LLMs Are Eating Entry Points While Apps Still Search for Their Place

This news is worth noting, but "lack of application vitality" doesn't mean nobody uses AI. The application layer hasn't stood firm yet. Large models are like water treatment plants; there are more and more outlets. Many AI applications just connect a pipe and sell a cup, so naturally, users feel they are dispensable.

The stronger the model, the more general-purpose entries resemble infrastructure. OpenAI started at the underlying model level but ended up becoming more like an operating system plus a super app. That observation group from 36Kr is quite harsh: truly few AI application companies have revenue exceeding $100 million. In MiniMax's semi-annual report, revenue was $117 million, up 283.1% YoY, surpassing the full-year 2025 figure of $79 million in half a year. However, losses were $358 million, still significant. Models eat the meat; the application layer might not even get a steady sip of soup.

Many teams building AI applications default to wrapping a model and adding a chat box. Launch events look good, but the next day the model upgrades, and the moat disappears. Zhou Hongyi said large models are like atomic bombs, and applications are like tea eggs. The metaphor is lively, but in business, you can't sell tea eggs just by setting up a random stall.

Vertical tools like Meitu Design Studio are worth watching. They embed AI into e-commerce design workflows that have budgets, frequency, and acceptance standards. Users pay for less overtime and fewer reworks. Models provide general capabilities; applications provide delivery certainty.

I've been trying RAG and search agents these past few days, and the feeling is direct. Making models stop hallucinating requires more than tough-talking prompts; retrieval, citations, boundaries, and reviews are what work. Applications are the same. Connecting to business data and making results traceable—these dirty jobs are the real barriers.

Another often overlooked problem with insufficient AI application vitality is that enterprise integration is troublesome. I previously wrote about cost-performance ledgers, meaning something similar. Model prices, quality lines, and review dates change together; you can't just look for the cheapest option. If the application layer only provides a pretty interface without helping teams solve quotas, permissions, costs, and logs, it will eventually get stuck.

Truly surviving applications must at least integrate into existing processes, not force users to open another window, manage data and permissions so sensitive content isn't exposed, and calculate costs and quality clearly so bosses know where the money goes.

This is why I've recently been looking at tools like viaSocket and Subpool. They are new to me, so I can't hype them as mature solutions. But if they can package authorization, retries, and quota allocation well, integration costs drop significantly. Integration cost is key. Whether a tool survives often depends on whether it can run smoothly on day one.

Large models keep charging forward, and application vitality seems lacking, sounding like bad news. I actually think this is the industry starting to split accounts. Model companies compete on parameters and inference, and also on entry points; application companies should compete on roles and workflows, and also on acceptance.

Some applications die because delivery is too light. Some launch for 180 days, raise tens of millions, accumulate hundreds of thousands of users, and finally shut down. After the noise, what remains are often small things embedded in specific work: an image editing workflow, a retrieval pipeline.

Surviving AI applications turn cleverness into chargeable daily routines.