--- title: "The Clear-Eyed Allen Zhu: An AI App Boom Is Imminent—This Year Likely Already Saw the Birth of the Next “Rednote”!" author: "JasonH1121" sourcePublication: "Unique Research Substack" originalPublishedAt: "2025-09-13T08:15:03+00:00" canonical: "https://ffcap.cn/en/research/the-clear-eyed-zhu-xiaohu-an-ai-app" source: "https://uniqueresearch.substack.com/p/the-clear-eyed-zhu-xiaohu-an-ai-app" language: "en" --- # The Clear-Eyed Allen Zhu: An AI App Boom Is Imminent—This Year Likely Already Saw the Birth of the Next “Rednote”! When it comes to the most interesting observers of today’s AI industry, Allen Zhu is unquestionably near the top of the list. The Managing Partner of GSR Ventures isn’t one for platitudes—his assessments are blunt, even a bit “cold-water”—and precisely because of that, they’re often the clearest wake-up calls for the industry. At the 2025 Inclusion Bund Conference, he once again laid out his thinking: the capability boundaries of AI are already visible, and the real opportunities lie **beyond** the models themselves. “Next year applications will absolutely explode; which means the next ByteDance, the next Kuaishou, the next Rednote was probably founded **this year**.” [![图片](https://substackcdn.com/image/fetch/$s_!908c!,w_1456,c_limit,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd6e22114-5a24-4c29-be95-2c3200ba1f37_1080x720.png)](https://substackcdn.com/image/fetch/$s_!908c!,f_auto,q_auto:good,fl_progressive:steep/https%3A%2F%2Fsubstack-post-media.s3.amazonaws.com%2Fpublic%2Fimages%2Fd6e22114-5a24-4c29-be95-2c3200ba1f37_1080x720.png) Allen Zhu’s views start from a cool reading of where we are. He believes the upper bound of current AI capability is largely in sight—paradoxically, that creates **more** opportunity for entrepreneurs. “GPT-5 finally arrived after endless anticipation, but to be honest, people were disappointed,” Allen Zhu says bluntly. Under the Transformer paradigm, the ceiling for AGI-like capability is already discernible; further gains in “core intelligence” are small. Improvements now skew toward **user experience** and **cost**. The bottlenecks are data and reasoning; blindly scaling parameters and data won’t boost intelligence and may even hurt performance. That may sound pessimistic, but Allen Zhu argues it’s good news for startups. As the pace of model-level breakthroughs slows, the risk of a startup being wiped out overnight by a new model diminishes—opening the floodgates at the **application layer**. Over the next 2–3 years, Allen Zhu expects **model miniaturization** to be a major trend: pruning and refining data, lowering usage costs, and improving UX—a direction that is more feasible and cost-effective, especially for Chinese founders. Across China and the U.S., token consumption is now exploding. The race has shifted from training ever-bigger models to **consuming tokens via real applications**. Last year’s breakout apps were text-first—meeting notes of every kind. In the U.S., Bridge (for doctor-patient conversation notes) did well; in China, DingTalk-based meeting notes built by local founders thrived. These “unsexy but useful” tools commercialize easily and aren’t technically exotic. This year, **voice** apps are clearly breaking out. Voice models are now so good that most users can’t tell AI from a human. Allen Zhu notes: if a user detects an AI call, ~80% hang up; if they **can’t** tell, the hang-up rate drops to ~25%—a huge gap. He even predicts that late this year or next, **video-based** AI apps will boom. As latency for generated video/voice approaches ~1 second—nearly negligible—these products will have massive potential and could upend how content is made. From UNIQUE RESEARCH’s long-running tracking of thousands of AI apps worldwide, we see the same pattern: the marginal “model dividend” is flattening. Winners are teams that **embed** AI into specific scenarios and relentlessly tune **experience and cost structures**. Allen Zhu’s core refrain: **long-term moats for AI apps are not technical**. “All AI applications are wrappers,” he says. They call base-model capabilities to do some job; that alone can’t be a moat. Model features iterate fast and are easy to imitate. Building a defensible moat on AI technology **itself** is nearly impossible. If AI tech can’t be the moat, look **outside** AI. - **Deep workflow + editing capability.** Even with strong generation, commercial delivery hinges on the last 5–10% of polish. Robust editing and post-processing matter. (Allen Zhu cites an AI creator community he invested in that focuses on advanced editing to perfect images.) - **Complex workflows or industry data.** Tackle real pain points—efficiency, cost, UX—in verticals where processes and compliance are hard to copy (e.g., medical scribing that auto-creates notes and files them into EMRs). - **Specialized hardware fused with scenarios.** Allen Zhu likes products that combine hardware + AI: - **Meeting-minutes hardware:** a thin card that denoises and captures audio better. - **AI name badges:** for sales staff—surfacing talking points, flagging mis-selling, and filing daily reports automatically. - **Why “dirty work” wins.** In the mobile internet era, only a handful of giants (Uber, DoorDash, Airbnb) emerged—and all did heavy offline lifting that big tech didn’t want to do. AI will be similar. To deliver **results** in the real world, agents need to do on-the-ground work. That’s startup territory. “A big model will eat 90% of agents,” Allen Allen Zhu warns. Many agent startups today look like early web “site admins”—toolish, opportunistic, lacking moats. Advanced models (e.g., DeepSeek) will absorb or obsolete most generic agents. His advice: **go vertical**. Build sticky agents in specific contexts to avoid being steamrolled by model giants like Google. For Allen Zhu, commercialization is the core metric. He’s wary of hype and insists founders and investors return to business fundamentals. From PC to mobile to AI, **retention** is the only metric that truly matters. Many AI companies tout “vibe revenue”: one-off spikes with no repeatability. Some pitch ARR by annualizing a single day’s number. Users pay in month one, feel “meh” in month two, and churn. In mobile, reacquisition can cost 10x—often impossible. Retention proves whether there’s a real future. Meeting notes, voice AI, contact-center conversation agents, on-call sales agents—even toys with dialog agents—these are practical, revenue-friendly scenarios. Agent app startups are cheaper to build, but competition is brutal. In Silicon Valley, many VCs now wait until a product ships and hits **$2M ARR** before investing. Allen Zhu cautions Chinese founders: if you can’t reach **$5M ARR within 12 months**, few will pay attention. Speed to paid retention and unit economics is critical. Despite the buzz, Allen Zhu is cautious about embodied AI’s unclear commercialization path; GSR has exited many such bets. He differentiates that from **purpose-built AI hardware**, which he still supports. U.S. and China are the two most competitive AI arenas. Chinese founders have unique strengths and a distinct path. In the U.S., 6 of the 10 fastest-growing AI firms are To-B with mostly non-Chinese founders. Chinese entrepreneurs, by contrast, excel at **consumer apps** and differentiated **UX/playbooks** (e.g., gamification). In smart hardware, China’s supply-chain cost/efficiency is a major edge. The Pearl River Delta’s dense ecosystem lets you solve production problems on the spot. Allen Zhu urges Chinese teams to **go global**—Japan, Southeast Asia, the Middle East—where competition is lighter and Chinese teams can execute. Steer clear of head-on model wars with giants. When everyone taps the same model capabilities, where does differentiation come from? **Outside AI.** The enduring needs of people haven’t changed in 30 years and won’t in the next 30; AI simply enables better experiences and product forms. Allen Zhu’s message is a warning and a guide: - **Embrace miniaturization & multimodality.** As model capability plateaus, **cost and power** decide winners. Voice and video are next. Focus on on-device deployment, latency, and data slimming. - **Go deep in verticals; build non-tech moats.** Healthcare, education, e-commerce, manufacturing—complex processes + rich data = durable moats. Workflow + editing tools beat “re-skins.” - **Fuse hardware to differentiate experience.** AI name badges, meeting-note cards—hardware is a key delivery vehicle and a moat. As AI moves edge-ward, **software + hardware** becomes a trend. - **Be sober about the race; close the loop.** Beware flashy tech with no retention. Early retention and paid renewal are what matter. As Allen Zhu said at the Inclusion Bund Conference: “Next year applications will absolutely explode; the next ByteDance, the next Kuaishou, the next Rednote was likely founded this year.” His encouragement to founders: have the courage to set sail for that “sea of stars.” So rather than chasing ever-more generalized model power, pour your energy into **deeply fusing AI with concrete industry scenarios** and solving real problems to create business value. Building moats in **data, operations, product, and user insight** will be the key for AI app companies to stand out in fierce competition. May these insights spark something for you—and may you find your own patch of that vast, starry sea. --- Original publication: https://uniqueresearch.substack.com/p/the-clear-eyed-zhu-xiaohu-an-ai-app On-site reading page: https://ffcap.cn/en/research/the-clear-eyed-zhu-xiaohu-an-ai-app