Summary

The GenAI leaderboard is unstable at the edge but increasingly durable at the center. January 2025 Similarweb visits and Sensor Tower monthly active users produced 17 new web entrants, while 16 web products have appeared in all four editions. For investors, products can still “rewrite the leaderboard overnight,” yet repeated rankings show that lasting consumer brands and businesses already exist.
Smith notes that consumer activity typically lags research by “six to nine to 12 months,” while Acharya still places many consumer-AI categories in the early-adopter phase. Current video systems make strong 3-, 5-, or 6-second clips, not complete movies, and consumer voice, browser-agent, and screen-aware products remain relatively underdeveloped at the application layer. Acharya’s deeper call is that intuitive assumptions may invert: AIs might prove “more human than humans,” or organize and delegate work to people rather than merely receive it.
DeepSeek showed that another horizontal assistant can still acquire mass distribution almost instantly. Free reasoning at scale and a visible thought process helped it reach No. 2 on web—about 10% of ChatGPT’s scale—with only 10 January days, and No. 14 on mobile with five days; Moore said five more days would have put it at No. 2 there. Day-30 retention was 7% versus ChatGPT’s 9%, though markets where the choice is “DeepSeek versus nothing” may skew the comparison.
ChatGPT’s growth reaccelerated as new models turned occasional experimentation into multiple recurring jobs. Web visits were essentially flat from February 2023 through February 2024, when data Moore had seen suggested students supplied 50%+ of traffic; visits have since doubled. ChatGPT’s own announcement reported weekly active users rising from 200 million to 400 million in six months—after the previous doubling took nine—as o1, GPT-4o, Advanced Voice Mode, Deep Research, Operator, and Canvas expanded utility.
Companion products are already significant, with multimodality still underexplored. Smith noted that three companion products were in the top 10, two NSFW-oriented, and that some users treat them as interactive fan fiction. Acharya said latent demand could grow substantially once products add richer multimodality and voice.
AI video is becoming a modular production stack rather than a winner-take-all model market. The ranking added Hailuo and Kling, both Chinese models, plus Sora. Google’s Veo 2 looked “next-level” in testing and was hoped for within three or six months. Different models excel at people, landscapes, anime, or hyperrealism, supporting aggregators such as Krea that reduce workflow seams and could consolidate 10 or 15 separate $20-a-month subscriptions.
Vibe coding is exposing enormous creation demand before it has produced its first wave of viral downstream apps. Cursor and Bolt reached the main ranking, Lovable reached the Brink list, and Moore cited Gary Tan as saying “95% of YC companies or something” now build with such tools. Lovable’s creation site draws significantly more traffic than hosted creations, validating “personal software” and “disposable software”—and foreshadowing an App Store that “is going to be chaos.”
Usage and monetization describe materially different consumer-AI markets. Only 40% of Sensor Tower’s mobile revenue and monthly-active-user rankings overlapped; tightly gated prosumer products can reportedly generate $50 million to $100 million of ARR from only one million or two million users. Paid acquisition can manufacture 10 million users at one-to-two-times payback, but the panel’s durable rule is product-first: without attention and retention, growth remains a “leaky bucket.”

Deep dive

1. Usage data says consumer GenAI is early and assumptions remain brittle

Moore grounds the ranking in behavior, not industry buzz: January 2025 Similarweb monthly visits determine the web top 50, while Sensor Tower monthly active users determine mobile. Both are then filtered for GenAI-first products; the separate mobile-revenue ranking captures what Sensor Tower can measure, typically subscriptions and in-app purchases, probably not advertising.
Churn coexists with persistence. Seventeen companies newly entered the web ranking, while 16 have appeared in every edition; the two five-product Brink lists caught former mainstays Runway, Otter, and Umax being edged out, alongside rising Krea and Lovable.
The adoption chronology began before ChatGPT, with Midjourney and Character.AI attracting niche communities in 2022. Snapchat’s My AI then reached roughly 150 million people; the Balenciaga Pope image, “BBL Drizzy,” and Coca-Cola’s Christmas ad with a lot generated by AI helped bring images, music, and creative AI into consumer or enterprise consciousness.
Smith notes that consumer activity typically trails research by “six to nine to 12 months.” Acharya places the market between infrastructure building and application building, with many categories still in the early-adopter phase. His signature inversion is deliberately hypothetical: AIs may handle some relationship-oriented interactions better because they have patience and “never have a bad day,” while humans may enjoy executing work that AIs organize.

2. AI video is becoming a modular production stack

Three video models entered: Hailuo and Kling, both Chinese models, and Sora. Google’s Veo 2 looked “next-level” in testing, and Moore hoped it would come out in the next three or six months. Acharya had expected more style transfer as a scalable video approach because it is more tractable and has lower inference cost; instead, researchers and product developers are pursuing raw text-to-video.
The panel attributes the Chinese models’ strength partly to training data that is “less copyright-sensitive”—prompting the aside, “a great euphemism”—which Moore connected to more realistic and prompt-adherent outputs. She also cited easier access to video-captioning labor and a larger pool of image-and-video researchers. Sora disappointed some users, whereas the Chinese systems exceeded expectations relative to the capital they had raised.
Models and interfaces are becoming opinionated: users can specify camera angles, shot width, and movement, while different systems specialize in people, landscapes, anime, or hyperrealism. Acharya sees Krea’s aggregation as “greater than the sum of their parts”: generate in Midjourney or FLUX, upscale, then use that image as a video’s beginning frame without crossing product seams.
Acharya singled out Ideogram for its distinctive text generation and aesthetic. Moore also described its image-to-text feature, which can turn a meme or copyrighted image into a text prompt for creating a new image.

3. Vibe coding monetizes latent creation demand before viral output

Cursor, an agentic IDE for technical users, and Bolt, a prompt-to-web-app product, reached the main list; Lovable reached the Brink list. The apparent mainstream breakthrough remains qualified: many users are still technical people prototyping quickly, exporting the code, and refining it themselves.
Acharya calls the emerging categories “DIY or personal software” and “disposable software.” A Bolt or Lovable project might be a compelling prototype, a tool for one person’s niche pain point, a 20-minute experience, or eventually a venture-scale company built by someone who never learned conventional coding.
Acharya initially assumed a few creators’ heavily trafficked sites explained the platforms’ rise. The data showed the opposite: Lovable.dev, where people create, receives significantly more visits than hosted Lovable.app creations. No major viral-app wave has arrived yet, so Moore expects awareness to rise further when it does: “The App Store is going to be chaos.”
Smith asks whether developers tailoring Deep Research to specific end uses is more likely than a horizontal tool; Acharya answers, “Not more likely, but I think it’s underexplored.” Market reports are the prescribed use, but tracing a meme’s origin revealed a “100x better” version of Know Your Meme; constrained applications could address Deep Research’s “blank page problem.”

4. Assistants are expanding through utility, differentiation, and companionship

ChatGPT’s web traffic stayed essentially flat from February 2023 through February 2024, when data Moore had seen suggested students contributed 50%+ of visits. It subsequently doubled, while its own reported weekly active users rose from 200 million to 400 million in six months after the previous doubling required nine.
Moore links the acceleration to capabilities that unlock distinct jobs: o1 reasoning, GPT-4o multimodality, Advanced Voice Mode, Deep Research, Operator, and Canvas. She moved from weekly-or-less usage to daily use across driving conversations, memo research, and brainstorming; reasoning especially made delicate work feel less risky than when ChatGPT could not reliably count the R’s in “strawberry.”
Companion products were another surprise: Smith noted three in the top 10, two of them NSFW-oriented, and said users were also treating them as interactive fan fiction. Acharya was surprised not to see more multimodality; he expects companionship demand to increase once products combine character with richer voice and other modalities.
Claude is not the conventional No. 2 with one-tenth the quality, Acharya argues; it is more beloved by fewer users, with stronger creative writing, a seemingly stronger personality, and “strangely much, much better” coding. That differentiation leaves room for ChatGPT, Claude, Gemini, and potentially other models to augment rather than simply replace one another.
Moore’s January cut captured only 10 days of DeepSeek: it reached No. 2 on web at roughly 10% of ChatGPT’s scale, then No. 14 on mobile with five days; she said five more days would have put it at No. 2 there. Its hook combined a reasoning model released free at scale—where o1 Pro had previously required ChatGPT’s premium subscription—with a captivating visible thought process. Day-30 retention was 7% versus ChatGPT’s 9%, though restricted markets may skew results because the alternative is “DeepSeek versus nothing.”

5. Revenue rewards concentrated utility more than raw reach

Sensor Tower’s revenue and usage lists overlapped by only 40%. The broad categories remained similar—photo and video generation or editing, beauty enhancement, and ChatGPT-like assistants—but the companies leading revenue were often different from those leading monthly users. Mobile use also favors phone-native behavior: avatar apps draw on stored selfies, voice-first products are natural to use on phones, and homework-helper apps are stronger on mobile than web.
Smaller audiences generated materially more revenue per user. Serious prosumer products such as Speak, Otter, and Captions can gate most utility behind subscriptions; Moore says some companies may reach $50 million to $100 million in ARR from one million or two million users, so they may not appear among the monthly-active-user leaders.
Mobile indie developers and international app studios can use App Store ads and other relatively low-cost acquisition channels to buy users. If they make back one or two times their acquisition spend, that can be attractive; they may reach 10 million users this way without producing as much revenue or ultimately as much profit as a smaller, higher-revenue product.
The optimal model depends on the product. A plant-identification app may never reach 100 million phones, yet committed plant or bird enthusiasts might pay $100 annually and use it daily or every other day. Moore’s closing rule: build toward the pain point or unique experience—even with an older model or one AI feature—because technical sophistication cannot rescue a “leaky bucket” product.

30. The Revival of Chinese Stocks and Alibaba’s AI-Stock Story: China Assets Enter a Long-Awaited Era of Confidence Ezra Klein and Derek Thompson: Politics, Trump, AOC, Elon & DOGE | Lex Fridman Podcast #462