---
title: "This AI Company Helps US Creators Sell on TikTok — Racing Toward $10M Annualized Revenue in a Year"
author: "Unique Research"
sourcePublication: "Unique Research Substack"
originalPublishedAt: "2026-10-28T01:49:43+00:00"
canonical: "https://ffcap.cn/en/research/this-ai-company-helps-us-creators"
source: "https://uniqueresearch.substack.com/p/this-ai-company-helps-us-creators"
language: "en"
---

# This AI Company Helps US Creators Sell on TikTok — Racing Toward $10M Annualized Revenue in a Year

[![](https://substackcdn.com/image/fetch/$s_!KieB!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Feceab680-722c-4e05-b891-c3fb37b4d12f_1280x720.jpeg)](https://substackcdn.com/image/fetch/$s_!KieB!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Feceab680-722c-4e05-b891-c3fb37b4d12f_1280x720.jpeg)

_Unique Research · AI Industry Watch · October 26, 2026_

7 million US creators hold TikTok Shop cart rights. Only about 100,000 can really sell. The rest — 6.9 million — mostly sleep. About 6.5 million have never closed a single order.

Nearly 10,000: that is how many ordinary Americans a ~40-person Chinese company has already walked through their first-ever sale. They do not even have to cut the videos themselves.

The company is **K2 Lab** (攀峰智能). The product is **Moras** — a Personal Commerce Agent that lives on a creator’s phone and runs the full affiliate-commerce loop automatically.

The industry is shouting cold. This company is stepping on the gas. The gap is all in the business model.

Last time Unique Research sat with founder **Wang Ming** was April. This time he opened with: AI applications have cooled faster than they expected.

He had assumed AI risk would transmit from the secondary market into the primary market first. Instead, secondary-market risk had not arrived yet — applications cooled first. Most apps, in his view, were not aiming to create value; they were helping users burn tokens. Users buy expensive tokens and get no result. That business is hard to sustain.

His own numbers tell a different story:

-   **July** — month two of public beta — annualized revenue crossed **$1 million**.
    
-   **August** — crossed **$2 million**.
    
-   **October** — racing toward **$10 million** annualized.
    

GMV is more concrete: the first public-beta quarter did **$10 million** GMV; Q4 is sprinting toward **over $10 million GMV in a single month**.

Moras’s playbook runs against most AI apps on the market.

You are a US creator with 5,000+ followers. You authorize your TikTok account to Moras. Product selection, creative, script, video generation, editing, copy, cart attachment, publishing, data analysis — the full chain is Agent-run.

You can leave it fully managed for a year, or talk to it more — feed it your background, taste, and lifestyle. The more Context you give, the more precisely it sells for you.

Most assets are AIGC-generated plus editing: 3–6 posts a day, thirty to fifty a month.

Where does the money come from? Products use merchant links from the TikTok affiliate network. Merchants pay commission; Moras and the creator split it — the creator takes about **30–40%**, Moras takes the majority.

The key: the creator pays **zero** from start to finish.

They did try subscription. Early on it was subscription plus commission. Then they found that charging subscription itself was driving people away. Originally they deducted an extra **$20** after a result landed; later even that $20 was cut.

Now there are two modes:

-   **Creator hires AI** — pure split. Roughly **80–90%** of users.
    
-   **AI hires human** — **$150/month** base salary plus a small commission; the creator follows instructions. Roughly **10–20%**.
    

A counter-intuitive finding: they expected overseas users to rush into being hired by AI — a guaranteed floor. The opposite happened. Over 80% chose to hire AI.

Wang Ming: if these were Chinese users, most would almost certainly pick the base-salary track for certainty. US users’ autonomy was unexpectedly strong — they still want their own agency.

Those who chose to be hired by AI still collaborate under AI instructions and respond when called.

Why does this business work? Look at the odd shape of US content commerce.

In China, a top streamer can move hundreds of millions in GMV alone. The US has no equivalent head; the affiliate ecosystem leans on mid-tier creators.

Of 7 million eligible creators, only about 100,000 really know how to sell. The remaining 6.9 million want to make money — but do not want to, and often cannot, put in that effort.

Where does the gap sit? Not laziness about filming. Wang Ming breaks it down carefully: a commerce video and a daily video are different species.

You walk the dog as usual, then suddenly say “check out the food our dog eats” — users do not buy, and the platform cannot latch onto you, because there is no commercial hook.

Product selection, script, trends, publish timing — which combination yields predictable returns is a full body of know-how. Ordinary people have nowhere to learn it.

So Moras’s most valuable metric is **video order rate**. At the start, only **3–5 out of 100** videos produced an order. Now the best creator cohorts clear **30%+**; overall average sits in the **teens of percent** — about **10×** the platform average.

Creator-level order rate is also about **50%**: of ~20,000 active creators, ~10,000 have ordered. Co-founder **Tang Minglei** adds a detail: of those ~10,000, **95%+** had never closed an order before — their first sale happened on Moras.

Creator profile is clear: mostly ordinary people with a few thousand followers; moms are an important cohort; an extra **$100–200/month** is real money.

The pyramid tip exists too:

-   Fastest creator: **$200k+** GMV in two months.
    
-   One creator went from **2,000 → 20,000** followers while doing **$200k GMV in a single month**.
    

By GMV: the top **10–20%** of creators can do tens of thousands of USD GMV per month. By income: about **10%** earn around **$1,500/month**; the top tier clears **$10k+/month** — corresponding to **$100–200k GMV/month**.

The merchant side is shifting too. Moras used not to need direct merchant relationships; now merchants come to them, because Moras has been leading TikTok Black Friday institutional ranking contests.

Merchants are willing to grant directed high commissions: the open pool may hang at **5%**; significantly higher rates go to Moras. High-commission coverage was under **5%** last month, hit **20%** this month, with a **30%** target next month.

Why can’t creators leave? Tang Minglei’s explanation: individuals have no negotiating power. In the open catalog you never get that price.

More important than the growth curve is the roller-coaster of **unit economics (UE)** — defined here as retained commission ÷ compute cost.

In the small-scale ops period, UE was extremely high: **$1** of compute returned several to tens of dollars in commission — almost unreal.

Acquisition then was Cold Email; a few ops people taught creators one-to-one how to use the product — even helped craft personas by hand.

When public beta opened in **June**, hundreds of creators poured in and UE crashed below **1**. Without enough human ops, the product was incomplete, users were unattended; some felt creation was so easy they posted 100 videos a day — 90%+ got throttled by the platform.

Looking back, scale-up itself was a stress test. The conclusion: tools are easy to teach; know-how is hard.

You have to persuade users how to interact with you, what Context to give, and how to understand the hire model. That needs stronger personal-agent capability than a pure visible-interactive tool.

Over the next three months they filled the hole: added coaching ability to the AI, taught creators content-commerce knowledge; gave emotional value — a video with good exposure gets immediate encouragement.

They also had to sell: users keep asking whether the money will eventually reach them, and why their split is the smaller share.

Work that used to be all human is now fully agentic — proactive service, encouragement, emotional value, patient ongoing explanation.

UE quickly returned above 1. Wang Ming’s scale logic fits in one sentence: as long as UE is not losing, he dares to scale. If users churn, they know why, fix it, and scale again.

Next year’s target: annualized revenue toward **$100 million**.

His definition of the business:

_The most important business logic of the AI era is not software — it is manufacturing, new manufacturing. New manufacturing is cost._

If you cannot even cover cost, the business cannot scale. So UE is the first metric they watch. It answers two questions at once: can retained commission cover compute, and is intelligence actually rising. Return consistently greater than investment is what “vertical intelligence going up” means.

Intuition says an AIGC commerce-video company should spend most compute on generating video. Moras’s actual structure is not that.

Of compute cost:

-   **AIGC: 40–50%**
    
-   **Scheduling / content adversarial checks: 20–30%**
    
-   **Long-chain multi-agent collaboration and reasoning: 20–30%**
    

Understanding accounts, understanding content, creation, pre-review, compliance checks, user interaction — the thinking costs more than the drawing.

Wang Ming’s litmus test: to judge a vertical-intelligence company, look at where it spends compute. If 100% or most of the cost is AIGC, it may be a wrapper.

Moras’s business chain runs **15–16 large Agents**, covering content understanding, creative decisions, compliance, and more.

The company is extending Personal Agent capability toward both creators and merchants — but having a back-end Agent system does not mean both end products are already fully open externally.

On cost optimization, Moras is gradually replacing parts of general models with self-built vertical models and distilled smaller models for industry-specific tasks.

Per Wang Ming, that brought all-in production cost per video from about **$3** down to **$1–1.5**.

Training self-built models also costs money, but over time it cuts dependence on general models and improves task-specific quality. Doing only industry harness and skills on top of a SOTA general model may not be good enough.

Video generation models, by contrast, they chose _not_ to train themselves — ByteDance and peers already have large-scale data, and models such as Seedance meet the need.

Moras is a ByteDance T1 discount customer and deploys heavy compute on Tencent Cloud; as scale grew, it also received corresponding resource discounts.

Wang Ming says monthly published videos are trending toward **500,000**.

**Caliber note:** $1–1.5 is a per-video cost estimate; ~500k is a trend description of monthly publish scale. Do _not_ multiply the two to invent a month’s actual compute spend. Whether scaled production truly pays still depends on actual publish volume that month, all-in cost, and the revenue those posts finally bring.

There is also a platform-level attack surface. A common AIGC-video problem is high views, zero conversions — content violates common sense; users watch but do not buy.

So Moras builds its own content adversarial and consistency checks. Platform APIs were too weak, so they built in-house.

Wang Ming’s judgment: _Manufacturing exposure and manufacturing conversion rate are two different roads._

Usually AI companies court platforms. Here it flipped.

In July alone they hosted **100+** visits from TikTok executives at various levels. That same month the platform asked Moras to lift the ecosystem at once — unlimited posting for creators. The team ran the numbers: compute alone would burn about **$6 million**. They declined.

Why the urgency? Two reasons.

First, an ecological vacuum: content platforms put most R&D into ad monetization. In Douyin ads, Julang Qianchuan accounts for over half; the platform looks like a content company but is really an ads company — ecommerce gets fewer resources.

Second, platform leadership KPIs: without content supply, commercialization rates do not rise. Whoever can wake the 6.9 million dormant creators is a platform hero.

Tang Minglei had just raised an API request; a senior leader replied in the group chat: must support, high priority.

The same ask in China flips. Wang Ming asked friends at domestic platforms whether they could do it. The answer: even if you try domestically, we will not let you on — we prioritize living creators making money; we cannot even keep human creators fed, and if they cannot earn they leave for Xiaohongshu.

Tang Minglei’s summary: platforms want GMV to keep growing, and also fear low-quality homogenized AI content flooding in and collapsing the human-creator ecosystem.

Their conclusion is blunt: _Not oppose AI — oppose low-quality AI._ Domestic opportunity is not replacing living creators; it is lowering the creative-expression bar enough.

Put Moras back into the industry and its position sharpens.

In Unique Research’s September global AI subscription-revenue stats, an a16z-style caliber showed: among leading AI-native products, **84%** charge subscription, **64%** sell credits by volume, and only **2%** take transaction commission.

Last-cycle consumer internet was free for users, paid by advertisers. This cycle flipped — everyone charges users a monthly fee.

Our judgment: subscription cannot carry AI into the mass-market second half; ads and transaction take-rates will enter.

Moras is already standing in that 2%. Same lane as Instinct, the personal agent on iMessage: do not charge the user; take a cut on transaction results.

Instinct already runs at about **$1 billion** annualized GMV; card-linked users spend about **$1,300/month** on average. One helps you spend; one helps you earn — both are new answers that skip the subscription question.

Tang Minglei draws a four-quadrant for the Agent track: save money, save time, make money, kill time.

Save-money / save-time is the most crowded; kill-time needs new worlds — AI games, AI social. They picked the most counter-consensus quadrant: make money.

The cost: you must complete a long, complex task — selection, script, publish timing; any broken link collapses the result. The upside is the same: the chain is long; entrants have to walk the whole process again.

As a former investor, Tang also summarized three death causes for this year’s AI-app chill: no moat — a model upgrade covers you; UE not positive — the more you scale, the more you lose; no ecological niche — forever living on someone else’s entry point.

Moras’s answers: model-application integration, pay-for-results, planted inside creators’ transaction chain.

His view: as long as AI apps tell a real story with a real business model, opportunity remains. He has heard that some investors no longer insist on ARR and even want to retreat to P/E — that, he thinks, is too regressive.

Capital is voting with its feet. The company closed **three funding rounds in six months**; the latest was **tens of millions of RMB**.

Near the end of the interview, Wang Ming offered a reversal that stuck.

Today Moras takes the commission majority, because AI contributes the largest value in the chain.

But for high-potential creators with unique taste and deep vertical know-how, he says, it should flip: based on the unique value their unique Context brings, they should take more than Moras.

The logic: today AI supplies the know-how ordinary people lack — so AI takes the majority. The ultimate value of AI is amplifying what makes a person distinct. When a creator’s expression, taste, and expertise become scarce assets, the split should reverse.

Pricing power for tools sits with the tool. Pricing power for productivity eventually returns to what makes a person different.

6.9 million dormant creators are waiting to be woken. The ones who wake all the way through will eventually cost more than the AI.

_Host: Delta Wu (吴畏), founder of Unique Research. Guests: Wang Ming, founder of K2 Lab / 攀峰智能; Tang Minglei, co-founder._

**On cooling:** Wang — applications cooled before secondary-market risk arrived. Most apps were not creating value; they were burning tokens. Surviving companies will be vertical-intelligence firms with ROI capability, not token-consumption helpers. Coding is a good scene for efficiency, but globally it has also cooled; large firms are limiting usage because token cost is approaching wage levels without incremental return. Short-form drama still works because consumers actually pay for the output.

**On “AI MCN”:** Wang — they are not an MCN and do not require creators to sign exclusively. Creators can partner with any MCN; on specific products, merchants grant Moras directed commission. Overseas MCNs generally also take commission; the China ecosystem differs — almost all top affiliate creators sit inside MCNs.

**Core benchmark:** Wang — first metric is compute ROI: can retained commission (after creator split) cover compute. AI-era logic is manufacturing, not software. If return does not stay above investment, vertical intelligence is not rising.

Wang: still hard to call it “law” — better call it capability. Matching improved: which creator persona fits which product, which script, which trend, which publish window — expected commercial outcome can be estimated. Started at ~3–5%; some cohorts now fit to 30% — ~10× platform average.

**P&L math:** GMV ≈ video count × video order rate × average GMV per ordering video; then × commission rate (which has been rising). Retained commission after creator split ÷ compute cost = UE; priority is whether the remainder after paying creators covers compute. If not, you are losing.

Co-creation period: UE very high ($1 compute → several to tens of dollars commission). Cold Email + point-to-point ops. June public beta invited hundreds; service and product maturity could not keep up; UE fell below 1 within weeks. Some users posted ~100/day; 90%+ got low reach. Platform does not want one account flooding. Ops then added publish strategy, lifecycle management, and coaching — UE recovered above 1 within about a month.

**Split:** ~30–40% to creator; Moras majority. Many users ask whether money will reach them and why the minority share — that sell used to be human. For high-potential users, Moras intends to let them take more; AI should amplify human distinctiveness.

**China:** Wang — cannot run the same play there; risk control. Needs a real US-region account and a local person. First beachhead is the US. Platform feedback: ~7M eligible / tested creators; ~100k can really sell; ~6.9M dormant — many want money without effort.

**Why DIY commerce videos fail:** daily content with a product mention has no commercial hook; platform push cannot latch; many creators do not know commercial-video structure; some products simply cannot sell. Content commerce on Instagram, YouTube, or TikTok needs persona–content–product–timing match.

**Fast cases (all USD):** co-creation period — some registered and cleared $10k+ GMV the same day on a hit SKU. Fastest: $200k+ GMV in two months. One creator at 2k followers hit $200k GMV/month and grew to 20k followers.

**Ordinary vs top:** pyramid — only ~10–20% reach tens of thousands USD GMV/month; many others land thousands USD GMV/month (~$100–200 income at ~3% effective take). Top income $10k+/month needs $100–200k GMV. Overseas users give less Context than Chinese users unless triggered by a sale or withdrawal prompt (”want $200–300 next month?”).

**Anti-abuse:** new users get a ~100-video quota; Moras coaches daily strategy via language interaction and sets GMV benchmarks to unlock the next stage. Quotas cannot be bought — pure CPS. Even if users want to pay subscription now, Moras refuses. Insight: reverse of classic tool apps — give quota first, unlock more after results; once a company can deliver results in a scene, it need not stay a tool.

**Hire modes:** expected laziness to drive “AI hires human”; reality — 80–90% still hire AI; 10–20% take $150/mo base + small commission and follow instructions (often higher GMV when highly obedient). Wang expected Chinese users would all pick the salary track; US autonomy surprised him.

**Subscription cut:** most users were on sub+split; subscription removed to cut friction after they could cover costs. Even the old “extra $20 after result” (not Stripe / App Store subscription) was later cancelled.

**Merchants inbound:** Moras leads TikTok Black Friday institutional contests; platform classifies them as a TSP-type selling entity. Merchants offer directed high commission (open pool ~5%; directed 20–30%). High-commission coverage: <5% last month → 20% this month → 30% target next. Individuals have no negotiating power. Moras charges merchants no placement fee — unlike many TSP agencies.

**Compute mix (also in Q&A):** AIGC 40–50%; scheduling/adversarial 20–30%; long-chain agent collaboration/reasoning 20–30%. Current UE about **1.2–1.5**; compute is **70–80%** of cost. Per-video cost ~$3 → $1–1.5 (~50% down). Monthly publish trending toward 500k; Tencent Cloud major customer with special discount; ByteDance T1 discount. (Again: do not multiply per-video cost × 500k into a monthly spend figure.)

**Views without conversion:** platform API for compliance/consistency was weak; Moras built its own. Goal: differentiation, consistency, fewer physics violations. High play with no commerce is common for pure AIGC; exposure manufacturing ≠ conversion manufacturing.

**TikTok stance:** actively promoting Moras as filling an ecological niche. July: 100+ executive visits in Hangzhou; same-month ask for unlimited posting declined (~$6M compute). Scale only while UE stays positive. Domestic platforms may reject AI-hosted accounts when human creators are still “enough”; overseas content supply is short. Tang: platforms want GMV growth but fear AI homogenization; he has heard a ratio requiring only ~20% of videos to be AI. If AI low-quality share rises and human creators leave, platforms cannot accept that. Domestic ecosystems will not forever reject AI — but oppose low-quality AI, not AI itself.

Tang’s four quadrants: save money / save time / make money / kill time. Crowded on the first two; kill-time needs new worlds. They chose make-money — long-chain complex task is both the pain and the moat. Three chill concerns: no moat under model upgrades; UE not positive at scale; no own niche. Moras explores model-application integration, results-based UE recovery, and A2A entry points. Hot tracks mint $10B+ valuations; AI-app PARR this year is softer. Retreat from ARR to P/E feels regressive. Vertical population / vertical scene entry points still have room if the model is designed and the data verified.

_Source: Unique Research WeChat essay, October 26, 2026. Mid 2247523141 · idx 1 · sn 6a30db04d96aed5b4326a75897933e01._

---

Original publication: https://uniqueresearch.substack.com/p/this-ai-company-helps-us-creators
On-site reading page: https://ffcap.cn/en/research/this-ai-company-helps-us-creators