On September 1, 2026, the 247th Order of the Chengdu Government officially came into effect. This is the country's first local government regulation aimed at promoting the development of the artificial intelligence industry, consisting of six chapters and 36 articles.
Most people scrolling past this will swipe over it as just another AI policy announcement. That's natural—there have been so many similar headlines over the past few years that eyes have long since grown calloused. But pause for a moment. The weight of this particular item lies not in Chengdu itself, nor in the specific wording of those thirty-six measures. It functions more like a signal card: the second face of the artificial intelligence era is now showing itself. The first face we've been watching for years—models growing more powerful, computing capacity expanding, applications multiplying across industries. The second face is only just beginning. Its name is unassuming: rules, starting to be written. In the past, we only watched whether machines would get smarter. Now we have to start watching how society catches the machine.
This article does not interpret the provisions of the regulations one by one, nor does it comment on the actions of any party. It is used as a mirror. We use the three lenses of management - strategy, organization, and governance - to look at the major event of the "arrival of the artificial intelligence era" itself. This is not a list of subsidies, nor is it a policy interpretation, it is an observation. After looking in the mirror, we will answer a question that is closer to us: during the rule-making period, for those of us who do artificial intelligence research and work in artificial intelligence, how should we write our own next article?
After reading this, you will have three pairs of glasses in your hand - one for strategy, one for organization, and one for governance, plus a action plan with a note that says "next article". Let's take a look together at the face that has just been revealed.
Another Face of the AI Era: Institutions Begin to Take Shape
In recent years, when judging the arrival of the "artificial intelligence era", we have almost only looked at the technical aspect. Models, computing power, and applications have taken turns occupying the headlines. The model aspect focuses on the scale of parameters, capability boundaries, and how to suppress those serious nonsense; the computing power aspect discusses how to build clusters, how to assemble chips, and where the electricity comes from; the application aspect discusses how intelligent bodies replace human labor, how generated content is laid out, and how models are embedded in terminals such as mobile phones and cars. This face is very bright, so bright that it dazzles, and for several years, we have almost put all our attention on it. Discussing artificial intelligence is equivalent to discussing technology, as if as long as the machine is smart enough, fast enough, and cheap enough, the era will naturally arrive.
Another aspect has slowly come into view. On September 1, 2026, the 247th government order in Chengdu officially took effect, with six chapters and 36 articles, incorporating dimensions such as innovation, industry, factors, and governance into a local government regulation. In other words, from how to encourage innovation to how industries gather, to how data and talent flow, and to who is in charge when problems arise, all are framed by a single regulation. This is not an isolated case. Looking back, Shenzhen's Special Economic Zone Artificial Intelligence Industry Promotion Regulations were implemented as early as November 2022. Shanghai's Regulations on Promoting the Development of the Artificial Intelligence Industry, which was implemented in October 2022, is the country's first provincial-level artificial intelligence local regulation, with six chapters and 72 articles. Looking at the present, multiple regions including Guangdong, Jiangsu, Zhejiang, and Fujian are also researching and formulating related laws. A line can be drawn from 2022 to 2026, from the coast to the inland.
the a of a top of a minimalized placement of a noored entity of also noored path of a noored a of a no unknown path of any path of noored path of noored path of noering path of noored path of noored path of noering path of noored it " place of noored path of noering path of no " : " place of noored path of " : " path of noored path of noored path of noored the noored path of noored path of noering common path of noering path of noored path of noering, " : " noored common path of noored work in a noored path of noored work in a no with a minimal path of a noored work in a noored work of noored path of noored common path of noored work in a noored work of noored work of noored work of no in a noored path of noored work of noored work of noored work of noored common work of noored work of noored work of of noering count of noored path of no noored count of noored work in a noored work of noored work of noored work of noored common path of noored
When these facts are put together, an inference emerges: artificial intelligence has been "caught" by a complete set of local legislation for the first time. It is moving from the "technology circle" to the "social circle". In the past, we discussed "who can create it", now we need to discuss "who can use it, how to use it, and who will regulate it". Moreover, this is clearly not an isolated action by a single city. Shenzhen and Shanghai have taken the lead in legislation, with multiple locations following up on research, and rules are growing simultaneously across the country. Different cities on the same timeline are taking action independently, without necessarily coordinating with each other. This "uncoordinated" effort suggests that the driving force comes from the technology itself, rather than a call to action from a particular document. The generation of institutions is a phenomenon of the era, not an initiative of a single city.
What's truly important to see is that the system is beginning to take shape, and it's not just one party taking action, but rather an inevitable step in the evolution of the artificial intelligence era itself. Any technology, once it has grown to permeate everyday life, will naturally develop its own set of rules - this is the path that railways, electricity, and the internet have all taken. Technology advances first, followed by the development of rules, a rhythm that has been repeated time and again. If we can't understand this trajectory, we won't be able to address the issues that follow; but if we do understand it, then we can start thinking about writing our own "next chapter".
Strategic Lens: The Rule-Setting Period is a Window of Opportunity
For an organization, changes in the policy environment are phased. Initially, it's a slogan phase, where everyone is shouting support, and documents are full of directional language, but no actionable details can be found - the information quality is very light, almost all noise, and it's difficult for the organization to make judgments based on this. Further ahead is the rule-setting phase, where the framework is established, and institutions start to be written, one by one, and scattered slogans begin to gather into a readable map. Finally, there's the implementation phase, where supporting policies, execution guidelines, and acceptance standards are gradually added, and the map becomes a path that must be followed. The government regulations in Chengdu are currently in the rule-setting phase, and are a representative sample.
When the timeline is laid out, the approach to reading this chapter becomes clear. The accompanying "Several Policies" (Cheng Ban Gui [2025] No. 6) was issued as early as August 2025, with eight articles and twelve clauses, and a total funding scale exceeding 100 million. The funding scale is roughly as follows: the annual maximum for computing power vouchers is 10 million, the annual cumulative maximum for a single computing power demand side is 5 million, and the deductible amount for computing power purchase fees does not exceed 60%; model subsidies are 30% of the annual model usage fees, with a maximum of 1 million; and no more than 20 benchmark scenarios are selected each year, with a reward of 30% of the actual investment. These numbers are not just slogans, but are tangible amounts that can be recorded in a company's account books. However, the method - or government regulation - was not released until July 2026, and will come into effect on September 1. The money was allocated first, and the rules followed later. This approach of using policy documents to test the waters, allocate funds, and put practices into place, and then upgrading these practices to regulatory systems, is a clear thread and
The rules are written behind the money, and the money also verifies the direction of the rules. During the system generation period, understanding which direction the rules are heading and seeing where the money is flowing are two sides of the same coin. Looking only at documents will miss the funding choices that are already happening in reality; looking only at funding will also fail to see which rules these choices will be incorporated into and solidified by. By looking at both, it is possible to piece together a map that is taking shape - the initial subsidies point to a particular field, and the subsequent regulations will likely reserve a place for it in that chapter.
This window period, in other words, is the gap before the rules are set in stone. The gap arises because when the game rules have not been finalized, whoever is the first to understand the direction of the rules and make a move will turn uncertainty into a first-mover advantage. The scenes of unveiling, tiered regulation, and factor guarantees - these items being written into the rules - are the same formula for people in different positions. For researchers, the blank spaces where the rules are not yet defined are the most fertile ground for topics - once the details are implemented and the blanks are filled, all that's left to write is commentary. For workers, the time when the rules are being written is the best opportunity to incorporate adaptability into products - if they wait until others have finalized their positions, lists, and standards before following suit, they will only be playing catch-up. For companies, the fact that the game rules have not been determined means that admission tickets have not been fully issued - those who position themselves on the right path first will secure a first-mover advantage; once the landscape becomes fixed, the cost of entering the market will be completely different.
This window of opportunity is not meant to be waited out. Instead, it provides a chance to read the rules, set coordinates, and occupy a position before others even begin to observe. Once the rules are established, the window closes, and being the first to act becomes the established pattern. The next lens to look through is when the scene is unfolding, rules are being written, and organizations must determine their own position.
Organizational Lens: The Government's Half-Step Back Sends a Signal to All Organizations
First, let's clarify what "揭榜" (unsealing the list) means. Article 19 states "establish an open system for artificial intelligence application scenarios," and the official interpretation explains it more straightforwardly: scenario construction is shifting from "government procurement" to "market unsealing." This means the government is putting up real-world scenarios - such as public services, state-owned enterprise operations, public services, and urban governance, which are everyday challenges - as a list, and companies can come to tackle them. The government is transitioning from previously being responsible for both creating and solving problems to setting rules and conducting evaluations. For companies and research teams seeking real-world scenarios, this is equivalent to gaining a institutionalized entry point: in the past, finding pilot projects relied on connections and luck, but now opportunities have been formalized into a public channel.
The difference between "揭榜" (public bidding) and "点单" (direct assignment) lies in whether the government needs to personally intervene. If the government were to directly assign projects, it may not fully understand the true demands of each subdivided market, resulting in unfulfilled orders - the project may look good on paper but ultimately fail to attract participants. If the government were to build its own computing power, the rapid pace of technological change, with updates every six months, would lead to inherent resource allocation lags, making mismatches a highly probable event. If the government were to designate a single company to take on a project, the lack of horizontal comparison would eliminate the pressure to deliver, making it difficult to evaluate the quality of the solution, and leading to a situation where good and bad performance are treated equally. These three bottlenecks point to the same solution: outsourcing uncertainty to the market, allowing multiple solutions to compete, and having the government only measure the results at the end.
The significance of the unveiling is not limited to just the government. It is a mirror that reflects not just the government, but the same shift in all organizations in the era of artificial intelligence: their core capabilities are transitioning from "doing it themselves" to "setting rules, building platforms, and receiving feedback." The reform of state-owned enterprises, which has progressed from "managing assets" to "managing capital," is the same logic from the previous round - handing over specific operational matters and retreating to the level of capital allocation and rule-making, from managing every single transaction to managing the flow and return of capital. The government's unveiling this time is a new example of this round. Although the actions are different, the underlying principle is the same: organizations no longer prove their value by taking direct action, but rather by defining their position, building rules, and amplifying their value. This applies equally to private enterprises and research institutions - are you setting rules, building platforms, or preparing to receive feedback? There are three positions and three ways of operating, with no hierarchy, but they determine your next move.
The rules have been distributed to each department, with the sixth article listing the responsible departments and specifying "who should take the ball and how" in the institutional text. The position is not only a responsibility, but also an interface - the department under whose name the article falls will have an additional external connection channel. When the rules are distributed, whether or not to take it and how to handle it after taking it becomes a new question that each organization must answer on its own, rather than a matter for a single department.
Insights from Organizational Lenses are reflected here. Regardless of whether you are in a state-owned enterprise, a private enterprise, or a research institution, the first question forced by the AI era is not "what should I do," but "how do I define my position" - as a rule-maker, a platform builder, or a receiver. The government taking a step back is not news, what is truly noteworthy is that all organizations must re-answer "where am I" in the new coordinates.
Governance Lens: Leaving Gaps in Governance, an Unfinished Scale
Article 32, for the first time, incorporates "tiered and categorized regulation" into the framework of local regulations. It states that regulation should be based on factors such as risk level, application scenario, and scope of influence, and can be dynamically adjusted. The significance of this statement lies not in what it prescribes, but in what it excludes - it explicitly rejects a "one-size-fits-all" approach. In the realm of AI governance, this is the first time a local regulation has provided a framework: different things can be treated differently, and the same thing can be treated differently over time. In the past, we were accustomed to waiting for a final, complete, and definitive plan, whereas Article 32 implies a different rhythm: first acknowledging that things are changing, and then providing a framework that can adapt to these changes.
The framework has been established, but the details have not been fully fleshed out. Article 32 mentions "categorization by risk level and scenario," but it does not specify how the categories should be defined, who should define them, or how often they should be adjusted - not a single word is mentioned in the article. Upon seeing this, some people may instinctively think that this is a loophole, that the legislators have not thought it through. But let's not jump to conclusions. During the formation of a system, leaving some leeway is the norm, not an oversight. When a regulation is established at a time when the technology is not yet fully understood, it sets out what can be determined and leaves room for what cannot be determined at the time. This approach itself is a deliberate writing strategy. The gap is precisely the area that observers should focus on: whoever fills it in first will have a hand in shaping the rules. For individuals and organizations, the gap is not a problem, but an opportunity - a space where they can make their mark.
The latest status of this gap is worth clarifying: the official list of application scenarios in the open system has not been publicly released as of early September 2026, when this article was written. However, the gap is not completely dark - officials have revealed the general outline of open scenarios on public occasions, while the formal list remains pending. This is the most accurate description of the gap: it is not a complete blank, but rather a partially revealed, yet unfilled, space. Observers should focus on the middle ground: the distance between the verbal disclosures made by officials and the formal list being put into writing.
Shifting the focus from "how to regulate" to "who is responsible", Article 34 takes things a step further. It states that the main responsibility for tech ethics management lies in setting up ethics review committees in sensitive areas. The underlying message is that the responsibility for governance is starting to shift from "the top" down to each individual organization that uses AI. Compliance is no longer just the concern of regulatory departments, but is becoming the responsibility of every unit that uses AI. In the future, an organization will not only have to answer "whether it uses AI", but also "what arrangements it has made for using it". The pen of the institution has been handed over to the organizations from this article onwards. The shift of responsibility is actually another way of saying "leaving room for improvement", where the room is left for each organization to fill itself.
The governance lens reveals a simple truth: incomplete rules mean everyone has the opportunity to participate in writing them. However, there is a prerequisite - you must first see the gap. Those who cannot see the gap will only wait for a complete solution to be handed down, and then passively accept it; those who can see the gap are already thinking about what they can add. This ability to "see the gap" is the first stroke of the next article for AI researchers and workers. Seeing is a prerequisite for taking action; this is probably the first lesson that the governance lens leaves for everyone.
Merger: Shengfu Shou is taking a position
The three lenses in front, each looked at one aspect. The strategic lens said that the AI era has issued a window period ticket to all organizations, and those who move first will occupy the position. The organizational lens said that this ticket is not a seat number, but the position itself is shifting, and where you stood in the past is not equal to where you will stand in the future. The governance lens said that the rules are being written, with gaps left, and responsibility is being passed down. When the three lenses are stacked together, they are actually pointing to the same thing: in the AI era, the winning hand is shifting from "model capability" to "the ability to coexist with institutions".
Saying that "model capability" no longer decides everything does not mean models are unimportant. Quite the opposite—it is precisely because models will become more numerous, more powerful, and cheaper. Once a technology enters a phase of inflationary commoditization, the window for using it to create differentiation gradually narrows. A few years ago, running a strong model required queuing for compute; today, open-source models are everywhere, API prices keep falling, and capabilities are rapidly leveling out—this is what technology inflation looks like. The powerful model you hold exclusively today may be available to everyone tomorrow, at a lower price. At that point, what truly distinguishes organizations and individuals is no longer whether you have a strong model, but whether you can read the rules, adapt your business to them, and even participate in writing them. This ability to coexist with institutional frameworks is the new, hard-to-commoditize source of differentiation in an era of technology inflation.
So, what this chapter is trying to say is not that "technology is no longer important". Technology is the ticket to entry - without it, you can't even get to the table. What it's saying is that after entering, it's about something else. In this period when the rules are still being written, where you stand and which direction you're heading will determine your position earlier than "how big your model is". This is exactly what the next chapter will discuss - what our next article should be about.
Our next article
The three preceding lenses have clarified the three ways organizations survive in the AI era - setting rules, building platforms, and playing by the rules. Applying the same logic to individuals results in three different starting points. In this period when rules are still being written, for those conducting AI research and those using AI to work, it is not an uncertain period to be endured, but a rare window of opportunity. The blank space is where the pen meets the paper. This pen, when held by different people, writes different things: when held by researchers, it turns blank space into problems; when held by workers, it turns rules into capabilities; when held by observers, it turns a recent event into a trail that can be followed continuously.
Look at the pen in the researcher's hand. The "unwritten" parts mentioned repeatedly in the previous chapters are problems in themselves. Article 32 states that regulation is classified by risk level, application scenario, and scope of influence, and not managed by a single standard from start to finish; however, the article does not provide answers to how the classification is done, who does it, and how often it is adjusted. Article 19 mentions establishing an application scenario openness system, which requires a scene list, but the list has not been released at the time of writing. The blank spaces in the rules are precisely where researchers can delve deeper: while others wait for the detailed rules, researchers can make the question of what the detailed rules should look like a topic of study. Even the issue of "system generation" itself is worth exploring: how long should it take for a new technology to develop rules, and what do rapid or slow development mean - the time difference between regulations and rules mentioned in the first chapter is a ready-made topic. Visible blanks mean the topic is already there, the difference lies in whether you are willing to treat "unwritten" as "incomplete" rather than "no signal".
Look again at the pen in the worker's hand. In the past few years, compliance has been a cost in many teams' accounting books, a hurdle to clear before going online. After the rules started being written down, this account needs to be recalculated: who can incorporate the logic of risk classification into models and processes, allowing an application to clearly explain its own risk level and boundaries; who can make content compliance and data requirements a part of the product, rather than an additional layer of review, will grasp a difference that others cannot make up for in this period of rule-making. This is no longer just risk control, but positioning. For everyone using AI to get work done, the next chapter of their profession will probably start from here.
Finally, there's the observer's pen, which writes the slowest but is the most durable. The growth of the system has its own rhythm: when the supporting regulations are implemented, which scenarios are included in the first batch of the unveiling list, and how the regulatory level is ultimately determined. By putting together information from different cities, you can see that the same thing has different writing styles and rhythms - the city that first gets the details, lists, and regulatory levels to really work will be the first to pave a referable path for others. This information is scattered in different files at different times, and only when it's continuously put together does it take shape. The approach is simple: create a tracking table of your own, leave a column for supporting policies, unveiling lists, regulatory levels, and progress in each city, and add new information as you see it. For example, if a city has already disclosed that it will open up thousands of scenario demands and focus on directions such as embodied intelligence, this kind of information should be added to the table instead of being overlooked. Looking at policies has now been upgraded to tracking the growth of the system. Those seemingly effortless judgments in management case studies are often not inspirations, but the accumulation of this kind of diligent effort.
The three paths are not mutually exclusive. Researchers will encounter real problems when developing products, product developers will see the direction ahead of time in tracking tables, and long-term observers will eventually write about what they have seen. People on all three paths are actually holding the same pen.
The rulebook and writing tools have been distributed to everyone. The difference lies in whether you wait to see the rules written by others or take the lead and start writing yourself.
After September 1st
On September 1, 2026, the 247th government order of Chengdu took effect. After this day, this regulation will likely quietly sit in the list of government information disclosures, like many of the city's regulations, being scrolled past by most people without a second glance.
However, there are always some people who repeatedly look it up. Researchers and others in relevant fields are checking the list to see when the first rankings will be released and which areas they will cover; those using AI for work are examining product compliance, focusing on the phrase "tiered classification and regulation," and trying to understand which category they will be classified under; observers who plan to apply are checking the supporting policies one by one to see if their projects meet the requirements. Once the document starts being consulted, referenced, and debated, the rules are no longer just written provisions, but begin to take hold in reality.
Rules start being written, never in a grand manner. Model releases have launch events, algorithm competitions have leaderboards, while the progress of a set of regulations is hidden in repeated solicitations of opinions, revisions of details, and review meetings. It doesn't make headlines, it just slowly, one by one, changes the map that people refer to when making decisions.
The first face of technology is exciting, but the second face of rules is rearranging the seating. We are on the eve of this rearrangement - the eve is always quiet, but quiet does not mean nothing is happening. When looking back on the day the seating is rearranged, people will find that the change had already begun in the documents that were repeatedly turned over after September 1.