The Quiet Engine Behind Cheap AI: AI Distillation

On 27 January 2025, Nvidia lost roughly $589 billion in market value in a single trading session. The trigger was DeepSeek-R1, a reasoning model from a Chinese lab that matched far more expensive systems at a fraction of the price. Two days later, OpenAI said it had seen evidence that DeepSeek had used its models to train its own, through a technique called distillation.

Most people had never heard the word. Within weeks, it was a question for markets, regulators, and courts. When I went looking for one place that explained what distillation is, what it costs, who does it, and whether it's allowed, I couldn't find one. So I built it.

What distillation actually is

Distillation is a way of teaching a small model to imitate a large one. The large model, called the teacher, answers a huge number of prompts. The small model, the student, is trained to reproduce not just its answers but the probabilities behind them: how confident the teacher was, and which wrong answers it nearly gave. That extra signal carries a surprising amount of the teacher's knowledge.

The idea is older than the current AI boom. Geoffrey Hinton, Oriol Vinyals and Jeff Deannamed it in a 2015 paper, "Distilling the Knowledge in a Neural Network," building onmodel-compression work from 2006. What changed is scale. Today distillation is how labs turntheir most expensive models into the fast, cheap ones people actually use. Google's technicalreports describe Gemini Flash and the open Gemma models as distilled from larger siblings.Meta built its smallest Llama 3.2 models the same way. DeepSeek released a whole family ofopen "R1-Distill" models, trained on output from its flagship.

Why it became a fight

Here is the tension. When a lab distills its own model, it's a product feature. When someoneelse distills that model through its public API, the lab calls it theft. Most major AIproviders' terms forbid using their outputs to build competing models, while those sameproviders rely on distillation for their own product lines.

The economics make the tension sharper. A frontier training run costs tens to hundreds ofmillions of dollars. Distilling some of that capability can cost almost nothing bycomparison: the NovaSky team reported training its Sky-T1 reasoning model for under $450.When copying is that much cheaper than building, it becomes a question of export controls,contract law and national strategy, not just engineering.

What the compendium covers

AI Global Distillation (https://global-distillation.com) is a free public reference that looks atthe subject from eight angles:

·                     Academic. The research record, from the early compression papers to the 2025–26 wave ofreasoning distillation, with 46 key papers and teacher-versus-student benchmark comparisons.

·                     Financial. Token prices for 63 models, documented training costs, and a cost model forwhen self-hosting a distilled model beats paying for an API.

·                     Political. 25 policies across jurisdictions and 14 named disputes, including the OpenAIand DeepSeek allegations.

·                     Company. How 18 AI companies use distillation, what their terms of service say about it,and whom they have accused.

·                     Developer. 27 tools and 8 step-by-step recipes for distilling a model, with realisticcosts.

·                     Customer. 74 models compared on price, quality, speed, context length and licence.

·                     Library. 27 distillation methods, each explained with its loss function, its data needs,and where it fails.

·                     Timeline. 130 dated events from 2006 to today.

How it's built

Every chart and table on the site cites its sources, 468 distinct ones in total, from arXivpapers and company documentation to government filings and news reporting. Where a numbercouldn't be verified, the site says "undisclosed" rather than guessing. Every morning anautomated job pulls fresh data from arXiv, Hugging Face, GitHub and Hacker News and rebuildsthe pages, so the live figures stay current without anyone touching them. The underlying datais published as plain JSON, free to reuse with attribution.

It also has limits worth stating plainly. Most benchmark scores are reported by the companiesthat built the models, not measured independently. The legal questions are genuinelyunsettled, and the site sets out the arguments on each side rather than declaring a winner.And any single compendium will contain errors, so corrections backed by a source are welcome.

Why it matters

If you build with AI, distillation decides which models you can afford to run. If you buy AIservices, it explains why a small model can come close to a flagship at a fraction of theprice. And if you follow technology policy, it sits at the center of how the US, China andEurope are trying to regulate AI.

AI Global Distillation aims to put the evidence for all three in one place. Take a look athttps://global-distillation.com, and if you find something wrong, tell me.