Last month, I watched a demo that stopped me cold. Two developers, working over a weekend with open-source tools, had built a chatbot that performed indistinguishably from GPT-4 on a set of business tasks. They didn’t raise a dime. They didn’t train from scratch. They just assembled pieces that were already free. And I realized: the $100 billion being poured into foundational AI models is chasing a mirage.
You’ve probably felt the anxiety. Every tech conference, every pitch deck, every quarterly earnings call is drenched in generative AI hype. The narrative is simple: whoever owns the best model wins. But that story is collapsing from the inside, and the relief you feel when you sense it might not be wrong. It’s the right reaction.
The best AI model in the world is the one you don’t have to pay a license for.
Here’s what’s actually happening. The marginal cost of running inference on a state-of-the-art model is plummeting toward zero. Open-source alternatives like Meta’s LLaMA family, Mistral, and a dozen others are closing the performance gap every quarter. Meanwhile, frontier labs are burning cash on training runs that cost tens of millions—only to watch their lead evaporate within months. This is not a temporary dip. It’s a structural shift.
The tension is delicious: generative AI is the most hyped technology in history, yet its core product—the model itself—is becoming a commodity with no defensible moat. Investors are still pouring money into model companies as if they were building high-walled castles. Instead, they’re building sandcastles at low tide.
But here’s the twist: worthlessness is a feature, not a bug.
Think about it. When was the last time you paid for an operating system? When did you last worry about which database engine to license? The greatest technologies become invisible utilities. Electricity isn’t a competitive advantage—it’s a precondition. The internet is not a product; it’s the floor. Generative AI is heading for exactly the same fate, and that’s the best possible outcome.
I spoke with a founder who raised $80 million to build a proprietary model. When I asked him what happens when open-source catches up, he paused. Then he admitted: “We’re already pivoting to a data-moat play.” The smart ones are running ahead of the collapse. They’re building application layers on top of commoditized models, wrapping them with proprietary data and workflow logic that actually solves real problems.
Your competitive advantage was never going to be the model. It was always your data, your distribution, your relationships, your ability to understand a messy domain and build a tool that doesn’t just generate text but transforms a process. The model is the engine; the value is in the car, the driver, and the route.
Open-source is eating AI’s lunch, and investors are still setting the table.
The emotional payoff here is not fear. It’s relief. The bubble narrative—that you must race to build a model or be left behind—is toxic and false. The sane response is to stop treating models like rare earth minerals and start treating them like plumbing. Focus on what only you can own: your customer’s trust, your domain expertise, your proprietary data.
This is not a prediction. It’s already happening. Anthropic and OpenAI are fighting over who can give away more free usage. Meta releases models with a license that basically says “go ahead, build a business.” The genie is out of the bottle, and it’s not going back in.
Worthlessness is the highest form of democratization.
So stop worrying about which foundation model to bet your career on. Stop agonizing over whether GPT-5 will be better than Gemini 3. Start asking a different question: What will you build when intelligence is free? When every startup, every mid-market firm, every school and hospital has access to the same raw capability, the only differentiator becomes execution. That’s not a threat. That’s the most exciting opportunity in a generation.
The day you realize that your model is worthless is the day you finally start building something valuable.
FAQ
Q: But aren't frontier models still getting better? Won't proprietary data matter less over time?
A: Frontier models improve, but the gap with open-source alternatives is closing faster than ever—often within months. Proprietary data will actually matter more as models become commodities, because unique data creates unique outcomes. The model is the engine; the data is the fuel that differentiates your car.
Q: What's the practical implication for someone building a startup or investing in AI?
A: Shift your focus away from model-weight moats. If you're building a startup, concentrate on vertical solutions that own the user experience and workflow, wrapping commoditized models with proprietary data. If you're investing, look for companies that own distribution, domain expertise, or unique datasets—not those hoping to sell a better model.
Q: Isn't this just another 'software is eating the world' platitude dressed up as insight?
A: No—this time the commodity is intelligence itself. The transition from proprietary AI to open utility will be faster and more disruptive than previous shifts because the cost of replication is effectively zero and the community is global. The incumbents' moats are evaporating in real time, and the winners will be those who embrace the utility mindset from day one.