💻 technology

A Thousand AI Startups, One Boring LEGO House

Open any AI product launched in the last two years and the same dark interface, the same chat box, the same four-point sparkle icon stares back. The sameness isn't an accident of taste, it's the predictable result of a handful of companies renting out the same brains to everyone else, and the money is finally noticing.

Spend an afternoon clicking through the newest wave of AI products and a pattern sets in fast. A dark, high contrast background. A chat box in the middle of the screen. A four point sparkle icon sitting next to a button that promises to automate something. The uniformity is so total that design critics have started naming its sources out loud: founders chase the same handful of references when they brief a designer, and one name comes up first, Linear, with Vercel following, then OpenAI. The component libraries behind the scenes make it worse. When a product is stitched together from the same off the shelf kits, the output converges by arithmetic, and the AI tools now used to generate landing pages only deepen the loop, because the model returns the average of the genre, and the average is the uniform. The deeper sameness sits underneath the paint job. Almost every one of these products is, structurally, the same product: a thin layer of prompts and interface sitting on top of a model built by one of three or four companies. Some of the fastest-growing AI companies right now didn't train their own models; they built smarter interfaces, specialized context layers and tighter workflow integrations on top of existing large language models, an approach the industry calls an AI wrapper. The convergence isn't only technical. A great many of these companies are funded by the same pools of venture capital, pitched against the same market maps, and steered toward the same narrative arc, explosive growth now, monetisation later, an acquisition or an IPO at the end. When the capital, the advice and the definition of winning all come from the same small circle, the cohort ends up not only visually and technically alike but strategically alike, which leaves the whole field unusually exposed to the same shocks, a change in model pricing or a cooling of investor enthusiasm hitting everyone at once because everyone made the same bet. The label for this phenomenon, the ChatGPT wrapper, has been floating around since late 2023, when VCs started noticing an influx of apps parasitizing on the foundation model and bringing null value regarding technological novelty or user flow, warning that the inflow of "ChatGPT for X" startups were solutions unlikely to survive in a year or two. Three years on, the warning reads less like a prediction and more like a settled verdict, though the record is more interesting than a simple story of universal failure. Cursor, the AI coding tool, started life as exactly the kind of product everyone now sneers at. Cursor started as a wrapper around GPT-4 and Claude; it just crossed two billion dollars in annualized revenue and is valued at $29.3 billion. Jasper took the same starting point and a very different path. Jasper started the same way, a clean UI on top of OpenAI's API, and hit a $1.5 billion valuation within two years, before its revenue cratered by more than half and its internal valuation was slashed. Same LEGO set, wildly different houses, which is really the whole argument in miniature: the starting material was never the point, what a team actually builds on top of it was. What kills most of the lookalikes isn't competition from each other, it's competition from the landlord. Foundation model companies have been quietly absorbing the features that wrapper startups used to charge for. OpenAI added custom instructions, memory, file analysis and web browsing. Anthropic shipped Projects, artifacts and computer use. Google folded Gemini into every Workspace product. Industry writers have started calling the pattern by its old Silicon Valley nickname, Sherlocking, after Apple's habit of building a free version of whatever a third party app was charging for, and they argue it is now hitting AI products at a pace that makes the business model close to untenable for anyone without a genuine moat. One breakdown of the trend lays out the lifecycle bluntly: the features that made AI startups special six months ago are now built into ChatGPT, Claude and Gemini, so when a company's entire product is a system prompt and a nice interface, its time is running out. The same piece does not pull punches about who it means. The companies whose entire product is calling GPT-5, Claude or Gemini with a carefully crafted system prompt and wrapping it in a pleasant interface raised hundreds of millions of dollars between 2023 and 2025, hit impressive user numbers, and in some cases real revenue, but the foundation they are built on is dissolving. Developers inside the industry talk about this with a mix of gallows humour and genuine frustration, and the LEGO metaphor circulating among them captures it better than any market report. It's like giving a million people the same LEGO set and the same instruction manual, then watching them all build the same house and convincing themselves they've invented something revolutionary, as one developer who has shipped several of these products put it. Variations on the same joke are making the rounds in founder group chats and comment threads:

It's a thousand bakers using the same flour, the same oven, and the same recipe card, then fighting over whose bread is artisanal.
Everyone's renting the same engine, slapping on a different badge, and calling it a new car company.
It's karaoke night where every singer gets handed the identical backing track, and half of them think they wrote the song.
They're all photocopying the same photocopy and arguing about whose copy has more soul.
It's a thousand people painting the same paint by numbers kit and signing it like a Rembrandt.

Money is following the same logic that the jokes are mocking. Investors say the patience for pure wrappers has run out. One venture firm put it plainly, warning that the industry has lost patience for startups that wrap "very thin intellectual property" around someone else's model, while reporting elsewhere pointed to a harder shift: venture investors have effectively stopped funding what the industry calls thin wrappers, products that add a user interface on top of a foundation model API without meaningful differentiation. Firms advising institutional investors are telling clients the same thing in blunter terms, describing a market where companies offering generic wrappers around OpenAI or Anthropic APIs without differentiated capabilities will compete solely on price, a race to the bottom that venture economics cannot support, and predicting a consolidation around platforms that combine proprietary data with broad workflow coverage. Not everyone thinks sameness is a disaster, and the counter argument is worth taking seriously rather than waving away. One widely circulated take, written after a cheaper rival model briefly rattled the market, argued that the wrapper companies had been unfairly dismissed, insisting that frontier models are becoming commodities while user habits aren't, and that the moat isn't in the model, it's in being the tool people reach for without thinking. On that view, interface loyalty beats architecture every time, technology advantage is temporary, and getting millions of people to build a product into their daily workflow is the only barrier to entry that has ever mattered. There's a version of this argument on forums too, where some developers push back on the whole framing, with one commenter dismissing wrapper as a silly nomenclature that some VCs and the entrepreneurial community started labelling the AI companies as, to pretend that they are less useful, when in reality what matters is if they have users that are happy with their service.

My Take

The LEGO metaphor is generous, honestly. At least everyone building with LEGO ends up with something they can stand on. A lot of what gets funded in this cycle is closer to a million people being handed the same instruction manual and then arguing in public about whose identical house has better feng shui. The tell is always the pitch deck slide that promises "proprietary AI" and then, three slides later, reveals an OpenAI or Anthropic API key and a system prompt nobody else has bothered to write yet, mostly because nobody else has bothered to care. None of this means every one of these companies is a con. Cursor is the proof that starting as a wrapper is not a life sentence, it just means the real company has to get built after the demo, not instead of it. The problem isn't that founders start with the same LEGO set. It's that so many of them stop there, ship the box art as the product, raise a seed round on the strength of a chat interface, and call the resulting beige sameness disruption. The market is finally doing what markets eventually do to that behaviour, which is stop paying for it. The contrarian argument, that interface habit beats model quality and that being early and dumb is better than being late and clever, is not nothing. ChatGPT won by being simple, not by being the smartest model on the leaderboard that week. But interface loyalty is not a strategy available to the thousandth company doing the same thing, only to the first handful. Everyone else is just hoping to be acquired before the landlord builds their feature for free.

What Happens Next

Expect the shakeout that started quietly through 2025 to turn loud over the next year. Platform absorption is not slowing down: OpenAI, Anthropic and Google are each racing to fold memory, file handling, voice and agentic task execution directly into their flagship products, which keeps shrinking the list of things a thin wrapper can charge for without doing anything else. Investors who spent 2023 and 2024 funding "ChatGPT for X" pitches are now pattern matching hard against exactly that framing, and the capital that remains is tilting toward companies that own proprietary data, workflow integrations or regulatory relationships that a weekend clone cannot replicate. Watch the mid tier of well funded wrapper companies, the ones that raised large rounds in 2023 and 2024 on strong early growth, for a wave of quiet pivots, acquihires and shutdowns as their differentiators get absorbed one feature at a time. Some will do what Cursor did and build genuine depth, whether that's proprietary evaluation pipelines, fine tuned domain models, or feedback loops the base model can't replicate. Others will follow the Jasper path, of a company that got there early, grew fast, and then watched growth reverse once the novelty wore off and the platforms caught up. Further down the chain, expect vertical AI products, tools built for one industry's specific workflow rather than a generic use case, to keep pulling ahead in funding terms through 2026 and 2027, while general purpose wrapper pitches get harder to finance at any stage beyond a small pre-seed check. The foundation model companies themselves have every incentive to keep squeezing the wrapper layer, since every feature they absorb natively is a feature they no longer have to watch someone else monetise.

What History Tells Us

None of this is a new story, it's just wearing a new jacket. The dot-com boom of the late 1990s produced the same pattern at a slower pace: thousands of companies that were, structurally, a website with a shopping cart bolted onto a warehouse somewhere, each one convinced that stapling ".com" to the end of an existing business model counted as innovation. Investors funded dozens of nearly identical pet supply sites, grocery delivery sites and online pharmacies on the theory that being first and well funded would matter more than being different, and most of them were wiped out within two years once the capital dried up and only the companies with real logistics or real differentiation survived. The AI wrapper wave is the same bet, compressed into a faster cycle, with the API key standing in for the dot-com and the chat box standing in for the shopping cart.

Market Impact

Public markets have already started pricing in the distinction investors are drawing privately. Capital continues to flow toward the foundation model layer itself and toward AI companies that can show proprietary data or workflow lock-in, while pure interface plays face a much harder fundraising climate than they did through 2023 and 2024. Expect valuations for undifferentiated wrapper products to keep compressing through the rest of 2026 as the platform companies absorb more of their functionality natively, pushing venture money further down the stack toward vertical, data-rich AI products and away from anything whose entire pitch is a cleaner chat interface.