February 12, 2026
Bloomberg: $35/month. Financial Times: $42/month. The Economist: $17/month. Original analysis by Tatsu with 40+ footnotes: $8/month.
Share this preview with others.
The AI bubble isn't popping. It's consolidating.
This is the final installment of our series on the AI Dollar, and it addresses the question every investor, founder, and employee in the AI space is asking: who survives? We've established that the United States controls the chokepoints of the AI economy, that China can't catch up, that Taiwan can't be taken, and that the Trump administration is (intentionally or not) rebuilding dollar hegemony on silicon foundations. Now the question becomes: where does the value accrue?
The answer is uncomfortable for most of the industry. The technology wins. The arbitrageurs lose. And the dollar wins either way.
Full investigation below. $8/month for novel, footnoted deep analysis.
The Wrapper Business Model
Let me explain what a "wrapper" is and why most of them are doomed.
A wrapper company takes an underlying AI model (typically from OpenAI, Anthropic, or Google) and builds a product on top of it. The wrapper adds a user interface, some prompt engineering, maybe integration with specific workflows. The customer pays the wrapper company. The wrapper company pays the model provider for API calls.
The economics work like this: charge the customer $20 per month. Pay OpenAI $2 in API costs. Keep $18 as gross margin. Build a business.
This arbitrage works until it doesn't.
The freemium trap. Most wrapper companies offer free tiers to acquire users. Free users consume API calls. The wrapper company pays OpenAI whether the user pays or not. If conversion from free to paid is slow, if users churn before converting, if the free tier is too generous, the company burns cash faster than it generates revenue.
The upstream threat. The model providers are watching what wrappers build. When a category proves valuable, nothing stops OpenAI from building that feature directly into ChatGPT. What took a startup twelve months and $10 million in venture capital to develop, OpenAI can ship in a sprint. The wrapper's differentiation evaporates overnight.
This pattern has a name: Sherlocking.
The Sherlocking Pattern
Apple perfected this move. For decades, they've watched third-party developers build successful utilities, then integrated equivalent functionality into the operating system, killing the original product.
Konfabulator created desktop widgets in the early 2000s. Apple shipped Dashboard with Mac OS X Tiger. Konfabulator died. Watson provided fast file searching. Apple shipped Spotlight. Watson died. Countless screenshot utilities existed. Apple built screenshot tools into macOS. The utilities died.
Apple executives have argued this isn't predatory; it's progress. Users benefit from integrated functionality. The third-party developers, well, they should have known the risk.
The frontier AI labs are doing the same thing, just faster.
Jasper and Copy.ai built businesses around AI-generated marketing copy. ChatGPT now does this natively. AI slide makers proliferated in 2023. Claude Artifacts now generates presentations within the chat interface. AI email writers were a category. Gmail's Gemini integration handles it. Meeting summarizers charged subscription fees. Zoom and Teams added native AI transcription and summary.
Every wrapper category that proves valuable gets absorbed. The question isn't whether absorption will happen but how long before it does.
Confirmed Casualties
The massacre is already underway. Let me document the bodies.
Builder.ai raised capital at a $1.2 billion valuation, claiming AI could automate software development. In 2025, the company collapsed into bankruptcy. Investigations revealed that the "AI" had been overstated; human developers were doing much of the work. Revenue figures had been inflated. The CEO was ousted. A billion-dollar valuation evaporated.
Humane raised $241 million for the AI Pin, a screenless wearable that was supposed to replace your smartphone. The product launched to brutal reviews. The hardware was slow, the AI assistant unreliable, the use case unclear. By mid-2025, Humane sold to HP for $116 million, less than half of what investors had put in. The founders cashed out. The employees got nothing.
EyeEm, the AI-powered image sharing platform, shut down in January 2026. It couldn't compete with the image generation capabilities being integrated into every major platform. The AI that was supposed to differentiate EyeEm became table stakes that every competitor offered.
These are the confirmed deaths. The walking wounded are more numerous.
The Walking Wounded
Jasper was valued at $1.5 billion in 2022, one of the first AI unicorns of the generative wave. By 2025, internal valuations had dropped to approximately $1.2 billion. Revenue was stuck around $88 million annually. Growth had stalled. The problem: ChatGPT does most of what Jasper does, for $20 per month or free. Why pay Jasper $40-80 per month for a wrapper when you can go direct?
Jasper hasn't failed. It's worse than that. It's stalled. The company has enough revenue to survive but not enough growth to justify its valuation. Employees hold options underwater. Investors are stuck. The optimal outcome is a modest acquisition, not the IPO that the 2022 valuation implied.
Copy.ai faces similar dynamics. The product works. Users exist. But switching costs are zero. Any customer can recreate their Copy.ai workflows in ChatGPT or Claude in an afternoon. There's no moat. There's no lock-in. There's just a UI that's slightly more convenient than the alternatives, and convenience isn't defensible when the alternatives are free.
The pattern repeats across hundreds of startups that raised money in 2021-2023 on the thesis that "AI wrapper for [category]" was a venture-scale opportunity. Most of these companies still exist. Most of them are dying slowly.
The Funding Cliff
The capital markets have noticed.
Venture capital funding rounds dropped 42% in 2024 compared to the 2021-2022 peak. The money that is being deployed is concentrating at the top: 60-70% of all AI venture capital is now going to megarounds of $100 million or more. The winners are raising at ever-higher valuations. Everyone else is struggling to raise at all.
The math is brutal. Most AI startups funded in 2021-2023 raised 18-36 months of runway. That runway is expiring in late 2025 and early 2026. They need to either demonstrate path to profitability (unlikely for most), raise another round (increasingly difficult), or sell (to whom, at what price?).
This is the cliff. Hundreds of companies approaching the edge simultaneously, with too few acquirers and too little new capital to save them all.
Who Survives
Not everyone dies. Some companies are building real businesses. The question is what separates survivors from casualties.
Infrastructure players win. OpenAI's valuation has reached $500 billion. Anthropic is at $183 billion. These aren't wrapper companies; they're the platforms that wrappers depend on. Every API call from every wrapper flows through their systems. They capture the margin that wrappers can't defend.
Cursor, the AI-powered code editor, went from $2.6 billion to $29.3 billion valuation in a single year. Cursor isn't wrapping someone else's model; it's building a deeply integrated developer tool that gets better with proprietary data from every codebase it touches. The moat is the accumulated understanding of how developers actually work.
Proprietary data creates moats. Bloomberg's AI initiatives work because Bloomberg has decades of proprietary financial data that no one else can access. A startup can't replicate Bloomberg's information advantage by wrapping GPT-4. The data is the moat.
Workflow lock-in creates moats. Figma's AI features work because Figma already knows your design system, your components, your brand guidelines. The AI isn't the product; the accumulated context is. Switching away from Figma means abandoning that context. The switching cost is the moat.
Vertical expertise creates moats. Legal AI companies that have trained on case law, built compliance features, and established relationships with law firms have defensible positions. The expertise isn't in the model; it's in the domain knowledge wrapped around the model. Generic AI can't easily replicate deep vertical integration.
Enterprise relationships create moats. Salesforce's AI isn't better than alternatives. It's installed. Enterprises have spent years integrating Salesforce into their workflows. Ripping it out to use a superior AI tool is a multi-year project that no CIO wants to undertake. Being worse but installed beats being better but requiring migration.
What Dies
The inverse is equally clear.
"GPT wrapper with a logo" dies. If your entire product can be described as "we call the OpenAI API and add a nicer interface," you have no moat. The interface isn't defensible. The prompts aren't defensible. The logo isn't defensible. OpenAI can ship your product as a feature.
Single-feature tools die. "AI that does one thing" is a feature, not a product. Features get absorbed into platforms. If your pitch is "we're the AI [narrow task] tool," you're describing a feature that will be integrated into the tools people already use.
UI-only differentiation dies. A better user interface is not a moat. User interfaces can be copied in weeks. If the only thing separating you from ChatGPT is design, you're a design project away from irrelevance.
Slow freemium conversion dies. If your free users don't convert to paid users faster than your burn rate, you die. The freemium model works when free users have clear upgrade paths and when paid tiers offer genuine value over free. Most AI wrappers can't articulate why anyone should pay when free ChatGPT exists.
The Dot-Com Parallel
History offers guidance here.
In 1999, the conventional wisdom was that the internet would transform everything. The conventional wisdom was correct. In 2000, the dot-com bubble popped and hundreds of internet companies failed. This seemed to invalidate the conventional wisdom. It didn't.
The internet transformed everything. The business models that failed in 2000 weren't wrong about the technology; they were wrong about the economics. Pets.com understood that people would buy things online. It didn't understand that commodity e-commerce couldn't support the margins required to justify its cost structure. Webvan understood that grocery delivery would exist. It didn't understand the unit economics of last-mile logistics in 2000.
Amazon survived. Google survived. The infrastructure players and the companies with genuine moats survived. The arbitrageurs and the feature-dressed-as-companies died.
AI follows the same pattern. The technology is real. The transformation is real. The companies that mistake proximity to the technology for possession of a business model will die. The companies that build genuine moats will survive.
The difference between 2000 and 2026 is speed. The dot-com shakeout took years. The AI shakeout is happening in months. The frontier labs move too fast, the Sherlocking happens too quickly, for marginal players to survive on hope.
Investment Implications
For investors, the lesson is straightforward.
Nvidia wins. Picks and shovels always win in gold rushes. Everyone needs GPUs. Nvidia supplies them. Whether AI startups succeed or fail, they buy Nvidia chips while trying. The concentration of AI compute in Nvidia hardware creates a toll booth on the entire industry. Every dollar spent on AI ultimately flows through Nvidia's revenue line.
Cloud providers win. AWS, Azure, and Google Cloud host the training runs and serve the inference. Whether the AI application succeeds or fails, it runs on cloud infrastructure. The hyperscalers are the landlords of the AI economy. Landlords collect rent regardless of tenant success.
Frontier labs probably win. OpenAI, Anthropic, and Google are building the models that everyone else depends on. They capture the API revenue from every wrapper. They absorb successful categories into their products. They have the capital and talent to maintain their positions. The "probably" is because there's genuine competition between them, and it's unclear which one or two will dominate.
Avoid wrappers without moats. Any company whose business model depends on margin between what they charge customers and what they pay OpenAI is vulnerable to compression from both directions. OpenAI can raise API prices. Customers can go direct. The wrapper sits in a vice.
The exception is wrappers with genuine moats (proprietary data, workflow lock-in, vertical expertise, enterprise relationships). These can survive and thrive. But they're rare. Most of what raised venture capital in 2021-2023 doesn't qualify.
What This Means for the AI Dollar
Here's how this connects to the broader thesis of this series.
The AI economy is consolidating into a small number of dominant players, almost all American. Nvidia, the hyperscalers, the frontier labs. The value is concentrating at the infrastructure layer, not the application layer. The wrapper massacre isn't destroying value; it's transferring value from thousands of startups to a handful of platforms.
This concentration reinforces the AI Dollar thesis. If AI value accrues primarily to American infrastructure companies billing in American dollars, then the global AI economy is denominated in dollars by default. Every company in the world that wants to use frontier AI pays Nvidia (dollars), pays AWS/Azure/GCP (dollars), pays OpenAI/Anthropic (dollars). The wrappers that survive will still depend on this infrastructure.
The dot-com bust didn't undermine American technological dominance. It concentrated it. The AI shakeout will do the same. Fewer companies, more American, more dollar-denominated.
Conclusion: The Confidence Trick's New Hardware
Let me synthesize the argument we've built across six parts.
The conventional narrative about American decline focuses on debt, deficits, and the erosion of dollar reserve share. This narrative is not wrong about the problems. It's wrong about the conclusion. The assumption that these problems inevitably lead to British-style decline misses what's being built while everyone watches the old metrics.
The new architecture of American hegemony runs on silicon, not oil. The chokepoints are compute, chips, and AI models, not tanker routes and oil fields. These chokepoints are more concentrated, more defensible, and more thoroughly controlled by American companies than the petrodollar infrastructure ever was.
China can't catch up. The brittle peer thesis holds: impressive on static metrics, hollow on dynamic capability. The chip gap is widening, not closing. The talent flows to America, not away from it.
China can't take Taiwan. The physics don't cooperate. The lift capacity doesn't exist. The weather windows are narrow. The detection is certain. China can punish Taiwan; it cannot conquer it.
The Trump administration is building (consciously or accidentally) infrastructure that extends dollar hegemony into the digital age. Stablecoins backed by Treasuries spread dollars on blockchain rails. The Strategic Bitcoin Reserve hedges the sovereign balance sheet. AI acceleration maintains technological supremacy.
And the AI industry is consolidating in ways that reinforce American dominance. The wrapper massacre clears out marginal players and concentrates value in American infrastructure companies billing in American dollars.
The confidence trick continues. The dollar remains the currency of account for the global economy, not because of American fiscal discipline (there isn't any) but because there's no alternative infrastructure. The pipes carry dollars. The chips process dollars. The models think in dollars.
This could still fail. Crypto volatility could trigger financial crisis. AI safety failures could be catastrophic. Corruption could hollow out institutional legitimacy. The transatlantic fracture could fragment the Western alliance. None of these outcomes are impossible.
But the doom thesis, the assumption that American hegemony is ending because the debt is high and the BRICS are meeting, misses the reconstruction happening beneath the surface. The old architecture is aging. The new architecture is being built. The transition isn't from hegemony to decline; it's from one form of hegemony to another.
The technology wins. The arbitrageurs lose. The dollar wins either way.
Notes
Notes
[1] Builder.ai collapse from TechStartups and financial press coverage.
[2] Humane AI Pin sale from company announcements and multiple reporting.
[3] Jasper valuation decline from Maginative and industry analysis.
[4] VC funding concentration from Crunchbase year-end analysis.
[5] OpenAI and Anthropic valuations from latest funding round reporting, January 2026.
[6] Cursor valuation growth from TechCrunch December 2025 coverage.
[7] Sherlocking pattern and Apple history from industry analysis and documented cases including Konfabulator, Watson, and others.
[8] AI startup survival predictions from industry analysis and venture capital commentary.
