---
title: "$100M Raised, $2.6B Valuation, 24% Drop: This Overseas AI App Rewrites the Definition of \"Growth\""
author: "Unique Research"
sourcePublication: "Unique Research Substack"
originalPublishedAt: "2026-06-20T15:00:17+00:00"
canonical: "https://ffcap.cn/en/research/src-20260620-01html"
source: "https://uniqueresearch.substack.com/p/src-20260620-01html"
language: "en"
---

# $100M Raised, $2.6B Valuation, 24% Drop: This Overseas AI App Rewrites the Definition of "Growth"

_Original · Unique Research · 2026-06-20_

_Editor's note: The first-person market commentary and its judgments belong to the original Chinese author. This English rendition retains the Genspark hook, the three structural signals, all product-level traffic and growth figures, the four-layer growth framework, the report's promotional and methodology sections, and the research disclaimer. Product rankings and figures are Unique Research's stated May data attributions, not independently verified metrics._

AI Industry Observation

May overseas data: 3D modeling erupts across the board, Agents double, smart search gets cut in half.

AI products are going through the critical leap from "fun" to "useful."

"

A product's traffic dropped 24% and its active users fell by a third, but it just raised $100 million at a $2.6 billion valuation.

You read that right, and I didn't write it wrong. This is Genspark, founded by former Baidu executives Liang Jing (景鲲) and Zhu Kaihua (朱凯华). Capital markets and product data are tearing apart on this company — investors are betting on its long-term story of transforming into an "AI Agent workspace," but users voted no with their actual behavior: traffic and activity both fell, and the drop wasn't polite.

This contrast was the finding that stunned me most in May's overseas AI data. It let me truly see that in the AI industry, capital narrative and user choice may be separated by a huge chasm.

We at Unique Research (非凡产研), together with Wu Xiaobo Channel, now publish a monthly "Global AI Growth Report" covering industry developments, product trends, and thought-leader views, hoping to capture frontier signals in the AI wave.

As usual, let's start with the overall overseas ranking. PolyBuzz is No. 1 again; it's held the top spot in the emotional-companion sector for so long I've lost count. But what I really want to talk about this month isn't it. The most worth-watching in May's data are three structural signals: a collective eruption in 3D modeling, an unexpected rise in the Agent sector, and a full retreat of smart search. Put together, these three signals point to one thing — AI products are going through the critical leap from "fun" to "useful."

3D Modeling, the Strongest Sector in May

To be honest, I couldn't quite believe this data when I saw it, because it has risen two months in a row. 47% — what does that mean? For a normal SaaS product, growing 10% is already popping champagne.

Meshy: Taiji Graphics' 3D generation tool, 8.97 million visits in May, +47.43% month over month.

Tripo: from VAST, 6.44 million visits, +47.32%.

Rodin: from Yingmei Technology (影眸科技), 1.3 million visits, +47.35%.

Three top products all rose 47%, precise as if they'd agreed. But this is by no means a coincidence. In plain language: they grew by nearly half.

On one hand, the embodied-AI wave pushed 3D assets onto the trend. Humanoid robots, autonomous driving simulation, industrial digital twins — these scenarios need massive 3D models, and traditional hand-building simply can't keep up. On the other, gaming and film are accelerating their embrace of generative 3D — cutting a character model from two weeks to two minutes is a temptation no one can resist. I've seen Tripo and Meshy's product power; input text or an image, and the resulting model goes straight into Unity and Unreal — not those "looks pretty but unusable" toys.

A deeper trend is that generative AI moving from 2D to 3D is only a matter of time. Image generation is already blood-soaked with involution, video generation is still burning cash on compute, and 3D is the sector with the most real productive value while its technical maturity has just crossed the usable threshold. At May's Google I/O, Gemini Omni released video generation, good for the whole 3D content ecosystem — the boundary between video and 3D will blur, and ultimately it's all "spatial content."

Yingmei's Rodin, though only 1.3 million in size, also grew 47%+, showing the dividends in this sector are broad-based, not only eaten by the top. It's a bit like the early image-generation sector — a few leaders first educate the market, then the whole pie expands together. I'm bullish this sector can surge at least one more quarter, for a simple reason: downstream demand is too rigid. As long as embodied AI keeps burning investor money, demand for 3D assets won't fall. This is a classic "selling shovels" business.

Video Generation Starts to Stratify

Though seedance 2.0's model is very hot, among overseas AI apps, KLING AI, with 12.16 million visits and 4.96 million actives, still holds the top spot. It seems Kuaishou's video model really does hold up overseas on product power; users vote with their feet.

But what's more interesting isn't who arrived, but who's rising and who's falling — this divergence is especially telling. Pollo.ai hit 7.7 million and rose 15.9%, Pika 2.24 million +19.6%, Vidu 2.85 million +34.8%, Hailuo AI 2.55 million +13.1% — all going up. On the other side, HeyGen 10.55 million fell 6.2%, Dreamina 9.96 million dropped 19.5%, PixVerse 6.25 million also fell 5.5%.

That's interesting. You can't say "the video sector has cooled," because the risers are rising nicely. What's really happening is users have started to be picky. From early curiosity about "who can generate a more jaw-dropping demo," it's turned pragmatic: "after it generates, can I actually use it?" The mid-tier surges explain it even more — CrePal +110%, Morph Studio +69.3%, Vivago +50.6%; none is a generic "video generation tool," but each has plunged into concrete workflow scenarios.

General-purpose ones are fluctuating; vertical workflow ones are emerging.

This signal is actually of a piece with the 3D modeling logic: users don't want "something that moves," but "something that can get a certain stage of my work done." The watershed for the video-generation sector may start this month.

Agents Keep Pushing

If 3D modeling is the "unexpected dark horse," Agent is the "long-planned explosion."

Coze: ByteDance's AI Agent platform, 690,000 visits in May, +141.98% month over month; active users 370,000, +129.48%.

Dola AI: also ByteDance's, 7.35 million visits, +64.61%, active users 2.7 million, +33.02%.

MuleRun: 910,000 visits, +36.08%.

This growth isn't "growth" anymore; it's "takeoff." I stared at this data for a long while to confirm I hadn't misread the decimal point.

The driving force behind is clear. May saw dense releases from three big models: OpenAI's GPT-5.5, Google's Gemini 3.5, Anthropic's Claude Opus 4.8. The three giants seemed to agree on dropping big moves in the same month. On the surface it's an arms race of model parameters, but for the Agent sector the meaning is entirely different — Agent's core bottleneck has never been how intimidating the parameters are, but how "obedient" the model is. After this update, tool-calling accuracy, long-horizon task planning, and multi-turn dialogue consistency all improved materially; Agent finally went from "PPT concept" to "actually getting work done."

ByteDance's layout is interesting. Coze targets the developer ecosystem, Dola AI targets the consumer end; two lines run in parallel, one building the stage and the other performing. About this dual-line strategy I have to say: shrewd. Look closely and you'll find Coze reels in developers to build the ecosystem while Dola runs volume and builds reputation on the C-end; both lines ultimately feed back into ByteDance's own large-model capability. Among China's big tech, ByteDance's AI product instinct has always been the sharpest, and this Agent-sector positioning proves it again. Tencent and Alibaba are still watching; ByteDance has already run a mile ahead.

Who's Cooling Down?

Wherever some rise, others fall — and the falls are brutal.

flowith, the smart-search product, 680,000 visits in May, -52.85% month over month, active users 290,000, -60.29%. Cut in half — no, cut at the knee. Genspark isn't much better; 15.35 million visits looks okay, but -24.01% month over month, active users 3.76 million, -33.32%. Active users directly lost more than a third; that drop is alarm level anywhere.

But something especially twisted is happening to Genspark. In the very month its data fell, it closed a $100 million B-round extension at a $2.6 billion valuation, with investors including Sozo Ventures, Weicheng Capital (威诚资本), and South Korea's Mirae Asset (未来资产). This is the fourth B-round extension, with cumulative funding over $645 million.

Capital isn't betting on the "search" sector at all, but on the new narrative of an "AI Agent workspace." The story Liang Jing and Zhu Kaihua want to tell is: Genspark isn't just a search engine, it's your AI collaboration platform for all kinds of tasks. This story is sexy enough to persuade investors to open the checkbook.

The only problem is: users don't seem to buy it. Traffic fell 24%, actives dropped a third; products in transition usually fluctuate, but a drop of this magnitude is hard to explain as "normal adjustment."

This is the most typical tension in AI right now — investors look at the story three years out; users look at today's experience. When a crack appears between capital narrative and product reality, history's side is obvious. But before then, plenty of people behind that $2.6 billion valuation are probably losing sleep.

Oreate AI, an efficiency tool, 1.75 million visits, -36.65%. Joyland, the junior in emotional companionship, 2.21 million visits, -26.31%. On image generation, ByteDance's Dreamina 9.96 million visits, -19.53%; Lovart 3.82 million visits, -16.26%, dzine also fell 25.4%; the tail is going through a shakeout.

The retreat of smart search isn't hard to understand. The large models' own direct-answer ability keeps getting stronger; do users still need a "middleman" to search for them? ChatGPT, Claude, and Gemini web versions are already eating into traditional AI search share. flowith and Genspark's plight isn't bad products, but their "value gap" is being filled by the models' native ability. It's like how vertical search engines were crushed by Baidu back then — it's not that you didn't work hard, but the giant's foot is already on your face.

The image-generation picture is more complex than "collective decline." True, Dreamina fell 19.5%, dzine dropped 25.4%, Lovart fell 16.3%; many products are ebbing. But look at the other side: SeaArt AI 18.29 million visits roughly flat, while active users actually rose 5.8%; OpenArt 10.91 million visits +19.4%, a striking growth rate in a mature sector; PixAI.Art 9.74 million also edged up 3.7%. On image editing, Cutout Pro 9.91 million roughly flat, Fotor 5.77 million actives slightly down 3.8%.

This shows image generation has entered maturity; underlying capabilities converge, and traffic is concentrating toward brands, communities, and creator ecosystems.

Products with differentiated operations still have room; OpenArt's growth is the best proof. Midjourney, Stable Diffusion, and DALL-E have eaten up the basic market; latecomers who only wrap APIs will indeed find it harder to gain a foothold. But products that can build communities and creator economies are still growing.

PolyBuzz sitting steady at No. 1, I think, has an under-appreciated underlying logic. People say they want AI to help them work, but the first thing they tap is usually the AI that can chat with them. PolyBuzz is strong not because it's more "AI" than others, but because it's more like an emotional-consumption platform. Users come back not because model parameters rose again, but because of the relational feel, the narrative feel, the immersion — things you can't find in technical metrics, but show up vividly in open rates. Model capability is just an entry ticket; what truly keeps people is that "feeling."

There's an interesting divergence in the emotional-companion sector. The top few — PolyBuzz, CrushOn.AI, Talkie — are as steady as diehard households, but small players are starting to fall behind. Joyland fell 26%, and HiWaifu's data doesn't look good either. This actually fits the universal pattern of social products: once users' emotional investment and social relation chains settle, switching costs are extremely high, and the network effects of top products keep strengthening. This sector is moving from "a hundred flowers blooming" to "the top takes all." Users' time, energy, and emotional investment are all finite, and winner-take-all is especially cruel here. In the coming months we may see more small players acquired or simply exit.

The generic-efficiency-tool category is going through a quiet "platform engulfment." Monica 7.34 million fell 6.8%, Sider 3.53 million fell 8.9%, PopAi active users dropped 37.9%. Their shared plight is: users do need these features, but increasingly they don't need a standalone product to provide them. ChatGPT, Gemini, and Claude are frantically adding features; browsers and office suites and phone operating systems are also embedding AI. Summarize, rewrite, translate, search — these once-independent startup opportunities are now being swept up as a matter of course by the big-model platforms and operating systems. Teams building generic efficiency tools have to ask themselves: if tomorrow Chrome or ChatGPT built in your core feature, what's left of you?

Layered Growth Has Begun

Putting the risers and fallers together, a four-layer structure becomes clear.

The first layer is high-frequency consumer scenarios: emotional companionship. PolyBuzz, CrushOn, Talkie, Emochi — these products have extremely high open frequency, but run on "emotional consumption" logic, not "efficiency tool" logic. Their core asset is the relation chains and emotional investment users have accumulated; model capability is just an entry ticket, and relational feel is the moat. This layer's trait: the top takes all, extremely strong network effects, small players will find it harder and harder.

The second layer is production tools that have hardened — 3D modeling, video workflows, audio. Meshy, Tripo, Fish Audio, CrePal are erupting because they solve real problems whose output plugs directly into professional pipelines. The 47% growth isn't propped up by "user curiosity," but by game developers rushing project deadlines and video creators rushing delivery milestones. This layer's trait: high technical barrier, but once it crosses the usable line, stickiness is extremely strong and willingness to pay is highest.

The third layer is Agent and workspace with platform-level imagination. Coze, Genspark, Manus, Dola — this line carries the greatest long-term imagination, but also faces the cruelest divergence. ByteDance's dual-line layout is taking off, while Genspark and Manus's voice is receding. It shows the "Agent platform" story isn't for everyone; users only recognize products that actually get workflows running. The competition in this layer has only just begun; the final winner may not have appeared yet.

The fourth layer is generic efficiency tools squeezed by platforms. Monica, Sider, PopAi — summarize, translate, rewrite are caught in a pincer between the models' native ability and operating-system built-in ability. Their standalone product space is narrowing fast: either find deeply rooted vertical scenarios, get absorbed, or disappear.

These four layers aren't static; the more time users spend in layer one, the stronger their productivity demand for layer two; if layer three works, it will encapsulate layer two's capabilities while swallowing layer four's features. Seeing which layer you're in matters more than seeing which sector you're in.

China's AI-going-global playbook is also changing. Early on everyone competed on model, speed, and cost; now it's gradually shifting to competing on scenarios, operations, and commercialization. Models get replaced; scenario understanding and distribution capability don't easily. Meshy and Tripo's deep work in 3D, KLING AI's breakout in video, Coze and Dola's positioning in the Agent platform, emotional-companion products' precise read on user emotion — these are the real moats. Good products increasingly look like a "scenario company," not a "model company."

"

But what this May data tells me intuitively is: the wind is shifting. Not from large models to larger large models, but from products that "amaze you" to products you "can't live without."

But what this May data tells me intuitively is: the wind is shifting. Not from large models to larger large models, but from products that "amaze you" to products you "can't live without." AI products don't lack capability anymore; what they lack is a reason to be continuously opened. This shift may matter more than any model release. WAIC, the World AI Conference, opens in Shanghai on July 17; by then the heat map of AI going global may look different again. I'm watching.

Unique Research (非凡产研) — the authoritative source of global AI market insight.

The detailed May 2026 AI overseas Web rankings published by Unique Research are as follows:

AI overseas website traffic ranking: the TOP 100 list of overseas AI Web products by visits and monthly active users.

All ranking lists are simultaneously provided with growth rankings, to help investment institutions discover promising projects and founders and developers discover potential opportunities.

Our influence and recognition:

Unique Research's data and insights have gained broad trust and citation from top domestic and international investment institutions, academic bodies, and media, becoming an important baseline for their market analysis, investment decisions, and trend judgment. Our partners and citing parties include (but are not limited to):

Brokerage investment banks: Sinolink Securities (国金证券), China Securities (中信建投), Huayuan Securities (华源证券), Haitong Securities (海通证券), etc.

Authoritative media: China Securities Journal (中国证券报), TMTPost (钛媒体), Jiemian News (界面新闻), etc.

We firmly believe that only data that survives verification by core industry participants truly has long-term value.

We believe in the power of open source and openness; Unique Research hereby makes three major commitments:

This AI product ranking is permanently open-source and free, with complete methodology and raw data supporting full reproduction.

The data sources for this AI product ranking are strictly limited to independent third-party monitoring platforms; enterprise self-reported data is not accepted as an evaluation basis.

The ranking system of this AI product ranking remains absolutely neutral and does not accept any commercial cooperation or paid behavior that interferes with ranking results.

How to use this ranking?

If you are an investor/analyst: this ranking gives you a pure data baseline that survives reverse-engineering verification — a reliable foundation for building investment models, discovering underwater projects, and cross-validation.

If you are an AI founder/product manager: you can not only precisely locate competitors, but deeply reproduce our methodology to analyze your own product's market performance and optimize iteration.

If you are a researcher/student: its fully open-source nature makes this report an excellent free material library for your academic research and market analysis. We encourage innovative research based on this data.

For the complete data behind the charts and rankings shown in this article, please visit 100aiapps.cn. All visible data can be copied and is open and free for non-commercial uses such as personal study and research; commercial use requires prior authorization from Unique Research. We welcome ranked teams to apply the rankings to business plans, recruitment promotion, and other scenarios under compliant use, providing authoritative evidence for your strength.

Join the AI think-tank open-source community

Unique Research not only focuses on AI industry research, but looks forward to building an open-source community that promotes transparency of AI market data. We invite developers, data scientists, and industry analysts:

Reproduce and verify: visit our Feishu knowledge base for complete data and methodology, verify our results, and offer improvements.

Co-build: contribute code to help us expand data dimensions (such as API-call-volume estimation, agent monitoring models, etc.).

Exchange insights: join our Feishu group to share original insights derived from Unique Research data with peers.

One person's insight is limited; a group's wisdom is boundless. Let us together define the standards by which the AI market is measured.

01\. Overseas AI WEB Ranking

1.1 Overseas AI WEB TOP 100 by Monthly Visits

Based on its independent research on the AI industry and continuous tracking of thousands of AI companies and nearly ten thousand AI products worldwide, Unique Research ranks by the month's cumulative visits of each AI WEB product, compiling the top 100 AI WEB products by Chinese or overseas-Chinese entrepreneurs targeting overseas markets, ranked by visits. The specific ranking is as follows:

1.2 Overseas AI WEB TOP 100 by Monthly Active Users

Based on its independent research on the AI industry and continuous tracking of thousands of AI companies and nearly ten thousand AI products worldwide, Unique Research ranks by the month's active user count of each AI WEB product, compiling the top 100 AI WEB products by Chinese or overseas-Chinese entrepreneurs targeting overseas markets, ranked by active users. The specific ranking is as follows:

02\. Research Statement

2.1 Data Description

The AI product revenue data involved in this report, articles, rankings, or charts is based on real-time tracking of redirect traffic from AI product official websites to payment gateways (such as Stripe), combined with time-series analysis to deeply examine historical visits, monthly active user fluctuations, and price-strategy changes, thereby building a dynamic revenue-estimation model. The model's core parameters are determined through a triple-validation system of paid-conversion-funnel monitoring, reverse calibration of enterprise-disclosed data, and comparison with industry-benchmark conversion rates, and are promptly compared and corrected against corporate financial reports, funding announcements, and authoritative media disclosures. Through sampling verification, the average absolute error rate between predicted and actual data is stably within ±10%, reaching industry-research-level accuracy.

The AI product traffic data involved in this report, articles, rankings, or charts is based on multi-source data integration and Unique Research's intelligent algorithm processing, with monitoring strictly limited to users' direct access through official apps and websites. Specifically, this data covers only users' direct access to AI products through websites and native apps, and does not include traffic from the following access methods: browser plugins/extensions, desktop client software, WeChat/Alipay mini-program ecosystems, embedded services in third-party platforms such as Discord, local deployment of open-source models, API calls, and other non-direct-access scenarios. This data focuses on monitoring the core end-side access behavior of AI products, aiming to objectively reflect the direct usage of mainstream user terminals.

Given the complexity and dynamism of data collection and processing, and the constant change of market conditions, the data shown may contain a certain degree of error and omission. Therefore, this data should be regarded as a reference basis for research and analysis, aiming to give users a window to deeply understand the AI product market, rather than a definitive basis for directly formulating specific investment strategies or providing advisory recommendations.

Data coverage expansion note:

We are actively developing and testing monitoring models in the following emerging areas, which are expected to be gradually incorporated in future quarterly reports:

AI Agents ecosystem

AI hardware device sales and activity

Large-model API call volume estimation

Stay tuned.

2.2 Concept Definitions

• Geographic dimension

Overseas AI products: products founded by non-Chinese entrepreneurs or teams and primarily launched for global markets (including but not limited to their home market).

Domestic AI products: products founded by Chinese entrepreneurs or teams and primarily launched for the Chinese mainland market.

Going-global AI products: products founded by Chinese entrepreneurs or teams but with their primary market target set to overseas markets (non-Chinese-mainland markets).

• Functional dimension

AI-native applications: applications that deeply integrate AI technology and algorithms from product design, whose core value, business processes, or user experience are entirely built on AI. These applications don't merely use AI to optimize or enhance existing features; they make AI their core component — without AI, these applications would cease to exist or lose their core value.

AI-feature applications: applications that, on top of existing business logic and application frameworks, integrate or embed AI to enhance existing features or provide new ones. These may already have mature business models and market positioning, but by introducing AI they significantly improve efficiency, user experience, or create new value points.

AI-ecosystem applications: platform-type applications that support, connect, or facilitate exchange among the AI ecosystem of AI researchers, application developers, service providers, and end users. These span technology, market, community, and resource sharing, aiming to promote AI's popularization, innovation, and cooperation.

2.3 Metric Definitions

• WEB data metrics

Visits: the sum of all visits to a website over a given period (such as a month), used to measure website traffic scale. Page views within a continuous active session count toward the same session, while a user re-activating after a gap of 30+ minutes or at the start of a new day counts as a new visit, thereby comprehensively and accurately measuring website traffic and visitor activity.

Unique Visitors: the number of distinct IPs visiting a target website over a given period (such as a month). If someone visits a website on multiple days within a month, they are counted as only one unique visitor.

ARR (subscription revenue): the total revenue a website earns over a period (e.g., a year) from subscription services provided to users, excluding one-time items such as advertising revenue, transaction commissions, and professional services.

• APP data metrics

Downloads: the number of times an app is downloaded by users over a given period (such as a month), usually counted as the download count of a single app on app stores (such as the App Store, Google Play). Downloads are a key indicator of an app's popularity and user appeal, and can reflect the effectiveness of app promotion and marketing.

Active Users: the number of distinct users who perform at least one activity over a given period (such as a month). Active users measure user participation and activity over a specific period, typically used to assess an app's user scale and engagement.

IAP (in-app purchase revenue): revenue from users purchasing virtual goods, extra services, premium features, etc. within the app, excluding advertising revenue, direct user payments (such as tips), and revenue from third-party Android app stores.

2.4 Disclaimer

This report is published by Unique Research (非凡产研); the copyright belongs to Unique Research. Any Chinese reprint or citation must note the report source; foreign institutions wishing to reprint or cite must contact us in advance for authorization. If citing data, please mark: \[Data source: Unique Research\]. If directly citing a chart, please mark: \[Data source: Unique Research, chart production: Unique Research\].

This report is Unique Research's independent original analysis as a third-party institution; its content does not represent the position of any enterprise and does not constitute investment advice to anyone. Investors should therefore note that any investment decision made on this basis is unrelated to Unique Research, its employees, or affiliated institutions.

To the extent permitted by law, Unique Research and its affiliated institutions may hold equity in companies mentioned in the report, or provide or seek to provide fundraising or financial advisory services for them, and its employees may serve as directors of companies mentioned in the report.

---

Original publication: https://uniqueresearch.substack.com/p/src-20260620-01html
On-site reading page: https://ffcap.cn/en/research/src-20260620-01html