Since 2025, the artificial intelligence (AI) industry has marked several significant milestones.

Early in the year, DeepSeek quickly rose to prominence, prompting the global tech sector to reevaluate the technological advancements of Chinese large-scale AI models. Subsequently, companies such as OpenAI, Anthropic, and Google continued to heavily invest in intelligent agents, with a growing number of products aiming to directly operate computers, utilize software, and accomplish multi-step tasks. Meanwhile, AI-enabled smartphones and smart glasses saw a flurry of releases, while humanoid robots frequently made appearances in factories, exhibitions, and capital markets.

These developments may seem to follow distinct paths. DeepSeek is a foundational model, intelligent agents are software applications, AI glasses are consumer electronics, and humanoid robots involve mechanics, control, and manufacturing. However, when viewed along the same technological evolution trajectory, it becomes evident that they all point toward a unified direction: AI capabilities are expanding from content generation to task execution, further extending into perception and action within real-world environments.

Over the past few years, the primary focus of the AI industry has been to enhance model intelligence. The next challenge is more specific: how to integrate these capabilities into workflows, form stable products, and transform them into calculable commercial value.

Thus, AI has begun to enter its next phase.

01 From ChatGPT, AI Truly Reaches the Masses for the First Time

At the end of 2022, with the release of ChatGPT, AI entered the daily lives of ordinary users in a natural and direct manner for the first time. Prior to this, AI was already widely integrated into systems such as search, recommendation, speech recognition, and machine vision, but most of these capabilities were concealed in the product backend, with users rarely directly perceiving the machine's involvement in decision-making.

ChatGPT revolutionized human-AI interaction. Users no longer need to understand code and algorithms; they can simply make requests in natural language, and the model can generate text, translate content, organize knowledge, and even assist in programming. Generative AI rapidly spread from the tech community to a wide range of scenarios, including education, office work, and content creation, making large models the focal point of competition in the global tech industry.

The industry buzzword for 2023 was the "hundred-model battle." Internet companies, startups, and research institutions intensively released foundational models, with initial market attention focused on who possessed their own large model. By 2024, competition shifted to capability comparison, with parameter scale, context length, reasoning ability, multimodal level, and benchmark scores becoming the most common metrics.

These two years marked a significant capability shift. While traditional AI excelled at recognition and prediction, large models endowed machines with stronger language understanding, content generation, and knowledge processing capabilities. AI began to transition from backend algorithms to frontend tools, enabling industries to reimagine product and process design.

However, challenges soon emerged. Models grew larger, requiring increasing computational power for training and inference, leading to rising R&D thresholds and operational costs. Improving a model's score by a few points on a benchmark might necessitate more chips, data, and funding. The industry gradually realized that if every capability enhancement relied on exponentially increasing resources, the pace of AI adoption and commercial potential would be constrained.

DeepSeek emerged as a turning point at this juncture.

02 After DeepSeek, the Rules of the Large Model Game Changed

DeepSeek garnered global attention, initially due to its model capabilities approaching international leading levels, but its deeper impact stemmed from cost and efficiency.

Prior to its emergence, the large model industry generally advanced along the path of scale expansion. More parameters, larger training sets, and stronger computational power were seen as the primary means of enhancing model capabilities. Leading U.S. model companies, leveraging advantages in chips, cloud computing, and capital, continually pushed training scales to new heights. This path drove rapid growth in AI capabilities but also made foundational models increasingly resemble a capital-intensive competition.

DeepSeek offered an alternative approach. Through optimizations in architecture, training methods, and engineering systems, high-performance models could achieve significant improvements in resource usage efficiency. It did not eliminate the importance of computational power—training cutting-edge models and large model inference still required robust computational foundations. However, it demonstrated to the industry that computational investment was not the only variable; algorithmic efficiency, engineering capabilities, and hardware-software synergy could also influence the final outcome.

This changed the dynamics of large model competition.

Previously, the industry often equated the strongest models with the largest investments. Now, companies must also compare model performance in terms of computational efficiency per unit, cost per unit, and practical task execution. For companies utilizing models, lower prices mean more applications can cross the commercialization threshold. Scenarios like customer service, programming, knowledge management, and data analysis can only transition from pilot to large-scale deployment when model invocation costs are sufficiently low and operations are stable enough.

Meanwhile, the importance of the open-source ecosystem was further amplified. Models like DeepSeek and Qwen lowered the barrier for enterprises and developers to access advanced AI capabilities, enabling more companies to deploy, fine-tune, and secondarily develop these models within their own data and business environments. Model capabilities began to transition from proprietary resources of a few leading companies to a widely accessible technological foundation.

As models grow stronger and cheaper, the industry naturally raises the next question: with such a brain, can AI directly get things done?

03 From Chat Assistants to Intelligent Agents, AI Begins to Enter Workflows

Chatbots primarily revolve around dialogue, with users posing questions and models providing answers. The tasks enterprises truly need are rarely this simple.

Conducting a market analysis requires collecting data, verifying information, summarizing trends, and producing a report; handling a customer service task may involve querying orders, modifying information, initiating refunds, and recording results; completing a software development project also entails understanding requirements, writing code, testing, and debugging. Single-instance content generation can only cover part of these tasks, with the remaining steps still requiring humans to repeatedly operate across different software and systems.

Intelligent agents aim to bridge this gap.

Built on large models, they incorporate task planning, memory, tool invocation, and execution feedback capabilities. After a user sets a goal, the system can attempt to break it down into steps, invoke search engines, browsers, coding tools, or enterprise software, and continue adjusting based on execution results. AI thus begins to enter workflows, undertaking tasks that previously required human coordination to complete.

After Anthropic introduced Computer Use capability, direct AI operation of computer interfaces became an industry hotspot. OpenAI, Google, and domestic vendors continued to advance similar directions, while banks, operators, and manufacturing companies began packaging intelligent agents as digital employees for customer service, auditing, data organization, and knowledge services.

This shift transcends the realm of chatbots but still falls short of fully autonomous operation. The more task steps involved, the higher the likelihood of error accumulation; once email sending, data modification, and financial operations are involved, issues of authority, security, and accountability quickly amplify. Enterprises cannot evaluate models solely based on whether they can complete a single demonstration; they must also assess accuracy, traceability, and exception handling capabilities during long-term operation.

Therefore, intelligent agents are more likely to handle clearly defined, structured tasks for an extended period, with humans still responsible for goal setting, critical judgments, and result reviews. AI will first alter task allocation in work before potentially impacting job structures and organizational models.

04 AI is Leaving Screens and Entering the Real World

While intelligent agents bring AI into software workflows, terminals and robots extend these capabilities into more concrete living and production environments.

Traditionally, ordinary users accessed large models by opening web pages or apps and actively inputting questions. AI smartphones and AI PCs aim to embed model capabilities into operating systems, enabling participation in file organization, meeting recording, image processing, translation, and cross-app operations. AI glasses take this a step further by continuously acquiring environmental information through cameras, microphones, and sensors, helping users identify objects, record content, and receive real-time prompts.

The key to these products lies not in how many AI functions they incorporate but in whether they can establish new interaction entry points. Smartphones remain the most stable personal computing devices due to their user base, data, and mature ecosystem; glasses, closer to human vision and hearing, offer environmental perception advantages but face limitations in battery life, weight, privacy, and user habits that restrict large-scale adoption.

Terminal competition addresses how AI enters personal life, while embodied intelligence seeks to further involve AI in real-world labor.

The rapid rise of humanoid robots in the past two years is directly tied to advancements in large model capabilities. Traditional industrial robots rely on fixed programs and structured environments, excelling at repetitive standard actions on production lines but struggling with open scenarios and ad hoc changes. Developments in large models, VLA, and world models now give robots the opportunity to understand natural language instructions, recognize complex environments, and generate actions based on goals.

This is why humanoid robots are regaining attention. Most real-world factories, warehouses, offices, and homes are designed to human scales. If robots can use human tools, navigate stairs and doorways, and complete diverse tasks, they could reduce the cost of modifying existing environments.

However, significant gaps remain between walking, dancing, and stable operation in terms of reliability, cost, and data. Digital models can access vast amounts of text and images via the internet, but robots require real-world interaction data involving actions, forces, and feedback. Even slight variations in a cup's position, material, or weight may necessitate readjusting grasping actions. Successful lab demonstrations rarely equate to continuous factory operation.

Thus, the most critical metrics for embodied intelligence are gradually shifting from action difficulty to work capability. Whether robots can operate continuously, handle exceptions, and outperform human labor and traditional automation in cost will ultimately determine whether this wave of enthusiasm translates into a genuine industry.

05 From Model Competition to Industrial Competition, AI Enters Its Next Phase

From ChatGPT to DeepSeek, from intelligent agents to humanoid robots, each change in artificial intelligence expands the technological usage boundary.

Large models solved the problem of machine understanding and information generation; DeepSeek prompted the industry to reexamine the relationship between capability and cost; intelligent agents began connecting models to software and business processes; AI terminals compete for personal entry points; embodied intelligence targets task execution in the real world.

This path also indicates that future AI competition will rarely be decided by a single metric. Computational power determines foundational supply, models determine capability ceilings, open-source ecosystems influence diffusion speed, industry data and business processes determine application effectiveness, terminals affect user entry points, and manufacturing and supply chains determine whether AI can truly acquire a "body."

For model companies, the long-term test will be whether they can convert technological advantages into stable revenue; for traditional enterprises, the availability of usable data, clear processes, and suitable scenarios for AI intervention will affect transformation outcomes; for different countries and regions, the combination of chips, energy, software, talent, and manufacturing capabilities will collectively determine their competitive position in the AI industry.

Over the past three years, the artificial intelligence industry has primarily answered whether machines can become smarter. In the coming years, the industry must answer more practical questions: whether these intelligent capabilities can operate stably, create calculable value, and integrate into everyone's life and real-world production systems.

These are the true challenges of AI's next phase.

Next, we will start from the foundational computational power at the bottom of this system and discuss the breakthrough paths China's AI infrastructure is pursuing.