Under the Anti-Moore's Law, no one can earn easy money without any effort.

In a café in Beijing, terms like "pretrain", "agent" and "hallucination rate" can be heard every few tables.

These people work in several office buildings no more than two kilometers away, and their product iteration cycles have been shortened to be calculated on a daily basis.

An investor reviewed his investment portfolio on Jike and found a rule that even surprised himself: almost all the AI companies with the best investment returns are located in a box of less than ten square kilometers, "south of Tsinghua University, east of Peking University, west of Xueyuan Road, and north of Dazhongsi".

Those high-quality companies that he failed to invest in are also gathered in this area.

This is the reality.

In the AI era, the iteration clock of core production factors has changed completely.

Moore's Law once set a stable rhythm for the entire technology industry: chip performance doubles every 18 to 24 months, and everyone can follow this rhythm to make product plans, financing arrangements and supply chain management.

But in the era of large models, the refresh cycle of algorithms, data engineering and prompt strategies is advancing on a weekly or even daily basis.

When the cognitive iteration speed at the software level far outpaces the physical constraints at the hardware level, the underlying logic of the industry changes from "seeing who can make more sophisticated products" to "seeing whose learning clock runs fast enough".

This phenomenon can be called "anti-Moore's Law".

It does not describe the decrease in cost of performance improvement, but the accelerated depreciation of cognitive assets.

A fine-tuning strategy that a team was using last week may have been directly replaced by a new open-source model this week; a carefully polished prompt template may become invalid due to an upgrade of the model interface; a paper that just came out of the lab may be disassembled, reproduced, packaged into a product and launched on the market within a week, and then forgotten a month later.

Under this rhythm, time has become the cruelest dimension of competition. Today, we try to re-decode the impact of this clock transition on the geographical distribution of startups, the sustainability of business models, and the game between capital and regulation.

Why Geographical Density Dominates AI Entrepreneurship Again

Remote work was once regarded as an inevitable trend of the digital economy.

During the epidemic, SaaS collaboration tools made "distributed teams" a widely recognized concept, and it seemed that geographical constraints had been completely lifted by technology.

However, the practice of AI entrepreneurship gives the opposite answer: the more an industry is driven by cutting-edge cognition, the higher the value of physical proximity.

Why does this phenomenon occur? When Marshall discussed industrial districts at the end of the 19th century, he noticed that after enterprises in the same industry gathered together, some kind of "industrial secret" seemed to permeate the air. Even without formal contracts, knowledge would flow in the region just like breathing.

Later people used knowledge spillover and informal information networks to explain this phenomenon, but their models usually assumed that the friction of information transmission is constant.

The special feature of the AI era is that the value decay period of information itself is sharply shortened, and any time lag and distortion in the transmission process will be compounded into a competitive disadvantage.

A discussion post on an open-source community may have derived a new technical consensus three hours after it was released.

A researcher at a company accidentally overhears a discussion about a solution to training instability at the next table during lunch. He can verify it after returning to his workstation in the afternoon, and integrate it into the experiment in the evening.

This efficiency of "aligning cognition with a single turn" is almost impossible to achieve in online collaboration, because online communication requires scheduling time, organizing language, and waiting for replies, and every link is consuming the validity period of cognitive assets. What is transmitted online is information, while what is exchanged face to face is the "semi-finished cognition" that has not yet been encoded, which is intuition, hunch and tentative judgment. These unformed ideas are the core raw materials that truly drive model iteration.

Moreover, this box does not only have an adequate supply of talents. Top universities and scientific research institutions in Beijing are indeed densely distributed here, but what really constitutes the "accelerator effect" is the positive feedback closed loop formed among capital, talents and technical consensus.

Within a radius of three kilometers, there are more than 30 VC institutions, more than ten government guidance funds, and resident offices of multiple overseas funds. When a founder has a new technical judgment, he can visit three investors on foot on the same day, and get informal capital feedback that night, and these investors will most likely pass this signal to other invested enterprises the next day.

The circulation speed among information, capital and talents is difficult to replicate in other regions.

This is the high-frequency resonance of "cognition-resource" convection.

Another deep consequence brought by this geographical agglomeration is that the screening mechanism for startups is moved forward to the site selection decision.

Within this box, enterprises are almost in a passive "accelerated survival" state: the teams around are iterating on a daily basis, and you cannot slow down; your competitor next door has obtained the latest open-source weights and started fine-tuning, but you are still waiting for cloud synchronization; the intern downstairs has heard of a breakthrough idea from a certain laboratory, but you have no channel to get the information.

In this environment, teams with poor adaptability will be quickly eliminated, and the remaining organizations are naturally screened in terms of perceptual sensitivity. The low survival rate in the region itself is pushing up the overall success probability of the region in reverse. The surviving companies already have a higher "cognitive basal metabolic rate" than enterprises outside the region.

The Pleasure of Arbitrage and the Real Time Barrier

When most AI companies are desperate to pursue "speed", another quite absurd phenomenon has emerged: in the AI entrepreneurship boom, the most stable, most direct and quantifiable source of profit is unexpectedly account reselling and resource reselling.

A Jimeng premium account was hyped up to nearly 20,000 yuan on Xianyu, seven times higher than the price at the beginning of the year.

Computing power points, GPT mirror site accounts, and batch-generated content materials, all AI resources that can be priced are being traded rapidly. A complete gray industrial chain has taken shape, ranging from "AI account opening studios" to "anti-AI detection de-traces services", with clear division of labor and rapid response.

From a purely commercial point of view, this is an extremely keen market discovery.

The training and reasoning costs of large models are rising, and the gradient of usage thresholds has created arbitrage space, and these small and medium players have filled the gap at an extremely fast speed.

However, this precisely exposes the common "computing power dependence" in current AI entrepreneurship.

Data shows that 89% of core businesses of domestic AI startups are directly dependent on the supply of GPU computing power, and most business models stay at the stage of calling public model APIs plus general frameworks for shallow scenario packaging.

Once the model capabilities are updated, these small businesses built on the characteristics of old versions will be silently crushed.

Here comes a distinction that needs to be carefully identified.

The media is used to generalizing "speed" as a single concept, but there are two completely different natures of speed in the industry.

One is the speed of arbitrageurs: they capture the window of information asymmetry, and exchange speed for price difference. The life cycle of this kind of speed is determined by the speed of information diffusion. Once more people realize the same opportunity, the profit will go to zero. The current boom in account reselling is exactly the peak of this nature, and also the harbinger of its decline.

The other kind of speed belongs to learning organizations. Its core is not to take the lead once, but to continuously absorb new cognition and transform it into product iteration. This kind of speed requires profound accumulation, which requires the organization to have built pipelines for digesting new knowledge internally, otherwise even if the most cutting-edge information is obtained, it cannot be transformed into valuable output.

Many AI content creators are facing a typical vicious circle: they invest 45% of their budget in AI tools, but the human input only accounts for 9%.

The tools are getting more and more powerful, and the output is more and more massive, but the content reach rate is less than 0.3%, and most people earn less than 3,000 yuan a month. If a creator finds that working on commercial orders is not as profitable as selling his account, it means that his labor value has been compressed to the limit of tool substitution under the current model.

Tools have lowered the entry threshold for production, but at the same time, they have also collapsed the price system on the supply side.

"AI content" without barriers has become a standard product, and the market price of standard products will eventually tend to the marginal cost, which is close to zero in the AI generation era.

What can really fight against this involution is usually summarized as "the sense of real people".

But I prefer to define it as "non-compressible experience stock".

A practitioner who has been deeply engaged in the maternal and infant field for ten years has an intuitive judgment on the crying characteristics of infant colic, and she knows which kind of anxiety is a real existing demand rather than a traffic gimmick.

When AI tools pull the basic content production to zero threshold, this judgment embedded in personal experience has become the most scarce asset instead.

Similarly, if an industrial inspection AI company masters the holographic defect database of a specific production line over three years, no general large model can bridge this data gap in the short term.

None of these barriers can be built in a day. They require time to settle, and time is exactly what those arbitrage players are most reluctant to invest in.

The real barrier is always to accumulate the thickness of experience in a vertical section that others cannot level instantly by purchasing computing power.

Capital and Regulation: Two Slow Constraints for Fast-growing Companies

Facing the dazzling iteration speed of AI companies, the capital market responds by accelerating capital injection to lock in the window period.

In 2026, artificial intelligence accounted for 67% of early-stage financing, and the Series A financing generally showed a "++" trend: institutions could not wait for the next round, and directly increased their bets continuously in the Series A round, for fear of missing the opportunity.

An investor even said bluntly that when reviewing projects now, if he encounters a startup in a non-AI field, he will subconsciously ask "why can't this be done again with AI".

The FOMO sentiment of capital has reached its peak, which in turn forms a "valuation siphon" in high-density areas, and the head resources and attention are further concentrated.

However, there is a subtle misalignment between the acceleration logic of capital and the real growth rhythm that AI companies need.

The timetable for capital to demand returns is usually deduced backward under the constraint of the fund duration.

And cognitive accumulation has its own rhythm, which requires trial and error, redundant exploration, and exploratory research that seems to produce no direct commercial return. When an AI startup raises huge funds and its valuation is pushed to a level where it must quickly prove its commercialization ability, the founder can easily fall into the temptation of realizing "research capabilities" into "project delivery". It starts to take a large number of customized demands, and is dragged into non-standard scenarios for adaptation. The cognitive bandwidth of the team is fragmented, and it gradually loses its sensitivity to the cutting-edge.

This organization may seem to have growing revenue, but its cognitive depreciation rate is accelerating in the dark.

When the next wave of architecture-level changes comes, it finds that it has no extra mental space to digest the new paradigm.

If the speed of capital is active, then the slowness of regulation is more structural.

The rhythm of legislation and rule-making in any country is based on the year as the basic unit, which forms a sharp contrast with the iteration of the AI industry on a weekly basis.

In 2026, the total number of AI-generated videos and audios has exceeded 2 billion. The highest-earning AI content channel on YouTube can reach an annual revenue of 2.65 million US dollars. At the same time, low-quality AI spam content floods the recommendation feeds. Platforms have begun to suspend the monetization eligibility of AI creators on a large scale, but this campaign-style rectification cannot replace a stable rule framework.

Under the uncertain regulatory expectations, content producers are more inclined to take short-term behaviors: quickly register accounts, quickly monetize, and quickly exit, which further poisons the entire ecosystem.

The difficulty of governance does not lie in "whether to regulate", but in what time dimension to regulate.

If the traditional command-and-control regulation is adopted, which lists in detail what is allowed and what is prohibited, then by the time this set of rules is promulgated, the technology has already evolved to areas not covered by the rules.

A truly effective way may be to establish a set of "principle frameworks" that are adjusted at low frequency but with high binding force, shifting the focus of regulation from pre-approval to post-event responsibility traceability and transparent labeling.

Regulators must realize that their feedback clock is naturally slower than the industry. Therefore, the key to rule design is not to "keep up fast", but to draw a bottom line that cannot be crossed no matter how the technology evolves, such as real-name traceability for deepfakes and mandatory assessment of general models that may cause systemic social risks.

Once this framework is established, it needs to have a stability expectation of more than five years, so that companies can iterate at high speed within a predictable responsibility boundary.

However, one element is that the rhythm of regulation and capital are not irrelevant. When regulatory expectations are unclear, capital will prefer light-asset, low-precipitation, easy-to-exit models, which is exactly the financial soil for the proliferation of many current arbitrage-type AI businesses.

Once the long-term rules are clarified and the violation cost becomes quantifiable, capital will recalculate the risk premium, and funds will naturally tilt to teams with compliance capabilities and willingness to make heavy investments.

The guiding role of the government and state-owned capital is crucial in this link.

In the past two or three years, many local governments have set up AI industry guidance funds. If these "slow money" can be targeted to match talent teams that carry out long-term infrastructure construction and basic research, they can play a role in calibrating the rhythm in the capital market.

Therefore, the current trend is that compliance capability itself is becoming the core asset of AI companies.

Companies that can prove they have systematic capabilities in data governance, model interpretability, and output security will gain structural premiums in large enterprise tenders and government procurement.

This advantage cannot be established in the short term by pouring money. It depends on process accumulation and team culture, which exactly requires a "slow" organizational philosophy.

At a time when the whole world is pursuing extreme speed, companies that can resist the obsession with speed and take the initiative to slow down in necessary aspects are more likely to survive the cycle.

Final Remarks

In the process of technological development, the same story repeats itself: in the early stage of every general-purpose technology boom, the first group of people who get rich are those who sell shovels and those who act as middlemen using information asymmetry.

In the railway boom of the 19th century, people who speculated in land made money faster than those who operated railways; in the Internet bubble era, players who resold domain names and traffic once lived better than companies that made products seriously.

AI is no exception. The bubble of account reselling and resource reselling is an inevitable product of this stage.

But history also shows that after the infrastructure is popularized and the information gap is smoothed out, the enterprises that remain are always those that transform technical capabilities into real improvements in vertical scenarios and build organizational knowledge barriers.

The most fundamental change in the AI era is that its iteration pendulum is no longer adjusted by the fixed cycle of hardware updates.

It requires an organization to have two seemingly contradictory capabilities at the same time: extremely high agility in external cognitive perception, and extremely strong concentration in the accumulation of internal core capabilities.

Slow perception will be eliminated, and thin accumulation will not last long either.

The cruelty of "anti-Moore's Law" is that it makes the leading position in any stage highly unsafe. Only by making learning itself the skeleton of the organizational structure and making experience density the business moat, can you find your own constant upward curve in the accelerated depreciation of others.

This is probably where the next batch of great companies really start.

This article is from WeChat official account "Dongzhen Strategy", author: Dongzhen Strategy, published with authorization from 36Kr.