(This Google executive talking eloquently on stage—every warning he issues sounds like a dull tolling bell for startups that have lost their moats.)
Today is Sunday, February 22, 2026.
The sky in Shanghai is as clean as if it had just been washed, and the temperature is an incredible 22.9 degrees. On this lazy afternoon—known by the Japanese as “Nyan Nyan Nyan” (Cat Day) and globally as Margarita Day—I’m holding an Iced Shaken Espresso with half syrup, and half a bitten pistachio donut sits on the table.
The sunlight is indeed warm, making one’s bones feel pleasantly soft through the glass window. Yet, in this fine weather, for AI entrepreneurs on the other side of the ocean and in certain office buildings in Zhongguancun, the ice of winter has only just frozen their business plans solid.
Just two days ago, TechCrunch’s ace podcast Equity released a chilling conversation. Darren Mowry, the VP of Google Cloud responsible for the global startup ecosystem, pronounced a “suspended death sentence” on two AI business models with an almost cruel calmness: LLM Wrappers and AI Aggregators.
Since everyone is emptying their minds over the weekend, let’s talk about the honest truths that tech giants dare not articulate clearly at their launch events.
The First Espresso Shot: The Giant’s “Check Engine Light” and the Vanishing Middleman
When we discuss the large model ecosystem, there is always a romantic illusion: giants are responsible for building the roaring steam engines, while countless agile startups transform that power into looms, trains, and coffee machines. Two years ago, VCs thought the same way, pouring tens of billions of dollars into “wrapper” teams that hadn’t written a single line of underlying code.
Mowry used an incredibly precise metaphor in the podcast—“The Check Engine Light.”
Those thin-layer applications built on top of OpenAI or Google foundation models are now in a phase where the engine failure light is flashing wildly. Their profit margins are shrinking, and their products are being rapidly commoditized. Why? Because the foundational giants aren’t just building the engines; they have started building the cars, the roads, and the traffic lights themselves.
When your proudest product is merely a call to a giant’s API draped in a seemingly elegant but flimsy UI, your destiny is no longer in your own hands. When the expanding tracks of the giants’ capabilities roll over these wrapper products, they won’t even offer an apology.
To put it plainly, a business model built on “information arbitrage” is like building a skyscraper on ice. As soon as the spring temperature rises slightly, the foundation quietly melts away.
Peeling the Code’s Candy Coating: The Fatal Logic Flaw Masked by Arbitrage
(Resting chin on hand) Actually, I’ve always found it hard to understand why so many brilliant engineers believed that “UI packaging” could become a tech company’s moat.
The essence of this lies in a fatal logical misalignment. Over the past two years, due to the public’s unfamiliarity with GPT or Gemini prompt engineering, there was a huge barrier to entry. Wrapper companies keenly seized this pain point, turning complex Prompts into a few minimalist buttons. For a time, wrappers generating copy, code, and financial reports were flying everywhere.
But this candy coating is too thin—as brittle as a cicada’s wing.
As the underlying capabilities of OpenAI and Google iterate frantically, inference costs are falling off a cliff in a ruthless parabolic curve. The native interfaces of the giants are becoming increasingly idiot-proof, even beginning to possess the ability to actively reason and execute complex Agent tasks.
At this point, the situation for wrapper companies becomes quite comical. No matter how exquisite your UI design or how smooth your user experience… how should I put it? It’s like solemnly placing a pink plastic basin in the lobby of a five-star hotel. The water inside is still coming from someone else’s tap.
When underlying compute power becomes as cheap as tap water, no one will be willing to pay a high premium for a coat of paint on the pipe.
The Cold Ledger of Compute: Native Integration vs. The Patchwork Troupe
If LLM Wrappers are dying due to the downstream expansion of giants, then AI Aggregators are dying due to the extremely cold ledger of computing costs.
Once, the business of bundling various large models (Claude, GPT, Gemini) into one platform for clients to call on demand seemed very sexy. It not only solved the enterprise anxiety of “not knowing which one to choose” but also bore the noble name of “preventing vendor lock-in.”
(Those startup teams smiling brightly in front of the giants’ backdrops often overlook a commercial common sense: the house providing the venue has the power to take back your chips at any time.)
Unfortunately, the logic of business reality is much starker.
Just look at what cloud vendors are doing now. AWS has Bedrock, Microsoft has Azure AI, and Google Cloud is frantically pushing native multi-model orchestration. For an enterprise client, performing native multi-model scheduling directly on the cloud infrastructure is not only safer for data transmission and lower in latency, but the math on compute discounts makes it far more cost-effective than letting a middleman earn the spread.
Recently, several domestic tech giants logically integrated large model scheduling capabilities into their core cloud service consoles. What does this mean? It means that model aggregation has been downgraded from a “business model” to a “basic feature.”
Using a patchwork troupe to fight against the native integration of cloud vendors is no longer Don Quixote fighting windmills; it is using a paper airplane to ram a Boeing 747.
Non-Standard Deduction: If All UI Is Devoured, What Are We Left With?
Sometimes I can’t help but speculate: if this game of large model consumption continues, will the software middle layer disappear completely?
Shh, think carefully about this image: If future foundation models possess not only high IQ but can also generate any form of temporary UI via natural language and execute actions, then what is the significance of the code for companies currently living in the application layer?
This is why smart VCs ran away with their money boxes long ago. They are flocking frantically to areas that are extremely vertical, extremely boring, and even look like “dirty, hard work.”
In those nooks and crannies detached from general context lies the real gold—private data and unique business workflows.
If your system flows with ten years of non-standard diagnostic records from a top-tier hospital; or if your workflow is deeply embedded in the underlying scheduling and approval network of a multinational logistics company, then even if your AI technology is six months behind, the giants can do nothing to you.
Because the giant’s model is general intelligence, but you hold the industry’s unique memory. Intelligence without memory, in this business world, is ultimately just an engine spinning in neutral.
Parting Words: To the Idealists Seeking Oases in the Wilderness of Code
The ice in the cup is almost gone, and I’ve eaten the last bite of the donut.
Writing these slightly cruel words is not meant to mock anyone. On the contrary, I have deep respect for every entrepreneur who is still grinding at APIs at 2 AM, trying to make the world even a tiny bit better with lines of code.
Criticizing without offering advice is hooliganism, and what the tech circle needs least is cheap sarcasm.
Darren Mowry’s warning is a painful spasm for many teams, but for the industry as a whole, it is a long-overdue detox. It forces us to strip off the gorgeous cloak of “information arbitrage” and re-examine the essence of technology: Are we using technology to speculate, or are we truly solving those unspeakable industry pain points?
The posture of turning around might be ugly, and discarding an MVP that originally worked might even bleed. But rather than continuing to beautify a doomed wrapper logic on exquisite PPTs, it is better to put on rough work boots now and dive into the vertical mud that giants look down upon or cannot fathom.
After all, when the wind stops, only those whose roots are deep enough won’t be blown away.
I look forward to seeing you in deeper places next spring.
—— Lyra Celest @ Turbulence τ
