Users, Revenue — and Still a Bad Business: Notes on LatePost’s AI App Deep Dive

This is a condensed summary of LatePost’s investigation 《有用户，有收入，AI 应用却不是好生意》(“Users, Revenue — and Still a Bad Business”) by Zhu Yingli, edited by Zhao Lei. All arguments and figures below are drawn from the original piece — this post is a reading note, not original reporting.

The thesis in one line

The classic software playbook — find a need, win users, generate revenue — is breaking down in the AI era. Hitting all three is no longer enough to sustain an independent company.

Because every independent AI app now faces three problems at once:

How soon will the model swallow this product? (Each model upgrade makes a batch of apps redundant.)
Can growth ever turn into profit? (Selling tokens means the faster you grow, the more you lose.)
Who owns the user entry point? (The labs selling you the model are coming for your users.)

Problem 1: Selling tokens means growth deepens the loss

The underlying business is reselling someone else’s tokens, so inference cost from upstream caps gross margin.

Knowledge-space product Kuse once made headlines for reaching $9M ARR in 60 days with no VC funding and no paid ads, later crossing $10M ARR. But founder Wu Xiankun is now deliberately suppressing growth: every new user handed a few dollars of trial credits drives inference bills to model and cloud vendors higher, and growth once pushed the company into heavy losses. Only after repeated pricing and packaging changes did gross margin turn barely positive.
ARR itself is losing credibility. Investors describe absurd annualization tricks: one-off payments counted as recurring, a single month’s annual fee multiplied by 12, or the best day or week of the year multiplied by 365 or 52. Even without inflation, renewal rates are far below classic SaaS — where roughly 95% of this month’s payers pay again next month.
Bessemer studied 20 fast-growing AI companies: many approached or hit $100M ARR within their first year of monetization, yet average gross margin was only about 25% — versus roughly 70% for mature cloud software.
Perplexity publicly cited gross margin around 60%, but only because roughly $33M of compute and network costs — mostly on free and trial users — was booked as R&D. Counted as cost of revenue, gross margin was negative.
Cursor’s quarterly gross margin for the period ending January 2026 was -23%; including free-user inference cost, about -31%.

Wu endorses Fireworks CEO Lin Qiao’s summary of the startup condition: “scaling to bankruptcy” — the more you grow, the closer you are to dying. Video app OiiOii stopped watching MAU/DAU and switched to paid conversion, renewal rate, and revenue per paying user. With Seedance 2.0 holding pricing power at ¥1/second, app-layer gross margin on paying users sits around 30% — and once free credits and marketing are counted, almost nobody in that market is margin-positive.

Problem 2: Every model upgrade makes a batch of apps redundant

The window for feature innovation keeps shrinking, and models close it faster each time.

Code search product Devv Search earned hundreds of thousands of dollars within six months of launch. Founder Zhang Jiayuan had mapped a three-to-five-year path: search first, then generation and debugging, then hands-free automation. Back-to-back Claude 3.5 Sonnet and Claude Code releases collapsed that roadmap — the search stage lasted only half a year before the market jumped straight to automated development.
An essay-writing product tells the story most starkly: in 2023, generating a paper cost one or two mao against a ~¥70 price, the company made ~¥10M with ~50% net margin, and receptionists once recognized the company name on an invoice because they’d used it for their thesis. By 2025, Kimi and DeepSeek wrote long-form content more cheaply or free, and revenue fell from tens of thousands of yuan a day to hundreds, then to zero.
Workflow products fared worse still: a team that raised seed funding on “AI workflows” in late 2025 found, a month later, that Skills had become standard in agent products. On the next raise, investors had already concluded “Skills will replace Workflows.”
Across a16z’s recurring consumer AI app rankings, only 14 of the top 50 companies appeared on every list between 2023 and 2025. Per Epoch AI, frontier model capability improvement roughly doubled after April 2024, from about 8 index points per year to about 15.

For these companies, the model doesn’t swallow the product overnight — all three pivots above happened while the original product still had revenue. What founders lose isn’t today’s users; it’s the future that might have been.

Problem 3: The company selling you the model is coming for your users

Whether “the model is the app” is the question hanging over every independent product.

When Manus went viral in 2025, founder Xiao Hong’s line “the shell has value” gave independent apps confidence. The rapid rise of Claude Code and Codex through 2026 has pushed the verdict back the other way.
The Brookings Institution argues that as model capabilities converge, selling API access alone can’t justify the enormous training and operating investment — shipping applications becomes a way for model companies to recoup that investment.
Anthropic has repeatedly run the playbook: once Cursor validated demand for AI coding, Anthropic shipped Claude Code and gradually became developers’ preferred choice; while deepening its Figma partnership, an Anthropic executive left Figma’s board and the company launched Claude Design, entering Figma’s core market. OpenAI follows the same pattern with Deep Research, Codex, and ChatGPT Sites.
In China it’s more direct. Tencent, Alibaba, and ByteDance all entered office agents with WorkBuddy, Qwen Office, and Doubao Office. Per Analysys, traffic to the office-agent category doubled from March to June, but two-thirds of that went to three big-company products. Of dozens of projects an investor saw early in the year, only single digits survived — some founders ended up reselling for WorkBuddy as agents.
Startups simply can’t match free-tier giveaways: WorkBuddy’s free monthly credits are worth roughly ¥25; at one million DAU that’s ¥25M a month just for free users, and each million invited new users can cost up to ¥100M in credits.

Escaping competition can mean escaping users

Many teams start from “what won’t the model companies do” — and fall into a different trap.

Wu picked the canvas form factor in 2024 precisely because the labs’ main entry point was chat. Lovart, Flowith, Miro, Subset, and even Baidu’s “Free Canvas” converged on similar thinking.
But designing to dodge the models pushes products toward more complex interactions and niche, novel-looking needs — every extra layer of differentiation shrinks the set of people who can understand and use the product. Kuse spent a long stretch perfecting PDF reading on canvas before Wu conceded a simpler truth: maybe not many people want to read PDFs on a canvas at all.
Wu attributes the miss to the “intention” behind the choice: starting from avoidance rather than from unleashing model capability narrows every path ahead. Devv 2.0 hit the same lesson — designing features against competitors produces differentiation that may not be what users want.
Yet meeting the competition head-on can also work: one founder saw Claude Code move into design and shipped an open-source alternative, pulling in nearly 50k stars in a week and revenue exceeding his previous workflow product.

Two escape routes: sell outcomes, or build the model

Founders still searching generally pick one of two directions.

Downstream — sell results, not products. OiiOii stopped fighting subsidy wars and moved into specific verticals, post-training small models to solve concrete problems and delivering finished video and workflows. In comic-to-animation, general video models overfit to realism or 3D and make 2D look unnaturally smooth, which reads as fake for Japanese anime; their trained approach reproduces authentic anime motion, and clients pay. The comparison is stark: two minutes produced with a generic tool costs about ¥200, while two minutes that actually meets the brief can sell for ¥10,000. The essay-writing team turned instead to its accumulated 5M+ student users, offering AI training and “one-person company” coaching. Wu went deepest: partnering with investors to acquire services companies, rebuilding their workflows with AI, and sharing profit through equity — one acquisition already closed.

Upstream — build your own model. Cursor, Krea, Captions, Perplexity, HeyGen, and LiblibAI have all moved from apps into the model layer. Yang Bolin’s team went from Cuflow (a courseware-to-video product, discontinued April 2026) to training a code-first real-time interactive video model. His description of the app-layer condition: “we’re only ever tenants.” Every upstream update forces a re-test of capabilities and costs, and some updates erase months of engineering outright — early long-document pipelines (OCR, chunking, separate image/text retrieval) became not just redundant but slower than simply feeding the whole document once context windows grew. Going back to the model layer felt like a higher-odds bet: 25 people building the post-training pipeline from SFT through preference optimization to RL, needing hundreds of GPUs, funded by a new round.

The one takeaway

None of these paths is proven yet. But the report lands the question every app-layer builder has to answer: when capability, cost, and the entry point all sit upstream, what does an independent application company actually get to keep? Either go deep enough to own outcomes and operations, or push scenario understanding back into a model of your own. That is a warning worth re-reading, for anyone building on top of foundation models.