Distillation and Open Source: How AI Capabilities "Leak" Downmarket

Here's what keeps AI investors up at night: capabilities that cost hundreds of millions to train show up, within a year, in models dozens of times cheaper — or free. Where's the moat?

Two Channels for Capability Flow

Capabilities flow downmarket through two channels.

Channel 1: Distillation

Channel 1: Distillation — photocopying a big model's abilities. The idea (Hinton et al., 2015): let a small model learn from a big model's outputs instead of raw data. In the LLM era it's even simpler — have a strong model generate Q&A pairs, reasoning traces, and code at scale, then train a small model on them. You never touch the teacher's parameters; you only need its answers. That's why, on many tasks, distilled small models now match or beat earlier flagships.

Channel 2: Open Source

Channel 2: Open source — publishing the knobs outright. Llama, Mistral, DeepSeek, Qwen keep raising the capability floor. When a free, self-hostable model nears flagship performance on key benchmarks, closed vendors' pricing power gets squeezed hard.

Legal Gray Zone

Why the constant legal fights? Terms of service generally ban using outputs to train competitors. OpenAI claimed in early 2025 that DeepSeek had distilled its models, and escalated in Feb 2026 with a memo to the U.S. House. But proof is the hard part: model outputs are everywhere, and tracing which model a capability was "learned" from is nearly impossible. Technically unpreventable, legally gray.

The Pharma Comparison

The economics? Compare pharma. A new drug costs $1B+ and recoups via ~20 years of patent protection; then generics sell for pennies. AI capabilities are protected mainly by ToS and trade secrets — no statutory monopoly. The leak cycle isn't decades. It's months.

The Moat Is Shifting

So the model itself is a fading moat. The real barriers are shifting: proprietary data, user distribution, engineering efficiency.