What exactly can AI do for the gaming industry? We had a conversation with Alibaba Cloud and TapTap Manufacturing | ChinaJoy 2026 On-site Session

Content Organized by | Liu Shiwu

The theme of 2026 ChinaJoy is "Playing Along with AI". Last year, everyone was still discussing whether AI could make games, but this year the question has shifted to how to make games, and how to get more people to play the finished works.

This dialogue is one of the sessions in 36Kr Games' series "ChinaJoy On-site Session: Who is Building the Next Decade of AI + Digital Entertainment". Chen Haisheng, Senior Solution Architect of Alibaba Cloud and Zhao Tianyi, Product Lead of TapTap Creator sat at the same table. One party builds game engines, the other provides cloud infrastructure. The two sides are exactly at the upstream and downstream of the game creation chain in the AI era, and both are vendors that lay the foundation for delivering better experiences to creators and players.

Three key judgments were repeatedly discussed during the conversation: the transition from Spark Editor to TapTap Creator is an inevitable result of the arrival of the AI era; as the AI era comes, the core contradiction of many UGC creators has changed from "unable to produce works" to "who will play these works", at this stage, developing engines or building cloud bases are both key links to solve the problem; the most critical principle for developing AI game tools in the current era is "not being swallowed by large models".

Selected Excerpts of the Dialogue:

36Kr Games: From Spark Editor to TapTap Creator, shifting from "visual editor" to "pure AI dialogue", what is the biggest challenge for you personally or your team during this transition?

Zhao Tianyi: When we were developing Spark Editor, the core goal was to solve the problem that "many people cannot make games because they do not master programming". So we built a Chinese visual programming system, allowing people who cannot write code to produce games.

However, after the AI era arrived, people no longer need to write code by themselves, the original development path of Spark Editor has been essentially swallowed by AI.

This "being swallowed by AI" is a reality that I will mention repeatedly today. The advancement of AI will keep absorbing many existing capabilities.

The biggest challenge we encountered when developing TapTap Creator is similar: AI is evolving so fast that we must find those capabilities that will not be swallowed by AI to develop, so that the input of small teams can be valuable. If the work you do will be replaced by AI in just one month, it is a waste of productivity. This has been the biggest challenge over the past six months.

────────────

36Kr Games: Last year at ChinaJoy, everyone was still discussing "whether AI can make games", and this year the theme has become "Playing Along with AI". You, Mr. Chen Haisheng, are a senior expert of Alibaba Cloud who has served the game industry for many years. Do you feel this strong change in your daily work? Are there any projects that impress you particularly?

Chen Haisheng: The feeling is extremely strong.

Last year, clients who came to Alibaba Cloud still asked "can AI really write game code"; this year, clients barely ask about the possibility anymore, they care more about how to use AI and how to make good use of it.

I think this change comes from the maturity of two technical lines:

The first one is MaaS (Model as a Service). Alibaba Cloud's Qwen3.8-Max model has been directly applied to core creation links such as game logic generation and code writing. Model capability is no longer a demo in the laboratory, but a stable service that can be embedded into the production chain of C-end products.

The second one is Agentic Cloud. In the past, AI mostly served as a chat assistant, but now Agents can actually write code, adjust tools, run sandboxes, and finally generate a playable game. Products provided by Alibaba Cloud such as Agent Run and Agent Studio Box are the infrastructures that enable Agents to shift from "can talk" to "can execute".

The project that impresses me the most is UGC creation platforms like TapTap Creator. It allows users with no programming experience to create the games they want through natural language. The smooth operation of this chain essentially relies on the model capability provided by MaaS and the execution and governance capability provided by Agentic Cloud.

────────────

36Kr Games: In the past one or two years, from the product perspective, what changes have taken place in the pain points and demands of creators? The same question for Alibaba Cloud, is the trend you observed as a cloud service provider consistent with this?

Zhao Tianyi: We have been developing UGC game engines, and the core demand we focused on in the past was that users "cannot make games". But this year, this problem has actually been eliminated — users can generate a demo with one sentence on TapTap Creator, and then develop a full-scale complete game through several days of dialogue.

In the first half of this year, we launched about more than 4000 such games. Then we encountered a new problem: who will play the game I made?

The market environment in China is relatively unique. Except for zero-revenue-sharing platforms like TapTap, traffic on other platforms has been fully priced, and it is almost impossible to allocate sufficient traffic to an imperfect game made by a new creator.

So after these creators finish making their games, the biggest problem they face is that no one plays or sees their works. I think this problem will plague the development of the entire industry for a long time.

Chen Haisheng: The trend we observed is highly consistent with what Tianyi described.

From the perspective of cloud vendors, this evolution is roughly divided into three stages:

Stage 1 (two years ago): Everyone was anxious about being unable to make games. Models can write code and generate images, but there is a lack of operation environment, tool invocation and quality assessment, so the creation chain cannot run through. Essentially, "there is MaaS, but no supporting Agentic Cloud".

Stage 2 (last year): Agents can perform actual execution. Alibaba Cloud's products such as Agent Run and Agent Sandbox allow Agents to write code, adjust tools, run sandboxes, and the closed loop of game creation is basically formed.

Stage 3 (current): The problem has shifted to distribution, compliance and quality grading. After the explosion of UGC content, platforms can no longer only focus on whether games can be produced, but also pay attention to whether the finished works can be launched, whether there is violation of regulations, and how to grade the quality. This requires the observable governance capability of Agentic Cloud — quality assessment at the Agent entry, and behavior audit through Agent Identity.

Our judgment is that the demands of game clients will shift from purchasing resources to purchasing the full-stack capability of MaaS + Agentic Cloud.

────────────

36Kr Games: For people without programming experience to generate games through natural language, what is the hardest part in the process of "lowering the threshold"? Same question for Alibaba Cloud, what new requirements do such products put forward for underlying cloud services, and what are the differences from the demands of traditional game vendors?

Zhao Tianyi: The hardest part may not be lowering the threshold itself, but while maintaining this goal, clearly realizing that we have another group of users, as well as a more abstract opponent.

TapTap Creator is positioned at "everyone can make games, you can realize your idea as soon as you think of it". This is one part of our users, they may have no basic game production capabilities, and the upper limit of AI is their upper limit. For this group of users, what we need to do is to raise the upper limit of AI running in our tools, so that the lower limit of the games made by developers can be higher. For example, we have developed a set of UI theme libraries for the professional team of XD, they can switch the theme with one sentence; we also provide sequence frame tools that can generate professional sequence frames.

But we have another important group of users — they already know how to make games. The core proposition for them is: TapTap Creator must not become a constraint that drags down their productivity.

How to solve this? We have done some work recently: allowing everyone to use their own Agent locally; the engine will be open source in the next period of time, so that everyone can deploy it anywhere at will. These are all measures to remove constraints for professional developers.

There is also a more abstract thing — the large model itself is a "ghost". It keeps making continuous progress behind the scenes, and every progress will swallow some existing capabilities.

Last year we thought system prompts, Skills (skill templates) and Harness (framework orchestration) were very important, but this year with the advancement of models, we deleted a lot of our own Skills, because we found that many contents do not need prompts, AI can write them well by itself. We used to have some templates to explain "how to make a complex game", but now we no longer need them, because AI already understands this logic.

So when we develop every requirement, we have to ask ourselves: will this requirement be quickly swallowed by large models in the future, or will it become a lever to amplify the progress of large models?

As a product manager, I think about these questions every day. We are a small team, so we need to invest resources in points that are truly valuable.

Chen Haisheng: The point you mentioned has pointed out the fundamental change that AI-native game creation platforms bring to traditional cloud architecture.

When traditional game vendors come to Alibaba Cloud, their core demand is stable operation resources — servers, databases, network security, to support the normal operation of games.

But platforms like TapTap Creator purchase the full-stack capability of MaaS + Agentic Cloud from Alibaba Cloud. It is specifically reflected in three aspects:

First, how to guide people who do not understand the game production process to create? At the product level, users are guided through MCP (Model Context Protocol), Skills and Harness; at the underlying cloud layer, Alibaba Cloud's Agent Run and Agent Teams provide Scale (elastic scaling) and MCP ecology. For example, encapsulate the experience of "how to make RPG games, how to make MOBA games, how to make platformer games" into reusable Skills, and Agents execute structured workflows according to Skills. Complex tasks can also be split into Worker nodes through Agent Teams for collaborative completion.

Second, how to guarantee the upper limit for professional developers? Alibaba Cloud adds degradation, current limiting and content security capabilities on the gateway to ensure compliance. Qwen3.8-Max directly improves product capability and efficiency — in our actual operation, it can complete game production in one attempt. At the Agent Cloud layer, Agent Run supports three modes: no-code, low-code and high-code, professional developers can choose Python, Java or any framework, and will not be restricted by the platform.

Third, how to ensure the stability of multi-terminal and mobile development environments? Traditional development environments are installed on local devices, with complex configuration and version conflicts. Alibaba Cloud's FC Agent Sandbox and ACS Agent Sandbox move the development environment to the isolated cloud sandbox, support more than 50 languages, can run browser automation, and provide second-level elasticity. Creators do not need to care about environment configuration, nor worry about mutual influence between AI-generated codes in multi-tenant scenarios.

To sum up in one sentence: traditional game vendors purchase cloud resources for stable game operation; platforms like TapTap Creator purchase the complete and governable infrastructure capability that "enables Agents to make games for users".

────────────

36Kr Games: In the era of zero-threshold creation, it is almost inevitable that a large number of works will pour in. How does TapTap Creator balance "creation freedom" and "content quality" through mechanisms? From the perspective of platform technology, what support can Alibaba Cloud provide for the platform to control content quality?

Zhao Tianyi: We spent quite a long time thinking about this issue in the first half of this year, and finally landed the solution in a very short time.

The reason why we thought for so long is that it shook the entire original traffic distribution system of TapTap. The number of games on TapTap in the first half of last year was 502, and this year it has exceeded 5000 — 4699 of these more than 5000 games are AI games made by TapTap Creator.

Both the game output and the number of games accessible to users have increased dramatically. Our review system received 260 new games on the busiest day, and the entire process of the review team almost collapsed.

Later we implemented a new system: we no longer give all games basic cold-start exposure as before, and then collect data. We use less traffic to judge the play experience of a game — collect data according to its play duration and retention rate, and then decide whether to expose it to more users.

This solves the problem: we only need very little exposure to judge whether a game is of low quality, so the harm for users to encounter low-quality games is reduced. After we accommodate all newly launched games with very limited exposure, we select high-quality games from them, and give more exposure preference to creators who make high-quality works. This is beneficial to the entire ecosystem and all developers.

36Kr Games: From the perspective of platform technology, what support can Alibaba Cloud provide for the platform to control content quality?

Chen Haisheng: Our idea is not to first lower the threshold with AI, and then use manpower to plug the loophole, but to make AI itself become the quality control link.

Technically, it is divided into two parts:

The first part is the multi-modal capability of MaaS. The multi-modal capability of Alibaba Cloud Qwen3.8-Max can check game screen recordings, screenshots and audios, judge whether the game can run normally, whether there are obvious bugs, what the art style and game type are, and even identify illegal content such as vulgarity, violence and infringement, and input these information directly to TapTap Creator to assist content grading.

The second part is the AgentLoop assessment system in Agentic Cloud. The core concept is Agent as a Judge — the evaluator itself is also an Agent. It can use pre-set assessment test sets, and also support custom assessment dimensions.

For example, if the user says "I want a Mario-like game", the Agent evaluator will check whether there are jump mechanics, enemies and level end points in the generated game. It can also perform hallucination detection, accuracy detection and task completion degree detection.

In terms of compliance and security, Agent Identity can perform behavior audit. Assign independent identities to each Agent, inject credential security when accessing the content library, and all behaviors can be audited. When the Agent detects abnormal behaviors such as unauthorized access and data exfiltration, the platform can send an alarm within minutes.

In summary: we do not restrict the freedom of creators, but provide the platform with a set of technical capabilities of "pre-assistance grading, real-time in-process assessment, and post-event traceability".

────────────

36Kr Games: There are more and more AI game creation tools, what is the core competitiveness of TapTap Creator that is the hardest to be replaced?

Zhao Tianyi: In the short term and even in the medium and long term, everyone will see more AI tools emerging. In the current era, the cost of building an Agent is very low — all models can be accessed, the cost of tokens is getting lower and lower, and there are a large number of examples of infrastructure such as Harness on the market, so it is not difficult to build a basic tool.

But these basic tools are very easy to be swallowed by the progress of large models. Maybe you have made efforts to make it perform better on the current model, but the model will absorb your efforts next month after iteration.

What we want to do are those capabilities that will not be swallowed by large models.

One is to raise the upper limit of the engine — make the progress of large models bring more productivity, instead of the progress of large models swallowing the capability of our engine.

The other is the TapTap ecosystem — solve the distribution problem after the game is made, so that developers can make games with peace of mind, instead of their finished games being ignored, their painstaking efforts being buried by time, and receiving no feedback. This chain needs to be well built, and commercialization also needs to be well developed. Our commercialization is still relatively basic now, but there are already games with monthly revenue of hundreds of thousands of RMB, and some games have a daily revenue of 15,000 RMB —