(The author of this article is Han Gong Observation, and Titanium Media publishes it with authorization)

Text | Korean Industry Observation

On November 5, Dark Side of the Moon rented out Harness. Instead of renting a model API - which has been sold by token for several years - it is to package and rent the "complete backend infrastructure for AI employees to work": cloud sandbox, session memory, tool chain, disconnection recovery, audit log, all included. The enterprise just distributes work, and it runs on the cloud 24/7, and it can continue from the breakpoint even if the power is cut off during the operation.

This is not a product feature release. This is a migration of pricing anchor points: AI manufacturers have changed from "selling electricity" (token) to "selling employee hours" (hosted agent). The underlying logic of this matter is far more serious than the words "cloud service upgrade" - it means that the budget items for enterprises to purchase AI are sliding from IT expenditures to human expenditures.

1. From selling electricity to selling employees: Five upward moves of the pricing anchor point

This is not the first time that humans have changed IT from "buying" to "renting". But the things I rented a few times before were essentially different from this time.

The first period is the License era. From the 1980s to 2000, when companies bought CDs, licenses, and CPU seats, they were essentially buying "copying rights" plus fixed assets. What Microsoft, Oracle, and SAP sell is a hoe. If it breaks, you can grind it yourself and upgrade it yourself. In fact, IBM tried mainframe leasing in the 1960s, and ASP hosting was also popular around 2000, but at that time, the rent was hardware or hosting rooms, and the operating environment was not priced separately.

The second period is the SaaS era. Salesforce shouted "No Software" in 1999 and made "renting access rights" mainstream. Enterprises no longer own the software and buy accounts on a monthly basis, and the manufacturer manages operation and maintenance, upgrades, and multi-tenants. Later it evolved into charging per seat/person per month - you pay how much you have to sell. The pricing anchor slides from "assets" to "access" to "seats", and CapEx becomes OpEx. But this is still renting a tool - a company buys a monthly gym membership, and whether or not to practice and what results it achieves is up to itself.

The third period is the era of AI usage. Starting from 2023, large model APIs will be billed by token, and enterprises will pay for "brainpower". But this is still the electricity sales model - you spend 1 cent to adjust the API, and the manufacturer earns 1 cent, which is the price of electricity. The company's perception is "how much money was spent", and the budget account is still listed under IT expenditures.

The fourth paragraph is now. Dark Side of the Moon Hosted Agents, Anthropic Claude Managed Agents, OpenAI Hosted Agents, global leaders commercialized "Agent runtime" almost simultaneously. What you rent is not a model or access rights, but a "job that knows how to work" - on call 24/7, you can split tasks, adjust tools, generate reports, send emails, retry if you make an error, and continue running if it crashes.

The pricing anchor point subsequently moved upward for the fifth time: from assets (License) → access (account) → seats (person/month) → usage (token) → working hours (Agent running time).

Why is it called "moving up"? Because the budget account has changed. The company originally spent 100,000 yuan to buy the CRM and 1 million yuan in sales salary; the agent directly cuts through the executive layer that is 1 million miles away. The market ceiling is not IT expenditure, but the portion of human capital expenditure that can be structured. This is not an amplification of magnitude, but a reconstruction of budget logic.

But history does not simply repeat itself. The previous times I rented tools and access rights, this time I rented "post"——A digital employee who can learn, scrap, and iterate. There is an essential difference between a digital employee and a physical asset: a chair will notOutdated, but the Agent will.

2. Dynamics: Why "things that evolve cannot be bought out"

When you buy something, the implicit premise is that it can be used stably after purchase. If you buy a chair, it will not evolve on its own, nor will it break down on its own initiative. But digital workers are different.

Dynamicity is not a matter of presence or absence, but a matter of intensity. The map app is also constantly updated, but if it is not upgraded for a year or two, the old map will still be usable, but the route will be inaccurate - this is a low-intensity dynamic that can be covered by a buyout and annual maintenance contract. If the Agent does not iterate for three months, the interface may have changed, the security holes may not have been patched, and the business system may not be connected. This is a highly dynamic situation, and it will be "wasted" without continuous investment. To judge whether the dynamics should be rented, it does not depend on the frequency of updates, but on "whether it will be useless if it is not updated."

First, it needs to evolve. The model is updated every three months. What you can do today may be done better tomorrow, or it may be degraded due to upgrades. You buy out a version and it becomes obsolete in three months. Manufacturers continue to iterate in the cloud, and the "buyout version" in your hand has become an island.

Second, it needs to be maintained. It's not "just buy it, put it there and do the work yourself". It needs to connect to new data sources, adapt to new system interfaces, fix bugs, adjust prompt words, and handle exceptions. Did you buy it out and take care of it yourself? Then you have to maintain a technical team, which is equivalent to hiring someone to look after this "digital employee".

Third, it needs to change with the business. Company processes have been changed, systems have been changed, and standards have been updated, and digital employees have to adapt accordingly. The buyout version will not automatically adapt to these changes.

So if you buy a digital employee, it’s like buying a person without giving him training, giving him a salary increase, or letting him learn new skills—how long can he work?

The essence of renting is to "keep it alive and working"The responsibility for "is thrown to the manufacturer. If the model is upgraded, the manufacturer is responsible for the update, and you have nothing to do with it; if the tool interface changes, the manufacturer is responsible for the adaptation, and you don't care; if a bug occurs, the manufacturer will fix it, and you don't have to worry about operation and maintenance; if it doesn't work well, you won't renew the lease next month, and you will have to change it. The lease naturally has an elimination mechanism, and there is no buyout.

This explains a counter-intuitive phenomenon: everyone intuitively feels that "expensive to buy, cheap to rent", but the general agent may be just the opposite - the more expensive, the more suitable for renting. Because if you buy the technology, it iterates every half a year, and the residual value will return to zero; if you rent it, the manufacturer will bear the risk of the technology becoming obsolete.

But "renting" is not one-size-fits-all. Whether the Agent has an entity, whether the data can be exported, and how fast the technology iterates determines the boundaries of renting and buying. To understand this, we need to break down the implementation form of AI into three layers.

3. Three-tier form: on-device purchase, cloud rental, and vertical hybrid

The first layer, pure agent in the cloud

This is the main battlefield of Kimi Hosted Agents - the Agent runs on the manufacturer's cloud, has no physical entity, and the enterprise only dispatches tasks. Sandbox, memory, toolchain, permissions, auditing, all managed.

Why is this kind of scene necessary to rent? Three reasons. First, there are no physical assets, and there are no problems with depreciation and residual value; second, there is no sense of technology iteration, manufacturers can upgrade remotely, and enterprises do not need to worry about it; third, general capabilities are commercialized, and the model weight itself is depreciating - the output price of DeepSeek-V4-Flash has been reduced to about US$0.28/million tokens, a single Atlas 800 can be privatized with eight cards, and general capabilities are being rolled into utility prices. You buy out a model and it becomes obsolete next year.

The pricing structure of Dark Side of the Moon itself is a signal: the sandbox is charged according to the running time (0.072 yuan/hour), the token inference fee is superimposed separately, and the plug-in is charged on a per-use basis. The company's perception has changed from "how much did it cost" to "hiring an intern who doesn't have a station and paying the bills on a monthly/per-day basis". During the same period, Anthropic Claude Managed Agents (released in April 2026) charges an operating fee of US$0.08/session hour, and OpenAI, Google, and Microsoft have also successively launched similar-level managed products. This is not the strategy of one company, but a collective shift in the industry.

The second layer, end-side equipment

On mobile phones, PCs, and NPUs in cars, the Agent runs locally. The core characteristics of this type of scenario are: hardware is an asset and intelligence is a service.

No one rents cell phones. The hardware is outdated - the chip has insufficient computing power, the memory is full, and the battery has deteriorated - it has to be physically replaced, and the manufacturer cannot "upgrade" an NPU remotely. So the equipment is purchased. But the above intelligent capabilities can be rented. Apple Intelligence, Huawei Hongmeng AI, and Alibaba Qianwen are connected to the National Bank iPhone (Beta in November 2026). Simple tasks are run locally on the device side, complex tasks are transferred to the cloud, and AI functions are subscribed on a monthly basis. The end result is a combination of "buying equipment + renting brains".

The Ministry of Industry and Information Technology has recently proposed network access management requirements for smart terminals. Mobile phone/wearable pre-set Agents must pass network access detection, OTA upgrade evaluation and permission invocation control - this means that the "rental brain" on the device side also has regulatory boundaries, and you can't just rent it if you want.

The third layer, physical carrier robot

This is the most complex mixed state and is divided into two levels.

General services - hotel food delivery, building cleaning, park patrol - data is not sensitive, equipment loss is large, technology iteration is fast, and it is suitable for renting. The health care robot will be exhibited at Wuzhen's "Light of the Internet" in November 2026. The industry is discussing the leasing direction, but it has not yet formed a large-scale commercial verification. The principle is the same as when an enterprise rents a printer: the manufacturer is responsible for maintenance and upgrades, and the enterprise pays according to use, and does not bear the risk of technological obsolescence.

For vertical precision categories—surgical robots, industrial robotic arms, and bank lobby intelligence—the data does not go out, which is a hard constraint and requires on-site training. New Hope Financial Technology and Yushu Technology built a bank hall solution based on the G1 humanoid robot. The computing power is deployed locally and does not connect to the public network. The design principle disclosed by the manufacturer is "no touching customers, no touching funds, and no touching data." This kind of scenario is not purely renting, nor is it purely buying, it is "Buy the main body + local training + maintenance service"A mix of equipment ownership is bought out, while model capabilities and on-site services are leased for continuous iteration.

The fourth level, sovereign level: non-rental chassis

There is another type of subject that cannot be discussed within the "rent-to-buy spectrum" - the state and the government system.

The national government affairs platform does not want to be rented, but its sovereignty cannot be outsourced: the computing power is provided by the national integrated computing network or the Eastern and Western computing nodes, the model is trained independently or the filing model is used, the data does not leave the domain, fine-tuning is kept in the ledger, and Harness is built in-house.

The "Guidelines for the Deployment and Application of Large Models of Artificial Intelligence in Government Affairs" issued by the Cyberspace Administration of China and the National Development and Reform Commission in October 2025 clearly stated: unified recruitment, reuse of superiors at the county level, no duplication of construction, and priority for filing models. In October 2026, the Ministry of Commerce plans to add a list of export restrictions for AI large models/agents/computing power cluster technologies, further making the boundaries of technological sovereignty more explicit.

They will also use commercial agent capabilities, but only in non-confidential general auxiliary layers - official document drafting, Q&A, document proofreading; core decision-making, confidential data, and regulatory judgments. They must "buy to the end": buy computing power, buy models, and buy controllable iteration rights.

Therefore, the real world is not "renting instead of buying", but three levels in parallel: rental in the general market is the mainstay, vertical industries are mixed, and self-construction in the sovereign domain is the bottom.

But the frame is empty. In the real world, what both renters and buyers are betting on is not price, but trust and sovereignty.

4. Game: What you rent is not a tool, but an agent

In traditional principal-agent theory, when companies hire people, they are afraid of people being lazy and hiding information. When renting an agent, you are afraid of hallucinations, overstepping of authority, prompt word injection, and value deviation—you ask it to review the contract, and it misses clauses that should not be deleted; you ask it to generate a report, and it compiles a set of false data.

A closed-door study on Agent trust mechanisms by the University of Hong Kong showed that 76% of practitioners said they had encountered trust concerns in actual deployment, 53% believed "low customer trust" was the main obstacle, and 42% believed "clear responsibility for errors". These three sets of numbers say three things:Frequency of concerns, customer perceptions, and subsequent attribution. Enterprise procurement of Agents is not essentially about buying capabilities, but about "entrustability" - the identity can be confirmed, the authority can be restricted, and the performance of the contract can be audited.

This leads to the data sovereignty game. Cloud Pai said: Hosting is the most worry-free, and it’s okay if the data passes through the cloud. Inference Hook can intercept sensitive fields. The local faction said: Finance, medical care, and government affairs are all vetoed. The bottom line is that the data does not go out of the domain. The hybrid faction is taking the middle road: sifting data on the device and strengthening the brain in the cloud. Honor's "Symbiosis Agent" and NVIDIA × Perplexity all follow this logic.

What enterprises strive for is commercial sovereignty, while what the country defends is data sovereignty plus technological sovereignty. The Intelligent Self-Discipline Convention released by Wuzhen in November 2026 proposed that "identity can be verified, authorization can be controlled, behavior can be investigated, and responsibilities can be traced." The national blockchain network serves as a trust anchor for Agents-this just shows that when Agents change from tools to agents, "entrustability" is not only a business issue, but also a governance infrastructure issue.

The real demands of manufacturers are opposite to those of users. Users want to "increase their capabilities and keep the original data away"; manufacturers want to "see more failure trajectories and better iterative models". Tencent would rather give up the original data and replace the failure log - because as long as the data does not go out, "failure trajectory" is the new oil for future training data.

The second dimension of the game is pricing power. Manufacturers want to charge based on results (RaaS), transfer the model cost risk to themselves, and exchange for higher unit price per customer. However, companies are concerned: Agent error rates are still high, and if the results are not attributable, they dare not pay according to the results. A UC Berkeley 2025 comparative study across seven multi-Agent frameworks showed that the multi-Agent task failure rate ranged from 41% to 86.7% (different benchmarks cannot be compared horizontally); the CMU TheAgentCompany benchmark measured that the failure rate of the top single Agent in real office scenarios was about 69% to 70%. Pay by results? Wait for stability to pass first.

The reality is: first rent a workspace on a monthly basis, then charge by working hours, and finally dare to pay by piece in a narrow scenario. There is a hurdle of "stability" in the middle.

The third dimension of the game is lock-in. Renting is worry-free, but if you rent for a long time, you will be tied up. After Agent is embedded in the workflow, business knowledge is deposited on the manufacturer's side, and memory, configuration, and plug-in ecology are all on Renjia Cloud. Rent Kimi today, want to move bean bags tomorrow? Data assets cannot be taken away, and switching costs are measured in years. Big companies are already vying for entrance: Doubao Tunfeishu and WorkBuddy are tied to the WeChat ecosystem, which is essentially "data sovereignty locking".

Therefore, the bottom layer of the rent-to-buy decision is not cost comparison, but risk allocation. Rent, externalize iteration risks and operation and maintenance costs to the manufacturer, but in exchange for lock-in risks; buy, hold data sovereignty and exit freedom in your hands, but bear the risk of obsolescence and the cost of raising a team.

If a manufacturer wants to charge based on results, the company should first ask: Who is responsible for the mistakes?

5. Opposition: Not everyone rents, and not everyone can pay based on the results now.

The core judgment of this article is that "the general scenario is the main channel", but this judgment has two limitations.

First, privatization is becoming cheaper, and the third path really exists. DeepSeek’s open source DeepInfer inference framework (October 2026) allows enterprises to build their own inference clusters, significantly reducing unit token costs in high-throughput scenarios. Daily Interactive launched a pure privatization solution for "Gizhi Workstation", using Huawei's Ascend architecture, which is known as "twenty times cost-effective". In fact, the implementation caliber of intelligence is more specific: in high-frequency production scenarios (an average of 100,000 calls per day), the comprehensive cost of the privatized DeepSeek solution can be reduced to 30% to 40% of the public cloud API.

But there is a premise: the computing power utilization must be high. A TCO calculation by AgentMarketCap in April 2026 pointed out that when the utilization rate is higher than 70%, the long-term cost of self-construction is low; below this line, hardware depreciation and operation and maintenance manpower will exceed cloud hosting. The bottleneck is not in the model itself, but in the engineering labor costs of agent orchestration, evaluation, and safety guardrails.

Second, leading companies do not rent but build their own. Tencent's internal CodeBuddy covers 95% of engineers, and WorkBuddy is an AgentOps platform; Byte, Baidu, and Huoshan each build an Agent middle platform and add ecological binding. Large companies have computing power, talent, and data. Renting is the main option for small and medium-sized enterprises, and self-building is the option for leading players.

Third, RaaS pays based on results, which only works in narrow scenarios.. Zhida Mingyuan AI Marketing RaaS writes the results into the contract, provides refunds if the standards are not met, and has a renewal rate of 96.8% - but only in scenarios where the results are easy to quantify, such as marketing customer acquisition. SDIC’s RaaS pilot program of 100 smart security robots is still in the “investment verification period.” Complex enterprise tasks are returned to "time rent" because the conditions of "errors attributable and results quantifiable" are not yet met.

So a more accurate statement is: renting is a trend, but buying will not disappear; RaaS is the ultimate direction, but there are two narrow gates of "stability" and "accountability" in between. Once the pricing anchor is switched from IT budget to human budget, Agent is not just a procurement issue, but a labor system issue.

The real problem is: when companies start to use "human resources budget" instead of "IT budget" to measure AI, the ceiling of the entire market changes - but the change is the structure, not the total amount.

6. Human budget replaces software budget？Social tolerance is the ultimate variable

This is not a continuation of SaaS, but a subversion of SaaS.

The essence of SaaS charging per seat is still tool logic - whether you use it or not, you will be charged per person. However, Agent charges based on working hours, tasks, and results, which is essentially labor logic.

However, what the Agent is looking at is not the total amount of the "manpower budget", butThe layer within the labor budget that can be structured. The tasks it can take on have three characteristics: repetitive, tireless, and rational—organizing clues, sending follow-up emails, filling out reports, reconciling accounts, and publishing morning reports. These tasks would be annoying for people to do, but not for agents. They are on call 24/7 and don’t get upset if you don’t ask for leave.

It measures the working hours of the executive and assistant levels, not the relationship and decision-making levels. Negotiations require reading expressions and tones, drinking is about building informal trust, and on-the-spot decision-making relies on field awareness—these "human presence" tasks cannot be done by agents, nor should they be done.

So it is not "the human budget is replaced", but "the human budget is internally re-layered":The execution layer is partially eaten by the Agent, while the relationship layer and decision-making layer are more valuable., because man frees up his hands to do what only man can do.

This analogy can be pushed down: now charging based on length of time plus token is equivalent to calculating wages based on working hours, and you will be paid according to the amount of work you do; in the future, paying based on results will be equivalent to piece-rate wages, and you will only be paid after the work is done. Enterprises are willing to pay this money because they save real money and labor costs at the executive level.

What is the market implication of the upward movement of the pricing anchor point? It is the company's psychological accounting for AI, which has changed from "buying a better hoe" to "hiring a part-time worker who does not get off work". You can buy a hoe, but you can only rent a day laborer. As long as renting is cheaper than hiring people, this logic will not change for a hundred years - what will change is the unit of measurement: from "monthly work station fee" to "wage per piece of work".

But buried here is a problem that is more difficult than technology: If the agent is not cutting the IT budget, but the executive level in the human budget, where will the cut manpower go?

The individual rationality of enterprises is very simple: use whomever is cheaper, and use Agent if it is cheaper than human beings. But the reflexivity of the system lies in this: if the executive level is replaced on a large scale, the consumer side shrinks, and the company's products cannot be sold, and it will have to hire people again. In the long run, it is not "infinite substitution";Dynamic balance of man-machine ratio. The business model of "renting an agent" is still mainstream in the equilibrium range - because renting is the most flexible, if the business is good, you can increase the rent by a few, and if the business is bad, you can cancel the subscription.

In 10 years, general office agents will be fully rented on a monthly basis, and will gradually switch to billing based on results; in 20 years, digital employees will have become a standard employment form, and the company's human resources structure will be equal to regular employees plus leased silicon jobs plus outsourcing. But before that, there is one thing more urgent than technology iteration: when the pricing anchor point of AI slides from software budget to human budget, where will the released executive-level manpower go?

The National Development and Reform Commission’s press conference on November 3 has put the AI ​​regulatory sandbox, employment structure monitoring, and the new career of model trainers on the table - the distribution issue is not a distant deduction, but an ongoing agenda.

Whether the allocation problem will arrive first and in what form does not depend on technical deduction, but onsocial tolerance. Technology puts pressure on the table. Whether to respond, how to respond, and when to respond are matters for society to weigh. This trade-off includes economic considerations, stability considerations, and employment considerations, which cannot be calculated by technology.

Technical problems always have solutions. How society accepts this solution is the hurdle that will come first.

【Information description】

This article analyzes the concept description:

1. "Dynamic": This article uses it to refer to the characteristics of Agents that continue to evolve, require maintenance, and follow business changes, in order to demonstrate that buyout is not economically feasible. This is a concept analyzed by the author, not a common term in the industry.
2. "Dynamic intensity": The criterion proposed in this article is used to distinguish the rental and purchase decisions of different technology products - low-intensity dynamics (such as map apps) can be covered by a buyout and maintenance contract, while high-intensity dynamics (such as Agent) can be covered by leasing. This is the author's analysis concept.
3. "Pricing Anchor": This article is used to describe the migration trajectory of budget account ownership when enterprises purchase AI (from assets to access to seats to usage to working hours to results). This is the author's analysis concept.
4. "Three-layer form framework": This article divides the AI ​​implementation form into three layers: pure agent in the cloud, end-side equipment, and physical carrier robot, and corresponds to different rental and purchase decisions. This is the author's original analytical framework.
5. "Sovereign-level intelligence": This article supplements the analytical concept and refers to the deployment form of computing power/model/data/Harness in government affairs, state-owned assets, and critical infrastructure that cannot be outsourced. This is the author's analysis concept.
6. "Harness": refers to the Agent running framework (session state preservation, cloud sandbox, tool invocation, disconnection recovery, retry current limit, etc.), which is a technical term in the official product documentation of Dark Side of the Moon.

Deduction prediction description:

The judgments in this article about "the pricing anchor slides from IT budget to human budget", "general scenario renting is the main channel" and "RaaS needs to pass through the narrow gate of stability" are all industry deductions made by the author based on public information and do not represent inevitable results. The vision for 10/20 years from now is a long-term deduction and there are major uncertainties.

Analogy explanation:

1. IDC predicts that "China's enterprise-level AI agent market will grow from 21.2 billion yuan in 2025 to 332 billion yuan in 2029." This is quoted from 36Kr Research Institute's "2026 China AI Agent Industry Development Research Report" (September 2026). The original IDC report has not yet been made public and the original text can be independently searched.
2. Data such as "76% of practitioners have encountered trust concerns" are quoted from KYA closed-door research at the University of Hong Kong. The full text has not been published and is for analogy reference only.
3. The multi-Agent failure rate is 41%~86.7%, quoted from UC Berkeley Cemri et al.'s paper Why Do Multi-Agent LLM Systems Fail? (arXiv:2503.13657, 2025), which cannot be compared horizontally across different benchmarks; the CMU TheAgentCompany benchmark (arXiv:2412.14161) measured the failure rate of the top single Agent at about 69%~70%, and the two cannot be mixed.
4. DeepSeek-V4-Flash pricing, DeepInfer framework, daily interactive "personal knowledge workstation", real smart implementation standards, etc. are all publicly disclosed by manufacturers or industry estimates, and have not been independently verified by the author.
5. 月之暗面Hosted Agents定价（沙箱0.072元/小时等）引自月之暗面开放平台官方定价页；Anthropic Claude Managed Agents定价引自其官方公告（2026年4月8日）。
6. 政策文件引自中国政府网公开公报：《国务院关于深入实施"人工智能+"行动的意见》（国发〔2025〕11号，2025年8月26日印发）。
7. 工信部终端进网管理要求为行业公开报道信息；乌镇智能体自律公约及区块链信任锚点为公开会议信息；发改委11月3日发布会信息为公开新闻信息。

本文基于公开信息独立撰写，文中涉及对企业的分析与判断，均为作者基于公开资料的研究性观点，不构成投资建议，亦不代表对相关企业的任何立场。相关数据以企业官方披露为准。文中对无公开证据的假设性推演已标注说明，读者请注意辨别。

更多精彩内容，关注钛媒体微信号（ID：taimeiti），或者下载钛媒体App