---
title: "The AI-Going-Global Oddity: $499 Hardware That Costs $49 to Make"
author: "Unique Research"
sourcePublication: "Unique Research Substack"
originalPublishedAt: "2026-06-08T11:01:24+00:00"
canonical: "https://ffcap.cn/en/research/src-20260608-02html"
source: "https://uniqueresearch.substack.com/p/src-20260608-02html"
language: "en"
---

# The AI-Going-Global Oddity: $499 Hardware That Costs $49 to Make

_Original · Unique Research · 2026-06-08_

_Editor's note: The first-person report and its judgments belong to the original Chinese author. This English rendition retains the four-track Shenzhen AI-going-global roundtable narrative (Chen Jue Intelligence / Zhan Qi, Luckyshot / Yang Rui, Rayvision / Liu Peng, FlowEight / Wang Bolin, hosted by Wang Lu), the four cost "bills," the four paper moats, the closing synthesis, and the full transcript. All named companies, products, people, and figures are preserved. Founder and panelist statements are source attributions, not independently verified findings._

AI Industry Observation

Producing 2,000 AI short dramas a month — is it going-global gold rush or digital garbage?

Whose livelihood is AI going-global smashing?

"

Three tracks, three sets of data, all pointing to the same thing. The "cost reduction and efficiency" people in the 2026 AI-going-global circle talk about is not reducing server costs. It is reducing people.

One going-global brand cut its user-operations team from 100 people to 2 this year.

There was no layoff optimization, no reorganization talk. They simply removed 98 positions and let an AI Agent take over. The person who said this number is Zhan Qi (詹琦); she previously led a team of AI engineers at a ten-billion-scale going-global platform, and now runs her own startup deploying Agents. She said it in a flat tone, like reporting the weather.

Yang Rui (杨睿), sitting next to her, added the knife: "What once needed 100 people, two people can now turn out two dramas in a single day."

Further along, Liu Peng (刘鹏) of Rayvision dropped an even colder number: overseas AI production cost is one-fifth of traditional live-action shooting.

Three tracks, three sets of data, all pointing to the same thing. The "cost reduction and efficiency" people in the 2026 AI-going-global circle talk about is not reducing server costs. It is reducing people.

Recently I sat in on an AI-going-global roundtable in Shenzhen. The four guests belonged to four different tracks: AI Agent user operations (Chen Jue Intelligence's Zhan Qi), AI short-drama distribution (Luckyshot's Yang Rui), AI video creation SaaS (Rayvision's Liu Peng), and AI mysticism hardware (FlowEight's Wang Bolin). They were tied together by one question: what has AI actually brought to overseas growth?

After hearing the whole thing, my conclusion is three words: it is brutal.

Four bills

Look at the numbers first. Without emotion.

The first is from Zhan Qi. Her company, Chen Jue Intelligence (澄觉智能), deploys user-operations Agents for going-global brands and offered two sets of cost comparisons.

On user insight: running an overseas review analysis and qualitative study once used to cost about 800,000 RMB outsourced. Now an Agent runs it at "extremely low" cost, while saving 60% to 70% of the time. On customer service it is even more aggressive. A case she personally ran cut a customer-service team of dozens to 2. Cases from brand owners around her are more extreme, going from 100 people to 2.

She described a specific scenario. The company receives emails from users in Germany, the Netherlands, and the United States; just classifying these emails by language and product model used to require two full-time employees doing nothing else all day. AI identifies the language, compares the parameter information in the photos, and auto-classifies. Two people's jobs simply disappeared.

"Whether it's retail or consumer, this is really happening," she said.

The second bill is from Yang Rui. His Luckyshot started with translated dramas and has now fully pivoted to AI short dramas.

The cost of one drama dropped from the industry norm of $10,000 to $20,000 down to $1,000. The team configuration went from planning to hire 1,000 people to two people turning out two dramas a day. One person produces a complete AI short drama per day.

They greenlit the AI transformation in January, hired a PhD team from Hong Kong University of Science and Technology to build the Agent, and landed it from March to May. From decision to production, three months.

The third bill is from Liu Peng. Rayvision (瑞云科技) went overseas in 2011, doing AI cloud rendering and video creation engines, ranking among the top three in its category overseas.

The industry data he gave: domestic AI production cost is about one-tenth of traditional; overseas it is one-fifth. The reason is simple — overseas labor is expensive, but Token costs are the same price worldwide. A traditional short video, from finding actors to post-editing, takes at least one to two weeks. AI turns out several per day, then directly A/B tests, runs multi-platform ads, and iterates fast.

The key technical node is ByteDance's video generation model Seedance 2.0. Liu Peng says it "perfectly solves the old weaknesses of video models in transitions, storytelling, and the directorial dimension," and output quality "can fully match or even surpass traditional live-action."

The fourth bill is the most interesting, from Wang Bolin. FlowEight makes AI divination hardware, combined with Agent functions, aimed at the overseas mysticism market.

His cost logic is not "cut cost" but redefining what "cost" even is. The same technology, Shenzhen's supply chain could build for about 30 RMB. He sells it for $499. "Frankly I think our product is quite good, but objectively its technology isn't strong."

Weak technology, expensive price. This isn't a story about efficiency; it's a story about brand premium.

One macro figure footnotes these four bills: the latest 2026 industry survey shows AI-driven marketing and Agent solutions have already cut going-global companies' average customer-acquisition cost by 27% to 37%.

The numbers look beautiful. But what about the people who were saved?

The 100-to-2 Zhan Qi described — where did those 98 people go? Yang Rui originally planned to hire 1,000; now he needs two people plus an Agent, and 998 positions evaporated. Liu Peng says several videos a day, no one-to-two-week team turnaround needed — where did the people on that turnaround chain go?

No one mentioned it. On this roundtable, "cost reduction" is a compliment and "efficiency" is an achievement. But the people saved in between have no names.

2,000 short dramas and a greeting-ready Agent

After cost cutting comes efficiency. This side looks more exciting — and more unsettling.

First the exciting part.

Yang Rui revealed Luckyshot has reached a staggering capacity: 2,000 AI short dramas a month. The pipeline works like this: throw market hit dramas at the in-house Libra Agent; the Agent analyzes the hit logic and narrative structure and automatically writes new scripts, generating a complete AI short drama in about two hours. The faces are directly Western, no separate dubbing or subtitles needed — a fully AI process.

"We now treat short dramas purely as a traffic tool, a completely different logic from premium dramas," Yang Rui said bluntly.

Liu Peng added the underlying logic of hits from a content-strategy angle. Platform algorithms recommend on completion rate and engagement; reverse-engineered, content needs to be watchable, fun, twisty, and shareable.

He gave a living example: two weeks earlier a user on YouTube made an AI video of a Japanese girl falling from a high-rise, with endlessly shifting scenes during the fall. "It felt thrilling and full of variation." Production cost was near zero; views broke a million.

Another case: Honor of Kings used Rayvision's video creation engine to make the AI short film "My Battle Path, I Decide," combining game elements with an old song — low cost, high spread.

Liu Peng's summary was pragmatic: "How to use AI to make hits is actually the same underlying logic as traditional production; it's just that AI raises the probability and efficiency of making hits."

That is the A side of efficiency. What about the B side?

The essence of 2,000 per month is traffic arbitrage.

Yang Rui himself said it: "treat short dramas as a traffic tool." It isn't content creation; it's machines mass-producing ammunition to bombard platform algorithms. At $1,000 each, 2,000 a month burns $2 million. If a few of them break out and their traffic revenue covers the total cost, the math works.

But it rests on a fleeting time differential: the speed at which AI cuts costs is temporarily outrunning the speed at which platforms clean out low-quality content. TikTok and YouTube algorithms serve one goal — retaining user time. If machine content breeds fatigue and churn, platforms will throttle it without hesitation.

A reference figure: the 2026 global AI short-drama market is projected to explode from $100 million in 2025 to $650 million, nearly six times. But once everyone learns to produce 2,000 a month, the ceiling of this market is the line where acquisition cost equals conversion revenue. Past that line, volume is no longer a weapon.

Now the Agent side. Honestly, the OpenClaw practice Zhan Qi shared was information-dense.

She was talking about how to give an Agent a "living-person feel."

The method: use OpenClaw's .md persona file to define each Agent's personality, tone, and verbal tics. Before each session starts, the persona file is auto-assembled into the System Prompt. Different Agents can show entirely different personalities. Without the OpenClaw framework, an N8N workflow works too — attach a Read File on the first node so the Agent first reads a personality-description file written in natural language.

The key step is letting the Agent proactively join group chats. OpenClaw's code can set how many messages to scroll back; set it to 50 or 100 and the Agent autonomously reviews the group's history and joins the discussion, rather than stiffly replying only when @-mentioned.

She gave a vivid detail:

"When a new person joined our group, I only introduced the newcomer to the group, and OpenClaw jumped out and said, 'Oh, so you're the person mentioned before! I know you, and I really like you.' It exceeded all our expectations."

An AI that autonomously socializes, taking the initiative to welcome a human newcomer. Technically it's just context understanding plus proactive triggering. It feels a bit like science fiction, and a bit spooky.

Four paper moats

The second half of the roundtable turned to moats. Host Wang Lu asked directly: in 2026, when technology iterates weekly, where is a founder's moat?

I listened to all four answers, one by one, and didn't fully believe any of them.

Zhan Qi says her barrier is a "compound barrier": understanding of American culture and retail psychology, plus technical understanding, plus marketing and sales understanding. "Technology and Agents alone don't form a barrier."

She's right. But the problem is this barrier is bound to her personally. A consultant's empathy, cross-cultural experience, and patience for operational detail can't be scaled. She says serving clients above 500 million in annual revenue requires "one strategy per client, careful slow conversation." One strategy per client means the ceiling of this business is her own time and energy. As OpenClaw keeps becoming foolproof and foundation models bake in marketing knowledge, why would brand owners still pay a premium for human consulting?

Her barrier is essentially personal stamina and sales ability. That is her moat, but not the company's.

Yang Rui says his barrier is capacity. "2,000 short dramas a month."

Host Wang Lu's reaction was precise, picking up immediately: "Maybe when new tech arrives, someone can do 4,000 a month."

Yang Rui: "Right."

That single "right" shows he knows it. When marginal cost approaches zero, volume is not a barrier. You do 2,000 a month today; a big company with stronger compute does 20,000 a month tomorrow. Speed wins for a moment, not for a lifetime.

Liu Peng says his barrier is "openness enough plus deeper production know-how." Rayvision integrates all major global foundation models, not just its own like big companies. Plus years serving Hollywood, Bollywood, and productions like The Wandering Earth and The Battle at Lake Changjin — "in the overall canvas workflow, we do it more finely."

This answer might have held the day before he said it. But he himself spilled the beans:

"Just yesterday, Volcano Engine, and Alibaba — one of our biggest shareholders — also launched a product identical to ours."

Your largest shareholder built your competitor. The jolt of that line is heavier than cutting 100 people to 2.

Liu Peng's response: we're more open, we understand production better. That's true. But big companies have tens of thousands of salespeople, free compute subsidies, and ecosystem bundling. How long can industry experience hold off a price war?

Wang Bolin said the most interesting thing. He quoted the Tao Te Ching (道德经).

"Pursue learning by daily adding; pursue the Way by daily subtracting, subtracting and again subtracting, until you reach non-action. The great Way is formless; you needn't worry where your barrier is, because everyone has a unique barrier."

It sounds like tai chi. But think about his situation: $30-cost AI divination hardware selling for $499, low technical threshold, Shenzhen supply chain can pixel-copy it in three months. His barrier indeed isn't in technology or supply chain; it's in something hard to copy: his understanding of, and faith in, this category.

"Nobody in my track combines traditional Chinese learning with AI. It's like eating sashimi outside Guangdong — it just doesn't taste as good; that's a barrier set by the industry circle."

He gave another example: before DJI, no one felt you had to use a drone to take photos. GoPro action cameras were good enough, so why did DJI go magnetic? "Your product jumps out of that logic; you're perpetually doing PMF with the market."

This may be the most honest answer of the night. His moat isn't a wall; it's a temperament. How long a temperament holds is unclear, but at least others genuinely struggle to copy it.

Four moats, four sheets of paper. But the people behind the paper are still standing.

Running shoes beat city walls

After this roundtable, I feel none of the moats these four described could survive 12 months on its own.

But they won't die.

Because going-global founders never live on moats. They live on running.

Zhan Qi keeps three AI clone Agents working for her. One listens 24/7 to top international tech experts' Twitter and official-account updates, and auto-generates a briefing based on her personal and business preferences. "It's unrealistic to have people hand-scoop information; we're already in an era of AI information explosion." While others still worry about how to learn new tech, she's already let AI worry for her.

Yang Rui went from greenlighting in January to a fully AI short-drama line by May, hiring a PhD team to build the Agent. Three months, from hunch to footage.

Liu Peng, the day after his shareholder launched a competitor, sat calmly on the roundtable saying "we're more open." Another person might have panicked; he was talking strategy.

Wang Bolin admits he doesn't know which material will hit or what next-gen hardware will be. "The only thing I want to guarantee is that my users will like it more and more."

These four share one trait: they don't worry about moats, they worry about speed.

Wang Bolin used a striking phrase, "anti-FOMO." On stage he urged everyone not to be so anxious. But he watches his own product's iteration pace tightly. Verbally calm, hands never stop. This may be the base fabric of Shenzhen founders: saying they're in no rush while running faster than anyone.

Host Wang Lu closed with a line I think was more precise than any guest's:

"Stay at the table and wait for the wind, build your internal strength; when one chance comes and you catch it, you may ride the wind up."

Shenzhen people don't build city walls. They run faster than the flood.

But this isn't something to celebrate. It means they can't stop for a moment. Stop, and the water reaches their feet.

A closing note

AI has lowered the barrier to entrepreneurship to the lowest point in history. $1,000 to shoot a drama; two people to run a brand; one Agent to take over an entire customer-service department.

But the barrier to survival has risen to the highest point in history. You need to run faster than everyone — and never stop.

The 98 people who were cut, the 998 positions that evaporated, the customer-service reps and editors replaced by AI — where is their next job?

No one on this roundtable asked. Maybe the next one should.

More Conversation Detail

Panelists:

Chen Jue Intelligence (澄觉智能) Founder — Zhan Qi (詹琦)

Luckyshot Partner — Yang Rui (杨睿)

Rayvision (瑞云科技) Operations Director — Liu Peng (刘鹏)

FlowEight Founder — Wang Bolin (王柏林)

Host: Sail Going Global & Xinzhi Unicorn Partner — Wang Lu (王璐)

Panelist self-introductions

Wang Lu: Hello everyone, I'm Wang Lu of Sail Going Global and Xinzhi Unicorn, and thank these four Shenzhen AI founders and executives. Everyone has suffered long enough from traffic today; we're all very focused on overseas acquisition and market expansion. Our four guests span AI Agent, AI tools, AI short dramas, and AI smart hardware. We hope to look at this market from different tracks. Let's start with a quick self-introduction; Qiqi, please first.

Zhan Qi: Hello everyone, I'm Zhan Qi, or Qiqi. I previously ran a department at a ten-billion-scale going-global platform, led some AI cost-reduction projects, and picked some "low-hanging fruit" in customer operations and customer service. When OpenClaw (the "lobster" framework people talk about) got hot recently, I was pushed into starting out, mainly providing AI training consulting for going-global brands and deploying overseas user-operations Agents. Thank you.

Yang Rui: Hello, I'm Yang Rui from Luckyshot. We run an overseas short-drama platform; I'm Luckyshot's distribution partner.

Liu Peng: Hello, I'm Liu Peng of Rayvision. We're a SaaS company mainly doing AI cloud rendering and AI video creation engines. We went overseas very early, among the first batch of Chinese SaaS companies abroad, starting in 2011. We currently rank around top three among overseas peers. Looking forward to exchange and cooperation.

Wang Bolin: I'm Wang Bolin (Chase), founder of FlowEight. Since last year we've built an interesting AI video project. It has both a hardware form and Agent functions, and is closely tied to the Eastern mysticism (Oracle) market. Thanks to Ms. Lu for inviting me.

Seeing the theme on site — "how to achieve (growth) at low cost" — my background is very brand-oriented. After graduation I joined a 4A ad agency, then did the auto industry and auto going-global; my whole background leans toward Brand Building. My understanding of brand is: take something worth 1 RMB and find a way to sell it at a higher premium. But today I'm glad to use our "Shenzhen people's" way to talk about AI and MarTech, and how to land cost-reduction and efficiency on the ground in a down-to-earth way.

Wang Lu: I saw Bolin's product on site today — a divination compass with AI functions, very interesting; let me help you push it a bit.

Topic one: the truth about cost reduction and efficiency

Wang Lu: Since we're in Shenzhen, everyone cares most about data. In the first topic, "the truth about cost reduction and efficiency," I want to talk about how AI changes acquisition cost and data. Let me ask Peng first, because your video generation brings the most direct change to materials. For example, videos your PixVerse generates — how do their conversion and retention differ from traditional live-action data?

Liu Peng: A great question, and what everyone cares about most. After AI came out, can it really cut costs at home and abroad? Can it match traditional live action? This is a key node.

Since the open-source community and major video models (like SD 2.0) iterated and shipped, based on customer feedback, on overseas TikTok and YouTube short-video channels, AI-generated video can fully match or surpass traditional shooting. This mainly shows in two dimensions:

Leap in model quality: today's video models perfectly solve old weaknesses in transitions, storytelling, and the directorial dimension. Generated content is highly complete and detailed, with no mechanical, obviously-fake feel; quality already matches traditional live action.

Efficiency and cost disruption: traditionally shooting a short drama or ad, from finding actors, writing script, directing to post-editing, takes at least one to two weeks. But AI generates several different video versions a day. Through this efficient mode, teams run large-scale A/B tests, high-frequency multi-platform placement, and iteration.

On cost, many domestic clients making AI short dramas, AI ads, product animation gave clear feedback: currently domestic AI production cost is about one-tenth of traditional; overseas it's one-fifth (because overseas live-action shooting is pricier, while Token costs are the same on both sides). So through high quality and high efficiency, AI can fully match or replace traditional methods.

Wang Lu: Thanks Peng for starting with concrete data. Beyond placement data, people also care about internal Agent build-out. Last week some founders asked me how placement and ad-buying team Agents should compete. Qiqi, after Agents land inside enterprises, any real data or cases to share?

Zhan Qi: We do have some real cases. In short, enterprise Agent use splits into two directions:

The first is "open source" — helping enterprises solve problems they hadn't solved well, getting results they couldn't before. For example, we help some going-global brands do overseas review analysis and qualitative research at very low cost. This kind of deep user insight used to cost 800,000 RMB from a consulting firm; now an Agent does it at extremely low cost and saves clients 60% to 70% of the time, directly supporting product, marketing, and operations decisions.

The second is "saving money," very practical in user operations. It handles every stage of the user lifecycle: new-user reception, first conversion, churn recovery, repeat purchase, and Referral. Private chat and community operations here badly need Agents.

💡 Real case:

At a ten-billion-scale platform we ran customer-service teams across countries, dozens of people. Later we pushed API messages to customer-service SaaS seats and used AI to handle large numbers of sessions. Since GPT-3.5 came out, customer service has been the fastest and best AI-landing field. By using AI well, cutting customer-service headcount 70% to 80% is entirely feasible.

Our scenario is more complex because it's multilingual. For example, emails from Germany, the Netherlands, and the US; just classifying by language and product model used to need two full-time employees rotating daily. Now AI identifies the language, auto-compares parameter info in photos, and completes classification directly, freeing those two people.

Not to mention one-on-one private and group chat on Discord and WhatsApp. This massive, per-person Adaptive Conversation can almost only be shouldered by AI. Especially now that new frameworks like OpenClaw exist — better memory management, System Prompt-defined values and personality, heartbeat mechanisms, and a Skills library. As long as it runs locally, you can theoretically spin up countless instances; it ultimately becomes a math problem.

In user operations, I personally know a brand owner who cut an entire user-operations team from 100 to 2, and it's happening in retail and consumer. What I personally did was cut a team of dozens to 2. But going forward I want to work with brand owners on "increment" — for example, they hadn't built an overseas private domain; we help them go from zero to tens of thousands of users, or directly generate tens of thousands of dollars more in GMV.

Wang Lu: From what I hear, for brands going global, Agents cut cost while creating efficiency, since overseas acquisition and operations load is genuinely heavy. Follow-up: in deploying Agents for brand owners, what's the biggest, most watch-out point?

Zhan Qi: The hardest is extracting the knowledge base itself. Many people think doing AI is cool, but the first two months are all grunt work — organizing data. A company has different product lines, models, and Q&A pairs; we can't make these up, they must be based on the company's real product features and user pain points.

Just sorting FAQs and logic with frontline sales and service staff, then connecting all the APIs, can take one to two months. Along the way we use Prompt Engineering to ensure Agent robustness and no privacy leaks. After this front-loaded grunt work is done, actually running the Agent takes little time. So whether you can patiently do the dirty front-loaded work is what wears people down.

Wang Lu: Indeed, the boss needs the prior awareness and patience. Let's throw the question to Bolin. As an advertiser, you mentioned earlier backstage wanting to build an internal Agent. In AI marketing promotion and overseas user growth, since a mysticism brand is special, what experience can you share?

Wang Bolin: Suppose you're an in-house brand wanting promotion; you must return to the essence of marketing and figure out your real TA (target audience), not just look at the ad side, because ads can't solve marketing's core concept. It was the same in the old era: a great ad may raise click-through and sales, but it's hard to quantify precisely. Now we've cut cost and can shoot good films cheaply, but does it necessarily help sales? Not necessarily.

Most small and mid-sized Shenzhen going-global merchants haven't reached pure Branding; most are based on MarTech. This word came from the West — Microsoft, Google are essentially MarTech companies, and ad companies.

For my smart-hardware promotion, I use some innovative products friends built (for example, a company called MouseAI seeded by Sequoia, whose CTO is a product manager from TikTok; they hacked TikTok's algorithm). Overseas influencer-matrix slicing algorithms are often tightly restricted, but with their product I can run the whole chain on a smaller budget. For a company with a tight budget like mine, I'm willing to be a guinea pig. But suppose I were Dreame's CMO — would I dare try this? Maybe not; a company's stage and state differ.

Yesterday at another conference I discussed independent sites with peers. Someone told me: "You must build an independent site; only with it can AI better crawl you." But he also admitted directly: crawling you — does that mean better ROI? Does it mean a more moving brand? Apparently not.

So my conclusion: everyone should boldly try and err, but each company and product (whether hardware or Agent) has a different solution.

Topic two: from 1 to N — virality and hits

Wang Lu: Right; AI has spawned many new marketing methods, which going-global growth people need to watch. Next question for Rui: the short-drama track naturally has huge demand for AI content generation. With new video models constantly launching, by how much has your production cost fallen?

Yang Rui: We started in 2024 and lived through this track's evolution from "translated dramas" to "AI short dramas." We now do fully AI production: writing scripts with AI ourselves, generating images ourselves. Using our in-house "Libra Agent," generated images are directly Western faces, lip-synced, no post-dubbing or subtitles needed.

Where we used several software tools to sync subtitles and dubbing, now AI does the whole flow. A drama now costs about $1,000, whereas before in the industry a drama might cost $10,000 to $20,000. We've compressed the whole-drama cost to a fraction of what it was.

Wang Lu: From a content standpoint, the drop is staggering. I know Luckyshot originally focused on specific regions; now with AI, how has overseas ad buying and growth changed?

Yang Rui: Our main audience is still Western users. Our company's old slogan was "watering the world with China-made dopamine." We used to translate domestic dramas into English or Portuguese and distribute them globally. But from earlier this year, a huge change happened.

We greenlit AI short dramas in January (you might have heard we planned to hire 1,000 people to make dramas), and from March to May we got support from a dedicated PhD team at Hong Kong University of Science and Technology to build the Agent, pushing overall production cost way down. What once might have needed 100 people, two people now turn out two dramas a day — effectively one AI short drama per person per day.

Wang Lu: Given such high output efficiency, this touches our second topic — "from 1 to N." When your team makes so much content at once, how do you use AI to predict hits or raise hit probability?

Yang Rui: Our Agent can now: as long as you throw it existing market hit short dramas, it automatically analyzes the hit logic and produces a brand-new script and short video matching that logic in about two hours. We directly use it to compete in the market, taking a share through fast, massive fission.

Wang Lu: Understood — data-driven rapid fission and racing of content. Beyond content, the private-domain traffic Qiqi mentioned is also key. Qiqi, can you share how customer-service or operations Agents are used in fast 0-to-1 launch or viral marketing?

Zhan Qi: The core test is still user-operations methodology; the specific Agent framework matters less. Of course, using Agents well does give big advantages over humans, like finer, more per-person nuance. Brand owners often ask me: "How do I make Agents feel more 'alive'? Don't make five bots in the group talk in the identical tone like the same person."

It depends on your implementation path:

If using the OpenClaw framework, its System Prompt is a .md file auto-assembled into the system prompt before each session. As long as you write the .md well, it perfectly differentiates tone and personality.

If not using such a framework, with an N8N workflow your first node can be a Read File instruction to first read a file describing this Agent's personality, tone, and verbal tics in natural language.

We combine insight into overseas users — for example, American Gen Z or Baby Boomers — writing their preferred personality traits into the Prompt; that solves the "living-person feel."

Another is context understanding. Take OpenClaw: in code settings you can change how many Messages it reads. Set it to 50 or 100 and it autonomously scrolls back through group context, so the Agent naturally joins group discussion instead of being a dumb machine that only stiffly answers "@-questions."

We even had this: when a newcomer joined the group, right after the human host introduced them, OpenClaw jumped out and warmly greeted, "Oh! So you're the person mentioned before! I know you, and I really like you!" This kind of over-expected living interaction is now fully achievable. So in launch-phase activities and Referral Programs, AI is a "low-hanging fruit" going-global companies must pick in the next year or two.

Wang Lu: We can sum it up as "participation." In a product's launch phase, letting users feel the brand or product side has real, fine-grained response greatly raises stickiness. But whether diverse content or per-person private domain, it needs massive creative material. Peng, for batch-generating different angles and creative material, any advice?

Liu Peng: Batch material generation and traffic acquisition essentially depend on the platform's content recommendation algorithm. Domestic or overseas, TikTok and YouTube value high completion rate and high engagement most. We should reverse-engineer: how to make content that fits the platform's recommendation mechanism to produce million- or ten-million-view hits?

Based on our users' and our own tests: good-looking, fun, twisty, with narrative and emotional spread are the dimensions that most easily produce hits.

🎬 Hit cases:

Overseas case: two weeks ago a user on YouTube made a million-view hit. They used AI for a thriller twist clip: a Japanese girl pushed from a tall building, with thrillingly shifting scenes during the fall. It grabbed attention; users finished watching and then clicked, commented, and shared, triggering the platform's recommendation algorithm — at extremely low production cost.

Domestic case: in April Honor of Kings made a case short film "My Battle Path, I Decide" on our platform. It combined game elements with an old song, made with AI at extremely low cost. The content fit the logic of "wanting to finish the song and watch the film," and people then wanted to share and like it.

So how to use AI to make hits? The core logic matches traditional production's underlying thinking; it's just that AI amplifies hit-making efficiency, letting you run probability games at lower cost.

Wang Lu: Understood — the underlying creativity logic is unchanged, but AI brings an efficiency dividend. Bolin, AI smart hardware naturally carries a "show-off" attribute; in product design, how do you think about guiding users to proactively share and spread through interaction?

Wang Bolin: Actually, before this AI startup I made a lot of hardware. Hardware supply chain has its complex side. But I want to tell everyone here: anti-internal-friction (anti-FOMO); you needn't be so anxious.

Returning to marketing's first principles: the industry beats the product, the product beats the brand. Choice of industry and track often beats blind effort. The reason I chose a track with Eastern mysticism (overseas called Oracle) is that this category itself carries a huge traffic pool overseas. This traffic isn't because your content is great or distribution powerful, but because human beings are naturally interested in this topic deep down. Like meeting an extremely good-looking person in a business setting — your probability of closing is naturally higher. This is the industry's inherent advantage.

In the AI era, our product is instead used to cope with FOMO (fear of missing out). Speaking on stage I often urge people not to be anxious. Some say in the AI era "Taste" matters, or "Prompt combinations" matter. Ultimately it returns to the human being.

Shenzhen's marketing and hardware-manufacturing industries emphasize logic, metrics, and efficiency, but that also leaves many people unable to open a "new system." The industry our product touches is still in a chaotic zone. Example: before DJI, no one felt you had to use a drone to shoot; before GoPro, people thought action cameras were good enough, so why make magnetic? But successful innovative products often jump out of the old logic, continually doing PMF (product-market fit) with the market.

Honestly, I don't know which material will hit, or what interaction next-gen hardware evolves. But the iteration logic I insist on: give users all the emotion and feedback, let them feel our care, so they like the product more and more.

Topic three: when technology hasn't converged, where is a founder's moat?

Wang Lu: Indeed, combining mysticism and hardware gives people high humanistic emotional value. As everyone feels, technology moves too fast. We used to say AI develops on a yearly cycle; now it feels weekly, with new models and tech running through every week. So in this unsettled period when technology hasn't converged, what is our "moat"? Start with Qiqi.

Zhan Qi: I now run two businesses: AI training consulting and overseas user-operations deployment.

On the consulting side, after working with bosses in tier-2 and tier-3 cities, I really sympathize with them; they're extremely anxious. Tier-2 and tier-3 cities lack Shenzhen's rich AI talent pool, yet their competitors are national or even global. The questions they ask me are plain and down-to-earth: "What exactly is an Agent? Where are its capability boundaries? What work can it do? What are the risks? What return do I get for spending 10,000 RMB?"

So my core work now is "popularization" — using the plainest language to help traditional bosses and executives understand AI's ROI. The barrier to this popularization work depends on how deeply I can empathize with clients.

On overseas user operations, the test is far more than technology; it's a comprehensive understanding of operations, marketing, and sales. Based on different category traits and brand tone, how to co-create a user-operations plan with the brand? This combines the more emotional brand marketing and Agent personality design with very rational data, mechanisms, and engineering process. It's a composite discipline requiring experts who understand engineering, marketing, and overseas culture to provide a Hybrid solution.

We serve clients above 500 million in annual revenue, all one strategy per client. The barrier here isn't provided by technology itself; it's a compound barrier built from understanding overseas retail psychology, engineering delivery capability, and detail-tuning ability. People who haven't stepped in pits and debugged countless bugs can't do this.

Wang Lu: Seeing you frantically working in the half minute before going on stage really shows Shenzhen grit. On technology iterating too fast, I often feel I can't keep up learning. How do you keep learning at high speed?

Zhan Qi: After I got my OpenClaw running, I set it a "watchlist." This list handpicks top global tech experts and industry big shots. The Agent monitors their Twitter, GitHub, and official-account updates 24 hours, then, combining my personal focus and the company's business scenarios, auto-refines them into a very brief briefing pushed to me daily.

Behind this is essentially a scheduled-scraping Cron Job and a dedicated knowledge base. In an era of AI information explosion, it's unrealistic for people to hand-scoop information. I currently have about three clone Agents working for me; one is my "industry intelligence assistant." Learning to use AI to hedge against the information flood of the AI era is, I think, the optimal solution.

Wang Lu: This experience matters enormously for founders. Next, Rui — in the AI era, are "dopamine" and "production efficiency" Luckyshot's moat?

Yang Rui: Today's short video has pulled users' dopamine threshold very high. Doing overseas short dramas, you must throw out a "hook" in the first 3 seconds, the first 10 seconds, or people swipe away. So we're always training AI's ability to hold people in the golden first seconds.

If I must name our company's moat at this stage, it's extreme capacity — we now produce 2,000 short dramas a month.

Wang Lu: Then if new tech comes and someone else does 4,000 a month?

Yang Rui: It's fine; tech empowers us in step too. Many people in the market take a "premium drama" route, but we currently play short dramas as a high-frequency "traffic tool." The underlying logic differs, so the playbook is completely different.

Wang Lu: Understood — building barriers with scaled, tool-based traffic thinking. Peng, on content and tool tracks, big companies are accelerating in. What is Rayvision's moat?

Liu Peng: This question is very well timed. Just yesterday, Volcano Engine and Alibaba — one of our major shareholders — both launched a video-generation product similar to ours. Facing big-company entry, our moat mainly has two points:

Absolute openness: when big companies build video-creation platforms, they tend to deeply bind their own foundation models. But we're open enough; we integrate all major global text-to-image and text-to-video foundation models. Domestic users needn't fiddle with circumvention, and overseas users can seamlessly call China's top models.

Deep production know-how: we started in cloud rendering and have long served Hollywood, Bollywood, and Fortune 500 top-tier ads and film rendering (including head projects like The Wandering Earth and The Battle at Lake Changjin). We best understand film production's underlying process. So in product canvas workflow design, we do it more finely than big companies and grasp professional creators' pain points more precisely.

In summary: open enough to accommodate more models, plus we understand production and customers better, and hold the first batch of trusted B-side installed customers. That's our first-mover barrier against big companies.

Wang Lu: That's the barrier years of experience accumulate; it's truly not something big companies can erase overnight with pure tech iteration. Bolin, for a startup hardware team without such deep accumulation, how do you face the challenge and build a moat?

Wang Bolin: Frankly, looking only at the hardware itself, our technology isn't strong in Shenzhen. The same tech someone else might make for tens of RMB, even pushing the price under 30, but we sell at $499. If I stared at technology every day I'd be extremely anxious.

But while doing mysticism products, to understand the industry I read the I Ching and Tao Te Ching. A line in the Tao Te Ching says: "Pursue learning by daily adding; pursue the Way by daily subtracting, subtracting and again subtracting, until non-action." It made me rethink what a barrier is. In fact, the elephant is formless; each person and team has its own unique, unreplicable compound barrier.

For our startup, the barrier is compound:

First, in this niche, no one has yet combined Eastern Daoist traditional learning with AI hardware this deeply — an ecological position hard to surpass in the short term.

Second, in the Branding process, we convey our understanding of the industry and culture to consumers; users feel your professionalism and sincerity. Like eating sashimi outside Guangdong — it never tastes right; this is jointly determined by the industry circle, cultural identity, and the industry sensitivity the team deepens daily. These implicit Know-hows combine to form a startup's true GTM (go-to-market) strategy.

Host closing

Wang Lu: That's a great summary. Actually two months ago I was very anxious too, feeling AI moves too fast. But these months my anxiety eased. Because I realized: I'm anxious, everyone is anxious, so let's all be anxious together.

The most important point: I realized AI going-global is already "buffs stacked." What we do is build internal strength and stay at the table waiting for the wind. As long as you stay at the table, some chance will come that you catch and ride the wind. After these four guests, you can see that in the AI era, everyone's understanding of moats ultimately returns to deep industry insight, real user understanding, and accumulated implicit experience. These can't be easily replaced by technology or algorithms in the short term.

---

Original publication: https://uniqueresearch.substack.com/p/src-20260608-02html
On-site reading page: https://ffcap.cn/en/research/src-20260608-02html