapsulates its essence. AI has officially transitioned from a feature to the default foundation for all new products, ranging from next-generation computing platforms like Nvidia Rubin and AMD MI400 to the Uber-led Robotaxi alliance, humanoid robots, ultra-thin wallpaper TVs, home health devices, and AI household robots. The entire exhibition hall exudes an atmosphere suggesting that technology is on the verge of 'large-scale transformation of the physical world.'

For this reason, CES 2026 serves as a reflective mirror.

On one hand, it mirrors the AI industry's shift from conceptual expansion to structural reconstruction; on the other, it compels all industry participants to confront an unavoidable question: in this industrial reshuffling driven by the computing system, edge devices, robots, and the global supply chain, where do they envision themselves standing?

I. AI Infrastructure Competition: From Performance to Delivery Capabilities

At CES 2026, the evolution of AI infrastructure was striking.

Nvidia focused on the new-generation Rubin platform, which significantly reduces token costs and overall operational expenses, rather than solely emphasizing 'speed.' It intentionally downplayed the narrative of 'single-card performance' and instead repeatedly highlighted system-level collaboration, where CPU, GPU, interconnection, networking, storage, and security capabilities are presented as an integrated whole.

AMD's approach at this year's CES mirrored this direction.

Rather than engaging in head-to-head competition in single-point performance, it introduced the Helios rack-level platform and centered its narrative around the Instinct MI400 series in terms of computing power and delivery forms. More symbolically, AMD specifically launched a product like the MI440X, clearly positioned as an AI chip suitable for traditional enterprise on-premises data centers, emphasizing the ability to complete training, fine-tuning, and inference without the need for AI-specific data center rebuilds.

These announcements collectively sent a clear signal at CES 2026: infrastructure vendors are no longer assuming their customers are solely cloud providers or research institutions but are instead treating 'enterprise data centers' as their core battleground.

Comparing this shift to CES 2025 reveals a clearer logical progression. Last year, Nvidia's spotlight was more on the expansion of the Blackwell architecture into consumer and developer ecosystems, with AI PCs, RTX 50 series, and Copilot+ PC-related chips becoming focal points. The main narrative then was 'AI will permeate all devices,' with computing power being distributed to endpoints and individuals like utilities. This contrasts with the current scenario.

Analyzing the reasons behind this change, it becomes evident that after a year of running large model applications in real-world environments, enterprises have gradually discovered that the bottleneck lies not in whether they can use AI but in whether they can use it continuously and at scale. While training costs are decreasing, the volume of inference calls is rapidly expanding. When businesses are operational, computing power, electricity, cooling, networking, and operations and maintenance collectively become long-term expenses. Enterprises are beginning to realize that the cost structure of AI is shifting from one-time capital expenditures to ongoing operational expenses.

Meanwhile, engineering realities are also coming to light. Enterprise data centers are not laboratories; expansions require scheduling, cooling systems need upgrades, networks must be reconfigured, and security and compliance must undergo step-by-step reviews. Simply having more powerful cards cannot solve these issues and may even exacerbate system complexity. What enterprises truly need is a computing power system that can be rapidly replicated, standardized for delivery, and stably operated within existing frameworks.

Against the backdrop of geopolitical and supply uncertainties, enterprises have become more cautious in their procurement decisions. They are no longer willing to bet their critical operations solely on a single form factor or component but instead hope for computing power systems that offer greater composability and substitutability.

This is why trusted computing and system-level security are re-emerging in the infrastructure narrative at this year's CES.

Overall, this aspect of CES 2026 reflects the core trajectory of the industry. AI infrastructure is not about who is faster or bigger but about who can transform AI architecture, industrialized delivery, and long-term operations and maintenance into replicable enterprise products. In the B2B world, this systematic delivery capability will become the key criterion for determining whether a supplier can truly make it onto enterprise budget lists in the next three years.

II. Edge AI: From Toy to Enterprise-Level Node

As infrastructure vendors begin designing computing power around enterprise delivery, the role of edge AI has naturally evolved.

At CES 2026, edge AI vendors started addressing a more enterprise-oriented question: whether these devices can be managed and deployed as IT assets.

The most representative change came from PC and enterprise terminal vendors.

Lenovo unveiled its cross-device AI system, Qira, at CES and deliberately defined it as an intelligent environment for multi-device collaboration rather than an assistant on a single device. Qira is designed to seamlessly invoke capabilities across PCs, tablets, phones, and other terminals and switch inference locations between local and cloud based on tasks. Meanwhile, Lenovo partnered with Nvidia to introduce the 'AI Cloud Gigafactory' solution, emphasizing the compression of enterprise AI environment deployment cycles to a weekly basis. This effectively packages the delivery chains for edge, edge, and cloud together, directly targeting enterprise IT departments.

Last year at CES, the main theme for edge AI was 'popularization,' with vendors most concerned about whether Copilot+ PCs, NPU computing power, and edge models could run. Now, the focus has shifted to how enterprises can manage these capabilities at scale.

Dell's attitude shift at CES is highly illustrative of this point. It publicly stated that consumers would not pay a premium for the concept of 'AI PC' alone and that overemphasizing AI might even confuse users. What truly drives procurement decisions remains battery life, performance, stability, and manageability.

The actions of infrastructure vendors are also echoing this change. AMD emphasized AI chips like the MI440X for enterprise on-premises deployment at CES, intending to keep more inference and fine-tuning within enterprises' own data boundaries. When Intel released the Core Ultra Series 3, it highlighted edge scenarios such as robotics, automation, smart cities, and healthcare, repeatedly emphasizing the performance-to-power ratio in video analysis and local inference tasks.

The collective signal from these announcements is that the edge is no longer just a supplement to the cloud but a key variable for enterprises in controlling costs and risks.

The reasons behind this narrative shift for edge AI are quite clear. First is cost pressure. Over the past year, after using cloud-based inference in real-world business operations, enterprises have gradually realized the unpredictability of inference costs. Once call volumes surge, expenses become difficult to lock into long-term budgets, making edge and local inference important means of stabilizing cost structures. Second is data and compliance boundaries. In industries like manufacturing, healthcare, and energy, much data is inherently unsuitable for leaving the premises. Local processing and edge inference are not technical choices but compliance prerequisites. Third is operational complexity. When the number of edge devices reaches hundreds or thousands, enterprises fear not a lack of computing power but the inability to uniformly update, audit, and assign responsibility, forcing vendors to make edge capabilities 'manageable systems.'

Under these realistic constraints, edge AI has begun to transition from 'being able to run models' to 'being accepted by IT systems.' Cross-device systems like Qira essentially respond to enterprises' needs for unified permissions, logs, task allocation, and data flow. The local inference and edge computing power emphasized by infrastructure vendors provide enterprises with more architectural choice space.

The resulting trend is also very clear. For some time to come, edge AI will not replace the cloud but will serve as a 'stabilizer' in enterprise AI architectures to control costs, mitigate risks, and reduce latency. The truly valuable aspect is not a single hit AI hardware product but the ability to orchestrate edge, edge, and cloud into a governable network. Whoever can unify devices, models, permissions, and operations and maintenance into a single system is more likely to propel edge AI from pilot projects to large-scale deployment.

III. AI Implementation: Entering a Replicable Stage

At this year's CES, it was evident that AI implementation in some areas had become more 'granular.'

Specifically, in the logistics and warehousing sector, AI is no longer just about demonstrating box-moving capabilities.

Arm's establishment of the Physical AI business unit hints that robots are no longer standalone products but are meant to be embedded into production lines, warehousing route planning, collaborative scheduling, and other systems. Companies like Unitree Technology and ZhiYuan Robotics showcased stable-running embodied intelligent products in manufacturing, retail, and warehousing logistics scenarios, indicating that this sector has moved beyond prototype demonstrations into engineering, mass production, and deployable stages. On-site, Aoshark Intelligence displayed exoskeleton robots, which directly enhance human-machine collaboration efficiency in light industrial and repetitive operation scenarios.

Compared to 2025, when robots were more like a 'robot show,' 2026 resembles an industry ecosystem exhibition on 'how robots integrate with businesses,' reflecting that logistics industry clients have shifted from curiosity to substantive judgments on solution implementability.

In the manufacturing sector, the narrative has shifted from 'what AI can do' to 'how AI does it.'

The most typical example came from Siemens, which unveiled the Digital Twin Composer at the exhibition. This platform tool integrates digital twins, real-time physical data, and design or engineering processes, enabling enterprises to verify risks and costs through software before transforming production lines. It is currently being piloted at PepsiCo's U.S. factory.

This move essentially addresses the core pain points of manufacturing enterprises' process transformations, which are high investment, high risk, and long cycles. While AI was previously used mostly as an 'auxiliary decision-making' tool, the emergence of the Digital Twin Composer indicates that AI is being leveraged to directly optimize core processes at the manufacturing execution and engineering levels, becoming a production-grade tool.

In the medical scene, AI is redefining service and research and development processes.

Withings' Body Scan 2 is an AI health device capable of measuring multiple physiological indicators, providing real-time health advice and long-term monitoring services. Enterprises can adopt it as part of long-term health management and corporate wellness programs rather than selling it as a one-time hardware product.

These scenario-specific cases demonstrate that AI implementation is gradually unfolding across industries, processes, and nodes. It is no longer a 'universal capability' but can be embedded into the core business linkages of industry clients.

It can be predicted that in the next one to three years, AI implementation in industries will not rely on single-point innovations to attract attention but will win markets through the replicability, deliverability, and long-term operational capabilities of industry solutions.

IV. Chinese Vendors' Presence: From 'Abundance' to 'Deliverability'

In these scenarios that have started discussing delivery, deployment, and operations and maintenance, the presence of Chinese vendors has become more tangible.

Notably, Chinese enterprises have started to demonstrate their prowess at several crucial industrial nodes. These nodes happen to be the most sought-after and influential positions as AI progresses towards implementation.

The most conspicuous transformation has taken place in the domains of embodied intelligence and edge AI.

At this year's CES, Chinese vendors of humanoid robots made up roughly 55% of all robot exhibitors at CES 2026. Chinese enterprises have emerged as a dominant force that cannot be overlooked.

For instance, Unitree Technology showcased a variety of humanoid and quadruped robots on-site, including the G1, H2, and R1 models. These robots cater to diverse scenarios such as research, inspection, services, and demonstrations, presenting a highly organized "product matrix" approach rather than simply displaying a single concept machine. ZhiYuan Robotics made its debut with a comprehensive lineup of humanoid robot products and revealed that its cumulative shipments had reached thousands of units, explicitly highlighting its delivery capabilities and scale experience. Additionally, companies like StarMotion Era and Zhongqing Robotics also made their presence felt.

This shift represents a stark contrast to the situation in 2025.

Last year, Chinese robot companies were primarily focused on demonstrating "what they could do," emphasizing technological breakthroughs and prototype capabilities. In 2026, they are addressing the questions that truly matter to enterprise clients, such as the availability of stable versions, the feasibility of mass delivery, and who will be responsible for maintenance when issues arise. This transformation itself indicates that Chinese enterprises have shifted their focus from "being noticed" to "being selected for procurement."

Similar changes have also occurred in the realm of edge AI carriers. Chinese vendors like Rokid showcased lighter, more independently operating AI glasses products at CES, emphasizing all-day wearability, independent communication capabilities, and multi-model access. The value of such products lies in their natural suitability for enterprise scenarios, such as remote inspection, instant translation, equipment operation guidance, training, and collaboration. These are precisely some of the edge forms most likely to generate scalable demand in the business-to-business (B2B) market.

If we take a broader view, we'll find that these changes are not isolated incidents but rather a more proactive "positional choice" made by Chinese companies in the global AI industry division of labor. At the level of foundational models and ultra-large-scale computing power, overseas companies still hold the dominant position. However, as AI enters the engineering and implementation phase, the factors determining industrial speed begin to shift towards hardware integration, systems engineering capabilities, supply chain efficiency, and cost control—areas where Chinese companies have more familiarity and greater accumulated expertise.

Therefore, the evolving role of Chinese companies at CES 2026 represents not just a shift from "quantity to quality" but also a move from "demonstrating capabilities" to "occupying key nodes." These nodes collectively form the core of the AI industry's second half, namely edge-side carriers, embodied intelligence, engineering delivery, and large-scale manufacturing.

It is foreseeable that in the short term, the competitiveness of Chinese AI on the global stage may not primarily lie in the advancement of foundational models but rather in the mass-scale implementation of AI. In areas such as robotics, AI glasses, commercial displays, and industrial equipment, Chinese companies are striving for greater definitional authority. If they can continue to address shortcomings in software ecosystems, enterprise-level management, and long-term service capabilities, China's position in the global AI industrial chain will no longer be merely that of a "participant" but will gradually move closer to being a "rule shaper."

V. 2026: The Arrival of the "Physical AI Era"

When we consider all the changes at this year's CES together, a clear trend emerges: the era of Physical AI is upon us.

This is evident from the conference content as well.

In past CES events, the core narrative around AI primarily focused on generative capabilities, Copilot interaction experiences, and how to produce text, images, audio, and other content more quickly and accurately. By 2026, NVIDIA CEO Jensen Huang, in his keynote speech, clearly defined the main theme of the entire CES as "Physical AI," representing his highest strategic judgment on the future of the AI industry.

The so-called "Physical AI" does not simply refer to deploying AI capabilities to terminal devices but rather to machines' ability to understand the real world, make inferences, and execute physical actions based on sensory results. In other words, AI will evolve into intelligent agents capable of "seeing, thinking, and acting," able to complete tasks in real-world scenarios rather than merely outputting content on screens.

The actions surrounding Physical AI at this CES were highly concentrated.

NVIDIA did not release traditional consumer graphics cards but instead dedicated nearly two hours of its keynote speech to outlining an integrated strategy around Physical AI, the Rubin platform, the autonomous driving Alpamayo system, and robotic inference architectures. It used a trinity framework of "AI brain + physical execution + simulation systems" to illustrate the future direction of the industry.

Arm launched its Physical AI business unit, explicitly elevating the embodied intelligence scenarios of robotics and automotive to a level equal with cloud and edge computing, and identified them as key engines for future automation outputs.

Numerous large automotive and autonomous driving systems have also incorporated this concept into their roadmaps. For example, NVIDIA's Alpamayo autonomous driving platform is defined as a next-generation automated parking and active driving system with inference and explanation capabilities, serving as a practical implementation of Physical AI in the mobility sector.

The common characteristic of these actions is that the industrial chain is shifting AI capabilities from "generating content" to "executing actions in the real world," such as robots moving on production floors, autonomous vehicles navigating and controlling complex road conditions, and home/industrial automation robots completing specific tasks.

Behind this trend lies a very practical industrial logic.

First, foundational AI models and large-scale inference capabilities have matured to the point where they can interpret the physical world, rather than merely producing textual results. Models and platforms like NVIDIA's Cosmos and Alpamayo are designed to enable AI to understand environments, make decisions, and execute actions—a capability that represents the key transition from virtual intelligence to embodied intelligence.

Second, over the past few years, the pain points in enterprise-level implementation have shifted from content generation to task execution and business impact. For example, robotic handling, inspection, reception, and cleaning, as well as urban-level operations and low-speed traffic management in autonomous driving, all require AI to reliably operate along the chain of real-world perception, inference, and execution. Physical AI treats this chain as a holistic system rather than fragmenting it into isolated capabilities.

Finally, market forces are driving this trend, as enterprises and cities have far greater demand for automation scenarios than purely digital needs. Labor shortages, rising operational costs, and the need for autonomous systems in complex environments have prompted both vendors and buyers to focus on AI systems capable of executing tasks at the physical level, rather than merely performing inference in the cloud or on terminals.

Thus, Physical AI is emerging as the next major arena for industrial competition. While AI procurement in the past focused more on cloud models or single-point devices, end-to-end physical intelligence systems will become the new standard for enterprise procurement in the future. Comprehensive physical intelligence systems, including robots, autonomous vehicles, and automation equipment, will become new long-term procurement projects for large enterprises.

Conclusion:

CES 2026 heralds the arrival of a new era: the AI industry is stepping into a true "Age of Actionable Intelligence."

For the B2B market, this will be a slower but more certain cycle of advancement. AI will not suddenly disrupt all industries overnight, but it will gradually and quietly embed itself into the core business processes of enterprises and the operational logic of the physical world. Those who can first complete the transition from "selling technology" to "selling results" will be more likely to secure a stable position in the industrial landscape of the next decade.

In the first half of the AI era, imagination was key; in the second half, the true test will be who can effectively bring intelligence into the real world and sustain it over the long term.