---
title: "Customers Want You to Fly Them to the Moon, but Only Pay for Same-City Delivery: Four ToB Veterans Lay Bare Enterprise AI's Pitfalls"
author: "Unique Research"
sourcePublication: "Unique Research Substack"
originalPublishedAt: "2026-07-22T10:25:23+00:00"
canonical: "https://ffcap.cn/en/research/src-20260722-01html"
source: "https://uniqueresearch.substack.com/p/src-20260722-01html"
language: "en"
---

# Customers Want You to Fly Them to the Moon, but Only Pay for Same-City Delivery: Four ToB Veterans Lay Bare Enterprise AI's Pitfalls

[![](https://substackcdn.com/image/fetch/$s_!fzCR!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2278dba0-abb1-4fe6-9f2a-008725949e73_2048x1152.png)](https://substackcdn.com/image/fetch/$s_!fzCR!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2F2278dba0-abb1-4fe6-9f2a-008725949e73_2048x1152.png)

_Original · Unique Research · 2026-07-22_

_Editor's note: The first-person report and its judgments belong to the original Chinese author. This English rendition retains the opening narrative, four themed sections, and the complete panel transcript including all named speaking turns. Customer-case figures, deployment claims, and speaker-reported customer/team counts are self-reports attributed to the named speakers, not independently verified findings. Company and person names are preserved as source attributions._

(Source: Unique Research)

AI Industry Observation

"Customers Want You to Fly Them to the Moon, but Only Pay for Same-City Delivery": Four ToB Veterans Lay Bare Enterprise AI's Pitfalls

"Technology stopped being the bottleneck long ago — the bottleneck is elsewhere."

"Customers hope you'll help them fly to the moon, but only want to pay same-city courier prices."

At an enterprise AI roundtable during Unique Bloom, Kyligence founder Han Qing borrowed this line from GeekPark's Zhang Peng, and the room laughed. The four ToB veterans on stage served hundreds to thousands of enterprise clients each — everyone had been shot by this bullet.

Han himself had a ready example. Two days earlier he accompanied a salesperson to meet a client whose CIO had decided to build data-direction AI. Han asked two questions. First: can your 2TB of data live in the cloud? No — too sensitive, it must stay on the intranet. Second: how many GPUs do you have? The CIO was pleased: four. Then added — I want everyone in the company to use AI.

2TB of core data locked on the intranet, four GPUs, and company-wide AI use — set together, that's an equation with no solution.

"But this is the most real situation of enterprise AI deployment today: technology stopped being the bottleneck long ago; the bottleneck is elsewhere."

Besides Han, the four guests were Cheng Kaizheng, founder of Shulie Tianxia (DataHunter), Cen Runzhe, AI product lead at Shushi Technology, and Sun Linjun, founder of Shizai Intelligence. They share one identity: none are AI-native founders; all are enterprise-services veterans who fought through the big-data era. Sun Linjun's Shizai Intelligence already has over a million digital employees deployed at clients; Han has built data foundations for over a decade, and many of China's large banks and insurers are his clients.

Precisely because they stepped into every pit of the previous era, when they talk enterprise AI, there's a rare honesty.

Technology Is Democratized, but Buyers Aren't Ready

The big-data-era ToB business was comfortable. Clients knew what they wanted — like buying a car, budget and performance were on the table, you just picked.

The AI era changed everything. Clients don't know what they want, and you don't know what you can give them. "It looks like you can do anything, and clients want everything."

Han says many clients now have the mindset of "I'll run an experiment first, you go make money elsewhere." The problem is every client thinks this way, and in the end it's the investors who get burned.

The more common scenario is: a leader snaps his fingers to do AI and lists a hundred scenarios at once; IT gets halfway and finds it can't get server resources. The business departments then complain that the AI IT built is useless — not accurate enough, impossible to actually use.

"The old centralized project-building approach has failed in the AI era. Han's judgment is that there's no standard answer right now — you can only keep trying. But if the early foundational work isn't done well and the trial results come out poorly, clients won't keep investing, which is waste for both sides."

POCs Look Beautiful, but No One Dares Sign Off at Launch

Cheng Kaizheng added another angle: inside enterprises today, the boss is the most anxious, while the business departments underneath are actually less urgent.

When the boss gets anxious, IT, to deliver, lists several pages of Agent scenarios. Then a very typical symptom appears: the POC (proof of concept) looks beautiful, but at the launch step, no one dares sign off.

Because everyone asks the same question: "Is this Agent reliable?"

"Shulie Tianxia's solution is plain: don't use perfection as the standard, use a human as the standard. How did the original person in this role do the job? What score did they get? Humans aren't 100% reliable either — they have moods, fluctuating states, all kinds of interference. Use the same evaluation system to test the Agent; if it doesn't lose to a human, it qualifies for launch."

Like Tesla's self-driving — at first no one dared ride, but data proved machines drive more reliably than humans, and then it smoothed out.

When Sun Linjun was doing algorithms at Alibaba, he validated the same logic: machines can't be free of hallucinations, but humans make mistakes too. The launch standard isn't "zero errors" — it's "reach human-comparable level." The concrete method is to break the process apart, see which stages allow hallucination (a bit more diversity is actually good), which stages must be precise, then validate across dimensions.

Cheng also had a stinging line: generic products are becoming less and less valuable, because the technical bar has dropped. What's truly valuable is your understanding of the scenario, and the semantics and ontology accumulated in that scenario.

The Truth About Hallucination: You're Forcing AI to "Guess One"

On hallucination, three technically trained guests gave answers pointing in the same direction, just phrased differently.

Cheng's phrasing is the most vivid. Why do large models hallucinate? Because the data and semantics aren't ready, yet you're forcing it to hand in the exam. "Like an exam where the teacher says, if you don't know it, guess — that's the AI's logic too."

The solution is to turn a closed-book exam into an open-book one: the enterprise prepares deterministic data and knowledge, and AI only does reasoning. He says that based on their deployment results, after this engineering treatment, hallucinations can basically disappear.

There's a frequently overlooked precondition here: semantics and ontology — i.e., "how AI understands your enterprise." A large model has eaten the entire internet's corpus, but it doesn't know your company's internal abbreviations, shorthand, and processes. Retail's Sell-in, Sell-out, the various MDs, TDs — if you don't teach it, it's a new employee who never went through onboarding.

Sun Linjun also brought a pit he personally stepped in. He previously used OpenClaw to scrape data and generate reports, then found the numbers in the report were off. Tracing back, the cause was both funny and exasperating: the screenshot resolution was too low, AI couldn't read it accurately, so it made up some data and slipped it in.

"Large models not only hallucinate on their own — they hallucinate when driving tools too, and they won't proactively tell you 'I can't see clearly.'"

Sun's solution is to break the task chain finely: the longer the task chain and the more compressed, the higher the hallucination rate. Better to break it into detailed workflows, boxing the large model into a specific range. For example, financial invoice review — don't throw raw data at it directly; have it write rules and pseudocode, then let a deterministic program validate, and accuracy jumps.

He also shared an example from his own company: the finance department used a new product and built an Agent in a few sentences, whereas previously an engineer spent a month and still didn't satisfy him. The engineer copied the logic by rote but had no "finance thinking," didn't understand the terminology, and kept reworking. Put simply, the industry knowledge the large model now holds exceeds the average engineer.

Han's phrasing is the most concise, good enough as a slogan: don't make the large model do the core computation. AI handles expression; the platform handles accuracy.

Computation goes to a stable data platform; the large model only does what it's best at — reorganizing dry data into the secretary-style reports the boss likes. When clients see a beautiful report, the first reaction is "is the data right?" the second is "can I verify it?" So Kyligence built a whole verifiable system starting from data lineage tracking, and is exploring using AI to rapidly validate AI.

"In the end, what cures hallucination was never a bigger model — it's finer engineering."

From Tool to Colleague: Agent Evolution's Next Stop

After "trustworthiness," Cen Runzhe pulled the gaze further out: after trust, then what?

His observation is that in early 2024, everyone felt a smart-Q&A Chatbot usable on a webpage was enough. But after tools like OpenClaw appeared, the form changed. On the web, AI is just a tool; but when an Agent moves into the enterprise IM, gets pulled into group chats, participates in collaboration, it goes from "tool" to "colleague."

He believes data-intelligence product evolution has three elements: enterprise-grade context sharing, standardization and governance of semantic data, and collaboration between Agents.

For the third he gave a cross-border e-commerce example. A client used a data-analysis Agent and gave feedback that "analysis alone isn't enough" — it found a financial anomaly in an overseas-selling product, then what? Where's the defect? How to fix SEO? That requires the data-analysis Agent to hand structured conclusions to a product-optimization Agent and a customer-service Agent to execute actions. Agent pushing Agent is what truly delivers business efficiency.

As for choosing between "more trustworthy" and "more intelligent," Cen's answer is a line ready to print on a PPT: trust is the floor, intelligence is the ceiling.

In financial scenarios, a relationship manager gives a high-net-worth user an asset allocation recommendation; one wrong number, and trust is gone forever — trust is absolutely first. But in retail, if a competitor runs a promotion and you don't, you've missed the window; then the trend being right and data slightly off is acceptable.

He also shared a counterintuitive case. A well-known tea-drink chain client found the Agent's analytical thinking was more thorough than a human's. A human-written prompt might only score 60, but the thinking the Agent extended from it reached 90. The Agent isn't just executing instructions — it's teaching decision-makers new analytical methods. That leverage is worth more than giving an accurate number.

Start Working First, Then Go to the Moon

Near the end of the roundtable, Sun Linjun told an old Alibaba story that I think was the best closer of the session.

He was then responsible for using big-data systems to handle rights-protection disputes, shrinking a 2,300-person team to 100. Machines' early decision quality wasn't high, so he let it silently learn human judgment in the background, round after round of A/B tests; data accumulated, and in many areas the machine reached or surpassed human level.

"'Last year's Agent might not have replaced humans yet; this year many jobs are already up to it.' He said this very calmly. So back to that CIO with only four GPUs. His problem may not be that there aren't enough GPUs, but that he got the order backwards — enterprise AI never goes to the moon first and then lands. You get every concrete role running first, and only then talk about going to the moon."

More Conversation Details

Speakers

Kyligence co-founder & CEO — Han Qing (韩卿)

Shulie Tianxia / DataHunter founder — Cheng Kaizheng (程凯征)

Shushi Technology AI product lead — Cen Runzhe (岑润哲)

Shizai Intelligence founder & CEO — Sun Linjun (孙林君)

Host

Unique Capital partner — Zhao Liang Abner (赵亮)

Technology Is Democratized, but Buyers Aren't Ready

Zhao Liang: Our flow today is to first have everyone give a brief self-introduction — your company's products, services, and client types — about one to two minutes each. Let's start with Mr. Han.

Han Qing: Very glad to share here. I'm Han Qing, founder and CEO of Kyligence, you can call me Luke. We've focused on the data foundation; over the past decade-plus we've mainly served clients in finance, manufacturing, retail and other industries, and many of China's large banks and insurers are our clients.

Over the past decade, most of what we did was help them manage the underlying data capabilities. Much of the data you interact with daily actually lives on our clients' platforms — of course, we don't own the data. In the past two or three years we've invested heavily in bringing large models into the whole data platform. Our positioning is simple: provide an enterprise-grade trusted data and context platform for AI.

AI is incredibly powerful — the leaps in intelligence the last two years are truly striking — but AI "runs its mouth"; it can't understand the real state of business inside an enterprise. How to give AI this layer of capability is where we're investing now. It's genuinely complex; we've polished it with clients for a long time, from underlying data cleanup and metric tidying to building upper-layer context capability. This is really an enterprise-organizational-level reconstruction, not just a simple technical problem. We keep exploring and hope to partner with more. Thank you.

Zhao Liang: Mr. Han mentioned providing a trusted enterprise AI Agent platform; we'll come back to that. Next, Mr. Cheng.

Cheng Kaizheng: Hello everyone, I'm Cheng Kaizheng from Shulie Tianxia (DataHunter), the founder. The name tells you we're in data; since founding, we've helped clients with data governance, cleansing, and organization.

Starting late 2022 / 2023, we also invested in AI accumulation, including AI Agents and the knowledge system AI requires — i.e., the semantic layer or ontology. Seeing today's guests, we actually have a lot in common on infrastructure and technical accumulation; hope for some collision of ideas.

Zhao Liang: Thank you, Mr. Cheng. Your company was founded in 2014, began operations in 2016 — over ten years. Mr. Han's company also started in 2016.

Cen Runzhe: Hello everyone, I'm Cen Runzhe, currently co-founder and product lead at Shushi Technology. Our company has also been helping enterprises build trusted data semantic layers, very similar to what several guests do.

On top of the data foundation, we provide enterprises with intelligent analytics and the now-popular "digital employee" concept, helping them with digital transformation. Our clients mainly include banks, brokerages, insurance, and manufacturing and retail. This year we find that besides using Chatbots, many enterprises want to hire digital employees (Agents) into their internal IM tools.

Besides data-analysis Agents, we're also building digital-employee Agents this year — moving beyond the web into IM tools, exchanging more context with the enterprise, truly helping reduce cost and increase efficiency. Very much looking forward to exchanging with guests on digital Agents.

Sun Linjun: Hello everyone, I'm Sun Linjun from Shizai Intelligence. Our company was founded in 2018, nearly eight years now. We started with digital employees, entering through RPA (Robotic Process Automation). RPA is one kind of digital employee — you can think of it as "blue-collar," handling repetitive, rule-based tasks like data movement and software operation.

By 2023, with large-model backing, digital employees upgraded from "blue-collar" to "white-collar," entering the Agent stage. The business model didn't change — still selling robots — but the tasks robots can complete are more valuable and more complex. We enhanced the robots' "hands and feet" (operational stability and performance) and developed a vertical large model called "Tasi" as the "brain." Our Agents were among the earlier releases in China in 2023.

Shizai now has over 5,000 clients and has accumulated a large number of scenarios. We've counted that the digital employees deployed at clients long ago surpassed one million. Our original goal was to contribute a million digital employees to society, and we've now exceeded it. Our new mission is "use digital labor to unleash human creativity," and our vision is "human-machine symbiosis, reshaping the work and life of a billion people." As work assistants for humans, Agents' capability boundaries keep expanding, and we're working on that. Thank you.

Zhao Liang: Mr. Sun has already created a million digital employees. So how many digital employees does Shizai itself employ?

Sun Linjun: For example, when clients used to visit our company, a tour took about 15 minutes; now it takes over an hour. Because every department you walk past, employees introduce the Agent they're using. Now when we talk with clients, we first study their org structure and roles, see which can be replaced by digital employees. Any given enterprise, we can offer hundreds of digital-employee templates for quick reuse.

Zhao Liang: You may have noticed that today's four guests run companies that aren't typical startups but relatively mature enterprises, well-known in the enterprise-services track, serving hundreds to thousands of clients each. What's commendable is all four have lived through the evolution from the big-data era to the AI Agent era, with products continuously iterating.

Next, please share your journeys and client-service experience. First question for Mr. Han: as an iconic figure in big data, in the process from system onboarding to result verification, what risks, challenges, or bottlenecks do enterprises hit?

Han Qing: I think there's plenty of "grass" to spit out. In big data, we caught the wave. Then ToB client demand was clear — they knew what they wanted, like buying a car, going to the market to pick price and performance.

But in the AI era things changed. Clients don't know what they want, and you don't know what you can give. It looks like you can do anything, and clients want everything. GeekPark's Zhang Peng once said: customers hope you'll fly them to the moon, but only pay same-city courier money. This process is enormously challenging.

The biggest client-side challenge isn't technology — technology has now been "democratized." Once the requirement is clear, we have the ability to build it quickly. The challenge is that client awareness and willingness to invest aren't equivalent. Many clients say "I'll run an experiment first, you go make money elsewhere" — and every client says this, so in the end it's the investors who get burned.

Two days ago I went with sales to a client; the CIO said they wanted data-related AI. I asked two questions: first, your data is about 2TB — must it stay on the intranet, or can you use cloud AI? He said the data is sensitive and must stay internal. Second, how many GPUs do you have? He happily said four (maybe H200 or H400). He wants everyone to use AI well, but there are only four GPUs — an irreconcilable contradiction.

Leadership says do AI, even lists a hundred scenarios, but in the end can't get server resources. Business-side leaders also complain the AI IT built is useless, because business people can't actually use it, or accuracy is insufficient. I think the biggest challenge is that the old centralized project-building approach must change in the AI era. There's no standard answer now; you can only keep trying. If early foundational work is poorly done and results are poor, clients won't keep investing — a waste for both buyer and seller.

Zhao Liang: Mr. Han analyzed challenges from the demand side and deployment. Same question for Mr. Cheng: Shulie has over a decade of experience, background in traditional ERP and BI. In the AI Agent era, from pilot to scaled deployment, where do you feel the bottleneck is? Do we need new validation standards?

Cheng Kaizheng: A very good question. We've always done classic ToB, so we understand enterprise-side demand well. In the AI era our most obvious feeling is: the boss is most anxious, while the business departments below are less so.

Boss anxiety usually doesn't solve actual problems. As Mr. Han said, IT departments, to please the boss, list several pages of Agent scenarios. But in actual deployment a typical symptom appears: the POC looks very beautiful, but at launch no one dares sign off, because everyone asks: "Is this Agent reliable?"

Our solution: first build an objective evaluation system. We look at how the original person in this role worked, what score they'd get. Humans aren't 100% reliable either — moods, family, all kinds of interference. We use the same evaluation system to measure the Agent; if it doesn't lose to a human, it has the logic to launch. Like Tesla self-driving — at first people doubted, but data proved AI drives more reliably than humans.

Zhao Liang: Won't this evaluation system be too customized? Every client and role is different.

Cheng Kaizheng: That's inevitable. Everyone in ToB knows many scenarios are hard to replicate quickly. Everyone wants to build generic things, but generic products are becoming less valuable because the technical bar has dropped. What's truly valuable is your understanding of the scenario, and the semantics and ontology accumulated from it — that's the core of Agent deployment.

Zhao Liang: Same question to Mr. Sun: you have a strong algorithm background, led algorithms at Alibaba for years. Shizai transformed from RPA to AI Agents and has served thousands of clients. In moving from POC to planned deployment, what "culture-shock" problems did you hit?

Sun Linjun: When doing a POC, you usually only focus on the main flow; you consider some exceptions, but at launch you find many details weren't considered, and these details consume enormous labor.

A few days ago our finance department used a new product and built an Agent in a few sentences, whereas previously an engineer spent a month and didn't fully satisfy them. We found the engineer copied the logic by rote but had no "finance thinking," didn't understand the terminology, leading to constant rework. Dev thought business was changing requirements; business thought dev didn't understand. Today the industry knowledge held by large models even exceeds the average engineer.

Besides, performance is a bottleneck. POCs have low concurrency; after actual launch, compute and data concurrency become sticking points.

On accuracy, I validated at Alibaba: humans make mistakes too, and machines can't be entirely hallucination-free. The launch standard should be that machines reach human-comparable level. We must break the process apart, see where hallucination is allowed (even better with diversity) and where precision is required, and force high standards through multi-dimensional validation.

Zhao Liang: Next, Mr. Cen: Shushi has strong internet and retail DNA. As product lead, what's your vision of the ultimate form of an AI data product?

Cen Runzhe: I think it goes through several stages. In early 2024, everyone felt a smart-Q&A Chatbot usable on phone or web was enough. But with tools like OpenClaw, the form changed a lot.

On the web, users feel AI is just a tool. But when an Agent moves into IM tools, penetrating the organization's capillaries, it goes from "tool" to "partner" or "colleague." You can pull the Agent into a group chat to collaborate; this collaboration is key to Agent evolution.

I think data-intelligence product evolution has three elements: first, enterprise-grade collaboration, achieving Context sharing. Second, semantic data standardization and governance — the basis for trustworthy conclusions. Third, Agent-to-Agent collaboration (A pushing A).

Take cross-border e-commerce: the data-analysis Agent we deployed can analyze financial data of overseas products, but clients feedback that "analysis alone isn't enough" — they want to know where the defect is and how to optimize SEO. That requires the data-analysis Agent to collaborate with a product-optimization or customer-service Agent. Data analysis finds anomalies, passes them structurally to other Agents to execute Actions (like optimizing ads, generating videos). Inter-Agent collaboration is what truly brings business efficiency, and it's our core focus this year.

Zhao Liang: Mr. Cen mentioned two points: from tool to partner, and multi-Agent collaboration. Several guests just raised trust, especially in enterprise settings where large-model hallucination, data drift, and context loss are big concerns. For Mr. Han and Mr. Sun, coming from technology and algorithms, how do you view the trustworthiness of AI decisions? Are there good solutions?

Han Qing: This is the biggest problem in deployment. Clients see a beautiful AI report and the first reaction is: "Are the data points right?" The second: "Can I verify it?" AI self-verification is unrealistic.

Because we build the underlying platform, we start from lineage tracking and definitions to build a verifiable system. Large-model hallucination is inevitable, so our principle is: don't let the large model directly handle core computation.

We hand computation to a stable data platform; the large model only does what it's best at — reorganizing dry data into the secretary-style report the boss likes. AI handles expression; the platform handles accuracy.

Also, manually validating every data point is unrealistic. We're exploring using AI to rapidly validate AI — through scenario polish and human spot checks, continuously raising thresholds to ensure final output is accurate and trustworthy.

Zhao Liang: Mr. Sun's view?

Sun Linjun: I previously used OpenClaw too, gave it a task to scrape data and generate a report. Then I found the data was off; tracing back, the screenshot resolution was too low for accurate recognition, so it made up some data and slipped it in. This shows large models also hallucinate when driving tools, and lack an "active hallucination-suppression" mechanism.

That's why we train large models specifically for process decomposition and execution validation. The longer the task chain, or the more a task is over-compressed, the higher the hallucination rate. A workable approach is to break it into detailed workflows, boxing the large model into a specific range. For financial invoice review, don't throw raw data directly — have the large model write rules or pseudocode and then validate; accuracy jumps.

As versions iterate, large-model hallucinations are dropping sharply. Last year's Agent might not have replaced humans, but this year many jobs are already up to it, and in the future it'll even exceed human level. I'm optimistic about this.

At Alibaba, I once shrank a 2,300-person rights-protection team to 100. We used big data to monitor disputes; the machine saw far more data than humans. Early decision quality wasn't high, but we let the machine learn human decisions in the background and ran A/B tests. As data accumulated and Scaled Up, the machine reached or surpassed human level in many areas. At today's model scale, this holds even more.

Zhao Liang: Next question for Mr. Cheng: Shulie has served traditional enterprises like Nestlé, Zeiss, Johnson & Johnson Medical. Before these enterprises plug into AI systems, what courses do they need to make up? Please share cases.

Cheng Kaizheng: Enterprises adopting AI have two must-have preconditions.

First is data completeness. This returns to the old informatization and digitalization topic — are the enterprise's IT systems (like ERP, MES) complete, and is the data preparation work in place.

Second, often overlooked — semantics and ontology, i.e., how AI understands your enterprise. A large model has general knowledge trained on internet corpora, but it doesn't know internal abbreviations, shorthand, and specific processes. For example, retail's Sell-in, Sell-out, or industry abbreviations like MD, TD. If you throw AI into the company as a new employee without this background knowledge, it can't understand the business.

When data and semantics are unclear and you force AI to complete a task, hallucination arises. Like an exam where the teacher says "if you don't know, guess" — that's AI's logic. If you turn the exam into "open-book," prepare deterministic data and knowledge, and only have AI do reasoning, the hallucination rate drops sharply. From our deployment results, after engineering treatment, hallucinations can basically disappear.

Zhao Liang: Last question for Mr. Cen: now people care both about large models being "more trustworthy" and "more intelligent." In enterprise scenarios, can both be had? If only one, which priority is higher?

Cen Runzhe: It depends on the business scenario. I believe "trust is the floor, intelligence is the ceiling."

For example, a relationship manager at a financial institution giving a high-net-worth user an asset allocation proposal — trustworthiness is absolutely first. If a number is wrong, the customer loses trust forever. But in retail, store-management scenarios, iteration timing matters more. If a competitor runs a promotion and you don't, you miss the window. Here it matters more to let AI use divergent thinking; data slightly off but the trend correct is acceptable.

In bank risk control, financial analysis, trust is the absolute floor. But for an Agent to actually contribute, the ceiling lies in "being more intelligent." Many enterprises find Agents don't just execute instructions — they teach new analytical approaches.

We served a well-known tea-drink chain; they found the Agent's analytical thinking was better than a human's. The original prompt might score 60, but the thinking the Agent extended reached 90. This leverage makes decision-makers smarter — more valuable than simply giving a number.

Zhao Liang: For time reasons, today's Panel has to end here.

---

Original publication: https://uniqueresearch.substack.com/p/src-20260722-01html
On-site reading page: https://ffcap.cn/en/research/src-20260722-01html