Commoditize your complements
A complement is anything customers buy alongside your product — hot dogs and buns, cars and petrol, consoles and games. Demand for your product rises when its complements get cheaper. So a smart company spends to drive down the price of its complement, knowing the demand it unlocks flows straight to itself. Joel Spolsky’s line: smart companies try to commoditize their products’ complements.
The mechanism
Two goods are complements when cheaper one means more of the other. Drop the price of buns and people buy more hot dogs. So if you sell hot dogs, you’d love buns to be free — and you’ll happily fund whatever makes them so. You’re not being generous; you’re moving the demand curve for your own product, then capturing the surplus where you have pricing power.
The classic moves: Microsoft benefited from cheap, commoditized PC hardware (the OS was the profit, the box was the bun). Netscape tried to commoditize the OS underneath the browser. Open-sourcing is the sharpest version — you don’t just lower a complement’s price, you drive it toward zero.
The AI case: NVIDIA commoditizing the model layer
NVIDIA sells GPUs. Its complement is the model — every capable model that exists creates demand for compute to train and serve it. So NVIDIA has every incentive to make capable models cheap and abundant, which is exactly what shipping its own strong open models (the Nemotron family) does. Released into the ecosystem, they pressure the closed labs’ pricing, proliferate AI deployment, and every incremental token of inference runs on… GPUs.
Jensen Huang says the quiet part plainly: “tokens are now profitable, so model makers are in a race to produce more” — and in that race, the picks-and-shovels seller wins regardless of which model is on top. Commoditize the model; sell more of the complement.
The other AI players running the same play
- Meta / Llama. Meta doesn’t sell model access — its business is engagement and ads. A commoditized, free model layer denies rivals (OpenAI, Google) a proprietary platform to control, while Meta gets frontier-ish capability for its own products. Open-sourcing weakens the complement that could otherwise have become a competing platform.
- Cloud providers. Hyperscalers want models cheap and plural because their product is the compute and the surrounding services; a commoditized model layer maximizes consumption of what they actually monetize.
In every case the pattern is identical: whoever profits from the layer next to the model wants the model itself to be a commodity.
How to use it
When a powerful company gives something valuable away, don’t take it at face value — ask the complement question:
- What do they actually sell? The free thing is rarely it.
- Is the free thing a complement to the paid thing? If cheaper-of-it means more-of-theirs, you’ve found the play.
- Who gets hurt? Commoditizing a layer destroys the pricing power of whoever depended on selling it.
This is the strategy a player runs; its market-level consequence — substitutes capping what anyone can charge — is commoditization & the substitute price floor. And it decides where value accrues in the stack: the commoditized layer bleeds margin to the layer that engineered it.
Any time Peregrinations explains why a company open-sources or gives away something valuable — NVIDIA’s open models, Meta’s Llama — it links here.