Over the past period, an anomalous industry phenomenon has increasingly drawn the attention of analysts: the performance curve of large language models has continued to climb at a nearly steep slope, with breakthroughs that redefine cognition (cognitive understanding) nearly every month, ranging from multi-step reasoning to long-text processing, from code generation to multimodal comprehension.

However, in stark contrast, the AI application ecosystem for end-users and vertical scenarios has shown a rare contraction.

According to the 'AI Startup Value Creation White Paper (2025)' released by the Oxford University research team, the total annual global AI investment has exceeded $400 billion, yet only about 33% of companies have successfully scaled their AI projects from pilot to widespread application. This means that for most B2B AI startups, while it is easy to secure proof-of-concept contracts, converting them into long-term, high-value renewal revenue is difficult.

More intriguingly, the number of standalone applications in app stores labeled 'AI-driven' continues to grow, yet the list of top products with monthly active users exceeding one million keeps shrinking. A significant number of applications fall into user churn and revenue stagnation within three months of launch. This divergent phenomenon of 'rising waters upstream and retreating boats downstream' is forcing the entire industry to re-examine a fundamental question: As models become increasingly versatile, what irreplaceable value do the applications built around them truly offer?

01. Application Scenarios: Vertical Territories Eroded One by One by General-Purpose Capabilities

During the early days of ChatGPT's emergence, the market saw a surge of AI applications built on the logic of 'filling model gaps.'

For instance, in the legal tech sector, startups gained paying clients among law firms by integrating professional regulatory databases and constructing custom retrieval-augmented generation (RAG) pipelines, effectively reducing the model's hallucination rate in legal citations. In medical consulting, some apps enhanced the reliability of AI diagnoses to near-primary-care physician levels by cross-verifying symptom libraries with model inferences. Marketing copywriting tools relied on carefully designed prompt chains and multi-round dialogue templates to align generated content more closely with specific brand tones.

These applications all seized upon the early weaknesses of large models, erecting temporary support structures around the models' vulnerabilities through engineering means. However, the release of GPT-4 virtually eliminated the accuracy gap in legal citations overnight. Claude's optimization for complex reasoning tasks reduced medical diagnosis error rates to levels comparable to vertical applications. After the expansion of context windows from thousands to millions of tokens, copywriting tools that once relied on segmented processing and external memory mechanisms found their meticulously maintained long-term memory banks rendered obsolete. This rapid collapse of scenario moats was not due to technical regression by vertical teams but because the models' general-purpose capabilities, advancing in unpredictable leaps, proactively covered territories once considered 'professional barriers.'

Even more concerning is the homogenization and contraction at the interface level. Every successful application in the mobile internet era boasted a unique interaction language—Tinder's swipe-to-choose, Shazam's audio fingerprinting, Instagram's double-tap-to-like—where interaction design itself was part of scenario definition, and users' mental models shifted when switching between apps.

Today's AI-native applications, whether for writing, painting, programming, or data analysis, have nearly all devolved into uniform dialog boxes with input fields, occasionally supplemented with a few shortcut buttons. When users interact with all AI products in the same way, the cost of switching between products approaches zero, and brand loyalty becomes elusive. Industry surveys show that over 70% of AI application users have switched their primary AI tools at least three times in the past year, mostly because 'they heard the new model works better' rather than 'the new app solves what the old one couldn't.' This dissolution of scenario identity means application developers can no longer retain users through 'unique usage experiences' but must passively follow the arms race in model capabilities—a War of Consumption (war of attrition) that small and medium-sized teams cannot sustain.

02. Business Models: Pricing Power Lost in the Cracks and Eroding Customer Stickiness

The erosion of scenario advantages has directly punctured the commercial foundations of the AI application layer. Traditional SaaS pricing logic is built on functional scarcity and switching costs.

AI applications, however, are vastly different, as their core capabilities rely entirely on calls to upstream large model APIs, with pricing power firmly held by a few foundational model companies.

Application layer entrepreneurs essentially act as 'intelligence wholesalers': they purchase general intelligence from OpenAI or Anthropic at several dollars per million tokens, then repackage it through prompt engineering, output formatting, and minor domain knowledge injection to retail it to end-users via monthly subscriptions.

The fragility of this business model lies in the Continuously declining (continual decline) of upstream wholesale prices—mainstream API call costs have dropped over the past year—while the retail end faces intense user incentive to 'bypass middlemen.' When users discover that direct subscriptions to large model vendors are cheaper, offer many open-source features, and even provide early access to the latest model versions, the paid conversion rate for intermediaries plummets.

Even more alarming to capital markets is the AI application layer's consistent failure to establish the 'data lock-in' stickiness effect seen in traditional software. In the past, when using a CRM system, enterprise clients' accumulated customer profiles, sales processes, and automation rules were stored internally, making migration extremely costly.

However, historical articles in AI writing tools, style reference images in AI painting tools, and code snippets in AI programming assistants can all be exported losslessly or retrieved directly from conversation history. The relationship between users and specific AI applications remains a shallow 'temporary invocation' rather than a deep 'strategic partnership.'

03. Future Forms: Transformation from Explicit Tools to Invisible Infrastructure

However, declaring the demise of AI applications is clearly an oversimplification. Reviewing computer industry history, operating systems did not eliminate application software but instead spawned a software ecosystem larger than ever before—except that successful applications no longer did what operating systems excelled at but instead targeted vertical scenarios unimagined by OS designers.

Today's AI models are playing a similar platform role, continuously absorbing general-purpose, template-able intelligent tasks while pushing application developers' creativity toward more niche, context-dependent, and physically interactive corners. The AI products that remain viable all avoid direct competition with general models in 'intellectual output' and instead establish new competitive dimensions in 'execution closure' and 'experience differentiation.'

In a deeper sense, AI applications are undergoing a transformation from 'explicit tools' to 'invisible infrastructure.' The most promising next-generation applications will no longer be standalone apps with AI labels, chat windows with robot avatars, or disclaimers stating 'generated by artificial intelligence.' They will become increasingly silent and background-embedded, operating continuously at levels imperceptible to users, much like electrical power systems.

For example, next-gen office suites won't require you to first open an AI panel and type 'summarize this document for me'; instead, summaries will automatically appear in the sidebar the moment you select text. Future ERP systems won't ask financial personnel to converse with large models to generate reports but will inject model reasoning directly into every node of the data flow, pushing explanations and action recommendations the moment anomalies arise.

04. Conclusion: Future AI Applications May Head in Two Directions

Thus, from a commercial logic perspective, surviving AI applications will evolve into two distinct directions.

The first is 'deep-diving' types, entering professional scenarios difficult for general models to penetrate, such as equipment failure prediction in industrial manufacturing, molecular dynamics simulation in drug discovery, and real-time risk control in high-frequency financial trading. These domains require not just natural language understanding but also deep coupling with proprietary hardware, industry databases, and physical simulation engines to construct composite barriers that model capabilities cannot easily infiltrate.

The second is 'thin-surface' types, which overlay an extremely thin but highly intelligent interaction layer atop general models, making users feel not 'I'm talking to AI' but 'I'm just getting work done, and everything suddenly feels smoother.'

The commonality of these two application types is that they both abandon the linear thinking of 'earning money by packaging model APIs' and instead treat models as a continuously depreciating public resource, betting true value on industry insights and interaction design beyond model capabilities. For entrepreneurs, this undoubtedly means higher barriers and longer return cycles, but it is precisely the Cruel Charm (ruthless charm) of this round of shakeout—it weeds out speculators who rush to market with just a few prompts, leaving long-termists who truly understand scenarios, respect workflows, and know how to build solid bridges between intelligence and reality.

- End -