Open-Source AI: You Get the Dish, Not the Recipe

Most "open-source" AI models aren't actually open source. They're "open weight" — and that distinction explains why giving away billion-dollar models isn't as generous (or as reckless) as it looks.

(Note: I use "open source" in the loose industry sense here; most of these releases don't meet the OSI definition.)

What Open Weight Actually Means

YOU GET THE DISH, NOT THE RECIPE

A model's capabilities live in billions of trained parameters — the weights. Open weight means putting that file online for anyone to download, run, fine-tune, and commercialize.

True open source would also include the training code, training data, data curation methods, and process details. Almost nobody releases those.

Think of it as a frozen meal: reheat it, add ingredients, serve it however you like. But without the recipe, you can't recreate it from scratch.

Two implications: → The crown jewels stay home. The data recipe and training craft — what enables iteration — never leave the building. → You're always giving away yesterday's dish. Once the next-gen model is trained, the marginal cost of releasing the old one drops sharply.

Who Actually Uses Open Models

WHO ACTUALLY EATS THE FROZEN MEAL? THREE LAYERS

• Enterprises: self-hosted deployment, so data never leaves the company. For banks, hospitals, and law firms with compliance needs, open models are increasingly the default answer.
• Developers: cheap building blocks. Download from Hugging Face, run locally via vLLM or Ollama, or use third-party inference platforms at a fraction of closed-API prices.
• Consumers: invisible coverage. Many apps quietly run Qwen or Llama in the backend; on-device translation and voice assistants are often compressed open models.

The result: the whole world's kitchens cook to your taste — while the better next-gen recipe stays in your hands.

Closed-Source Pricing Power Loosens

THE SHOCKWAVE: CLOSED-SOURCE PRICING POWER IS LOOSENING

When "good-enough intelligence" costs a fraction of the price, closed APIs hit a visible ceiling. Companies are moving from token-maxxing to strict ROI reviews — public reports suggest Uber and Microsoft now restrict pricey AI tools, and Coinbase reportedly cut internal AI spend substantially by switching to open models.

Palantir's Alex Karp calls token billing a "wealth tax" on enterprises. Self-serving? Sure — Palantir sells the alternative. But the underlying trend is real: open models force closed labs to keep proving the frontier premium is worth it.

Where Value Will Ultimately Settle

THE REAL BET

Closed labs bet a recipe lead keeps commanding token revenue. Open-weight players bet the giveaway buys something more valuable — ecosystems, chip sales, or a rival's eroding moat. Deployment vendors bet what matters most is whose turf the AI runs on.

Where will AI's value ultimately settle — models, compute, ecosystems, or deployment? No answer yet. One certainty: that free frozen meal has already reshaped the pricing of the entire menu.