The All-You-Can-Eat Trap: When AI Companies Mistake Discount-Market Fit for Product-Market Fit

Why Cursor might be more like Red Lobster than Netflix

Photo by Edwin Petrus on Unsplash

When I was a student at the University of Texas at Austin, I discovered CiCi's Pizza buffet. For $6—less than what a single slice costs at most places today—you could eat unlimited pizza, pasta, salad, and dessert. My friends and I would show up on Saturday afternoons and eat with complete abandon, stacking our plates high, going back for seconds, thirds, sometimes fourths. It didn't seem legal. How could any business survive letting college students—notorious for their bottomless appetites and empty wallets—eat as much as they wanted for the price of a fast-food meal?

But CiCi's wasn't naive. Like other buffet restaurants they had figured out two dynamics that make the business work. First, not all customers eat like college students. Most people think they'll eat four plates but actually eat two. The psychological satisfaction of "unlimited" drives the purchase, but human biology limits actual consumption. CiCi's knew that my friends and I would be subsidized by the families with young kids who barely finished one plate. The heavy eaters brought in the crowds, but the light eaters paid the bills. Second, buffet operators control their costs ruthlessly—strategically placing expensive proteins in hard-to-reach spots while making cheap starches like bread and pasta easily accessible.

This same principle powers some of the most successful companies in tech. As Netflix built their streaming platform, they worked ruthlessly to control their input costs—just like CiCi's creating more garlic bread and pasta instead of expensive protein. When Netflix spent 100 million on House of Cards, they were investing in their own "pasta"—content they owned and controlled instead of paying $100 million just to keep Friends for one year. Friends was the pepperoni pizza—expensive, popular, but outside their control. Now Netflix spends the "vast majority" of its $17 billion content budget on originals and uses algorithmic recommendations to nudge users toward their cheaper in house content instead of expensive licensed hits.

Spotify faces the same challenge but can't replace Taylor Swift—she's the pepperoni pizza that brings people to the platform. Instead, they've quietly implemented what they call "Perfect Fit Content" program, systematically replacing well-known artists like Brian Eno and Jon Hopkins on mood playlists with cheaper stock music. When you search for "Ambient Chill," you're getting the musical equivalent of garlic bread—it fills you up, but costs Spotify almost nothing in royalties. Spotify pays out 74% of its revenue in music royalties—$10 billion in 2024 alone—so every stream they can shift from Taylor Swift to their house-brand alternatives directly improves their margins.

Both companies understand the same customer segmentation that made CiCi's work: heavy users are subsidized by light users, but the key is controlling what the heavy users actually consume. Netflix knows some subscribers binge entire seasons while others watch one show per month. Spotify knows some users stream music all day while others listen occasionally. The magic happens in the algorithmic nudging—steering the heavy users toward the cheap pasta while keeping just enough pepperoni pizza to maintain the illusion of unlimited choice.

The AI Wrapper Gold Rush: When Everything Seemed Perfect

Then came the AI wrapper companies—startups that built products on top of frontier models from research labs like OpenAI and Anthropic. These companies didn't train their own large language models from scratch. Instead, they created brilliant user experiences that made cutting-edge AI accessible through monthly subscriptions. It seemed like the perfect business model: let the research labs handle the expensive model development while you focused on building great products.

Cursor became the poster child for this approach. The AI-powered code editor that became the darling of the developer community built their platform on top of the world's most powerful models—GPT-4, Claude, and others—integrated seamlessly into a coding workflow. For $20 a month, Pro users got 500 fast responses on advanced AI models, then unlimited responses at a slower rate. It seemed like magic: write a comment describing what you want, and watch as sophisticated AI generated entire functions, debugged complex problems, and refactored legacy code.

I was one of those early subscribers. Like thousands of other developers, I was amazed by what Cursor could do. The experience felt transformative—suddenly I could code faster and tackle problems that would have taken hours of Stack Overflow searching. The $20 monthly fee seemed like a steal for access to the most advanced AI models on the planet.

The business model looked brilliant, almost Netflix-like in its sophistication. Cursor was simultaneously building their own "home-grown models" for features like tab completion while providing access to expensive third-party models for complex tasks. The plan seemed obvious: attract users with unlimited access to premium models, then gradually transition them to Cursor's own cheaper alternatives as the technology improved. Just like Netflix had weaned users off expensive licensed content by creating irresistible originals, Cursor would eventually replace costly Claude and GPT-4 calls with their own models. The heavy users would stay for the experience, the light users would subsidize the costs, and everyone would be happy. The numbers seemed to prove it worked: from $1 million to $100 million in annual recurring revenue in just 12 months—the fastest growth in SaaS history.

It was the perfect buffet business for the AI age. Or so it seemed.

When Input Control Breaks, Everything Breaks

But buffet businesses are not easy. They are a delicate balance and when the balance is not there the business falls over. When companies can't control their input costs, the results are catastrophic, and the pattern is always the same: a death spiral where the heaviest users stay while everyone else leaves.

Red Lobster learned this lesson the hard way. In 2023, they made their limited-time "Ultimate Endless Shrimp" promotion permanent, charging $20 for unlimited shrimp. The result? They lost $11 million in a single quarter from this one decision. Traffic increased only 4%, but costs exploded as heavy shrimp consumers discovered the arbitrage faster than Red Lobster anticipated. The company filed for bankruptcy within months.

The deeper problem wasn't just the promotion—it was that Red Lobster couldn't control what customers actually wanted. Unlike CiCi's, which could nudge customers toward cheap bread and pasta, Red Lobster customers came specifically for expensive shrimp. There was no cheaper substitute to offer, no way to engineer portion psychology. When you promise unlimited access to a premium product you can't control or replace, you're not running a business—you're running an unsustainable discount program.

MoviePass followed the identical trajectory. They charged $9.95 per month for unlimited movies while paying theaters full ticket price—typically $12-15 each. The math only worked if customers watched fewer than 0.77 movies per month. Instead, actual usage averaged 2.11 movies per month by 2018, burning through $40 million monthly. Light movie-goers quit because the value proposition was poor for them. Heavy movie-goers stayed and increased their usage. The company desperately tried blocking popular movies and limiting showtimes, triggering a user revolt that led to bankruptcy.

The pattern is mechanical: when you can't control input costs, adverse selection kills you. The college students discover the discount and keep showing up but families stop coming.

The AI Trap: Nobody Wants Yesterday's Model

This might be the trap that AI-wrapper companies have walked into, and they're discovering it in real time. The key insight is: Nobody wants yesterday's model.

When Netflix launches a new original series, users continue to watch old series. When Spotify adds new music, listeners also listen to older songs. But in AI, the moment a new frontier model launches, demand for older models collapses. Users immediately notice the difference between GPT-4 and GPT-3.5, between Claude 3.5 Sonnet and Claude 3 Haiku. There's no hiding inferior performance behind clever UX or algorithmic nudging.

Cursor thought they were Netflix, building their own content to reduce costs. But they were actually Red Lobster, promising unlimited access to premium products they couldn't control or replace. Users didn't come for Cursor's models—they came for Claude and GPT-4, the expensive shrimp of the AI world. And just like Red Lobster's customers, they could tell the difference immediately.

The blowback was swift and brutal. When Cursor tried to move away from unlimited usage, implementing caps and higher pricing, the beloved developer community turned hostile overnight. Reddit threads that once praised Cursor now called it "absolute garbage" and accused the company of a "bait and switch." The company was forced to issue refunds and apologize for "missing the mark."

When the VC-funded discount on frontier model access goes away, does the AI wrapper offer enough value to keep using it? Or will users migrate directly to products from the model providers themselves—Anthropic's Claude Code, OpenAI's Codex CLI, and Google's Gemini CLI? The model providers can control their costs because they own the models. They don't need to pay API fees to themselves.

Looking back at my college days at CiCi's Pizza, the business model worked because CiCi's controlled both sides of the equation—what they served and what customers actually consumed. They could nudge us toward the cheap pasta and away from the expensive pepperoni. But when you can't control the input costs, when your customers demand the premium product and can immediately tell the difference, the math stops working.

The question that should terrify every AI wrapper company executive is: Do you have product-market fit, or just discount-market fit? Will people keep coming when you stop serving free shrimp?

Related