“The hardest part is not going overseas. The hardest part is knowing exactly who you are the moment you step out.”
Agnes Video 2.5’s list price for generated video is 0.15 RMB per second. Real deals are usually cut in half — about 0.07 RMB, or roughly one US cent, a second. And if you don’t want to spend even that, there is a completely free Flash tier.
That was the first set of numbers I wrote down at a roundtable on AI going global. The four people on stage happened to sit at four different positions on the industry chain: two model vendors, Agnes AI and Shengshu Technology (生数 / Vidu); one platform distributor, China Reading (掌阅科技, iReader/iDrama); and a leading TikTok short-drama production house, Guanghe Power (光盒动力). The topic was “going overseas”: which market to enter first, which platforms to use, and how to split the budget.
Put one line from Shawn Xiao (肖伟), head of overseas business at China Reading, first:
“The hardest part is not going overseas. The hardest part is having a crystal-clear positioning when you do.”
He talks positioning at every meeting, and the reason is practical enough. The short-drama export chain has three parties: scripts, production, and distribution. He has seen too many companies try to grab several lines at once, end up with a low hit rate on all of them, burn through their money, and still not have found a position.
China Reading knows exactly what it is. It will not out-engineer the model companies. It will not out-spend competitors on user acquisition, and it will not fight for production compute jobs. So what does it do? It pulls all of them onto its own boat. Shawn’s words: “Nobody is my competitor.”
This is not a pose; it is built from real money. I noted the numbers on stage: an S-tier script is guaranteed at up to 100,000 RMB; B- and A-tier scripts are bought for 10,000 to 30,000 RMB. Finished dramas are procured at more than 500 per week, and a top exclusive S-tier finished drama can command 80,000 to 100,000 RMB. Production is quoted by the minute, at 400 to 1,500 RMB.
More interesting is the quality-control design. China Reading runs a fifty-plus-person QC team that production houses say is ruthless. But here, QC is actually a price ladder: move from B to B+, and the per-minute rate rises from 400 to 600 RMB; move from B+ to A+, and it rises again to 700.
“You improve, I pay more” — giving partners a stable income line. As Shawn put it, handing out policy is not enough; you have to know how to help these people actually make money. In the words of the host Zhao Liang (非凡资本 partner), China Reading wants to build a boat and pull everyone overseas together. Counting from the iReader launch in 2015, this road has taken more than a decade; the secret to getting this far boils down to that one sentence.
The passenger running furthest ahead on the boat is Guanghe Power. In 2024 it became the number-one TikTok short-drama production service provider. It entered TikTok-native (端原生) content this March and has already listed several hundred native AI short dramas.
But founder Jim Zhan (詹卓安) opens the positioning conversation by demystifying it: going global was a passive choice. By 2024, the domestic short-drama fight had already turned brutal; staying home was not an option. He accepts the popular line: “Go overseas, or go out.”
What is interesting is his path design:
First be the “DreamWorks” or “Foxconn” of AIGC, then slowly grow into the Disney or Marvel of AIGC.
In plain terms: first be a vendor and push content capability to the extreme, then turn around and build owned IP. Genres — sci-fi, werewolf, cute pets — are filtered by content tests and paid-traffic feedback; once it finds a direction where it has a real shot at the top, it builds an IP matrix there.
His way of using TikTok is also worth unpacking. In his hands, TikTok is a test field: one good script gets 4–6 native distributions, watching comments and data while it runs, then going back to revise the script.
One set of content assets is split into character material and highlight clips, repurposed as serialized e-commerce creative, and even used with view data to pitch interactive traffic-driving on Steam. Same drama, same distribution method — he says TikTok’s data is currently better than YouTube’s.
One production, many distributions, many monetization paths. This is the production-house’s sense of positioning: the content asset is yours; the platform is only a channel.
On the model side, Hanson Zhang (张毅泷), co-founder of Agnes, gave the most counterintuitive answer.
Start with the business model. Large model companies sell APIs at list price with discounts. Agnes went one step further: its Flash series is free outright, at the cost of rate limits — one request per minute for video, at most five reference images, resolution locked to 720P. It only launched its API in June this year, but volume grew fast thanks to the free entry point.
Where does the money come from? Freemium. On the consumer (C) side it builds the brand: roughly 15% of videos generated each day come from paying users. On the business (B) side it makes the money: customers are API relay stations, platform companies, and short-drama studios.
Its own app Pavo, with no promotion, adds more than a thousand daily active users. The short film “Cloud Palace” (云上仙宫), which a foreign ministry spokesperson recommended to overseas viewers, was made with Pavo.
The real counterintuitive point is this: the Agnes brand originated in Singapore and started from Southeast Asia, accumulating more than ten million overseas users. But Hanson says that as a model company, its largest market is China, and this year’s growth is mainly domestic. Put accurately, Agnes went out first, then came back home.
A group of content companies are learning how to go out, while a model company that was born overseas is walking back. Why? Because the model business follows compute and customers, while the content business follows users and wallets. Positioning sets the direction; geography is only the result.
Its overseas brand-building is equally counterintuitive: it barely buys traffic, slowly building word of mouth in Discord, Reddit, and developer communities. That is long-term work, testing sustained operations.
The roundtable was genuinely dense on “where to go,” and I assembled everyone’s pieces into one map.
Want revenue? Go to North America. Willingness to pay is strongest, but customer acquisition cost has already risen from $10–20 to $40–50; this is where you make premium IAP.
Want volume? Go to Southeast Asia and Latin America. Southeast Asia is populous but fragmented — Indonesia alone has 170 million people, users are accepting of Asian faces, and some would rather watch a 45-second ad than pay $1 to unlock. It is the place to test scripts and directions.
Latin American users are like Bilibili viewers: they love danmaku and comments, rarely pay on their own, but cluster together — if one person decides a drama is good, a whole group may pay together.
Jim offered a reference point: domestic short-drama users already spend more than 20% of phone time on it; Latin America is about 15%, North America about 5%, and the global average under 2%. The ceiling is still very high.
Japan and Korea are an interesting market: once you get in, it is hard to leave — the hard part is getting in.
Middle East users lean toward impulse spending; once cultural fit is reached, the whales really spend like water, but rich people may not have time for short dramas — and when they watch, they may just buy an annual subscription.
Europe is the hardest: conservative markets plus fragmented languages. If you really do it, pick one or two countries such as the UK or Germany for a point breakthrough.
The Spanish-speaking market deserves special mention. China Reading only cracked localization after opening a local studio in Spain; the cost is high, local staff work eight hours a day and overtime must be paid overtime.
But the monetization pattern is clear: Spanish users will pay $9.9 for an episode and keep paying that way, they just dislike monthly or annual plans — they buy whichever episode catches their eye.
Shengshu’s Liu Changyan (刘长艳) added a dimension many overlook: overseas content control is extremely strict — copyright, race, family ethics, and especially youth-related content, with many review red lines. Compliance must be budgeted before going overseas.
The map is detailed. But notice: everyone’s map is different, because everyone’s positioning is different. The map is a projection of positioning.
On platform and budget, Shawn’s framework is the most practical.
There are two overseas monetization models: IAP, users paying to unlock; and IAA, ad monetization. His advice: don’t start with the platform, reverse-derive from your content.
Making premium content and in it for the long run? Go to IAP platforms. Making batch content — at least five dramas to start — publish to TikTok-native and TikTok Minis, and eat the IAA traffic pool.
YouTube’s payout cycle runs as long as three months; if you cannot keep shipping batch content, traffic may fall off a cliff and later revenue share suffers.
What if you are a small studio with only 1,000,000 RMB in budget and little AI-drama experience?
Shawn’s answer is a 70/30 split. Put 70% into IAA — the base that pays back and replenishes cash over the next three to six months, so compute the payback rate carefully. Put 30% into IAP — use 300,000 RMB to make five or six dramas and try to open the premium channel. Don’t abandon either line. The reasoning is blunt:
Staking everything in one place is no different from betting on a hit; the outcome is uncertain.
Notice the structure of the answer. He did not say which platform to pick; he gave a logic for allocating risk. Platforms can change; the logic is yours.
The roundtable nominally discussed three questions: which market, which platform, how to split the money.
But stack the four answers together, and they converge on one question: who are you, and which dollar are you earning?
China Reading figured it out, so it dares to pull model vendors and production houses onto its boat. Guanghe Power figured it out, so it is content to be Foxconn first. Agnes figured it out, so it dares to give away its models for free and is not afraid to admit its biggest market is at home.
Domestic short-drama time share exceeds 20%; globally it is under 2%. That ten-times-scale space will draw more and more people overseas. The map will be redrawn again and again; platform policies will change again and again.
The only thing that cannot be copied, and never expires, is your own judgment of your positioning.
Hanson Zhang (张毅泷), Co-founder, Agnes AI
Shawn Xiao (肖伟), Head of Overseas Business, China Reading (掌阅科技)
Jim Zhan (詹卓安), Founder & CEO, Guanghe Power (光盒动力)
Liu Changyan (刘长艳), Southwest Regional Business Lead, Shengshu Technology (生数科技 / Vidu)
Zhao Liang (赵亮), Partner, Unique Capital (非凡资本)
Agnes Video 2.5 list price 0.15 RMB/sec, real deals ~0.07 RMB/sec; Flash tier free at 720P, rate-limited.
Agnes video model: ~30,000 minutes of calls per day; Pavo adds ~1,000 DAU daily with no paid promotion; ~15% of daily generated videos come from paying users.
China Reading: S-tier script guarantee up to 100,000 RMB; B/A scripts 10,000–30,000 RMB; finished dramas procured 500+/week; top S-tier exclusive 80,000–100,000 RMB; production 400–1,500 RMB/min; QC ladder B→B+→A+ raises per-minute rate 400→600→700 RMB.
Shengshu Vidu S-series world model: digital-human response as fast as 3.2 seconds; S2 split into Avatar (real-time interaction, larger movements, stability) and Editing (real-time background/outfit/prop swap and style transfer). A new MoE-architecture generative video model is expected in early October.
Short-drama time share: domestic China >20%, Latin America ~15%, North America ~5%, global average <2%.
Spanish users pay ~$9.9 per episode, episodic rather than subscription.
Small-studio budget rule of thumb: 70% IAA / 30% IAP.
