The most talked-about idea in artificial intelligence right now isn't a new chatbot. It's a technique for shrinking existing ones — and it's quietly rewriting the economics of the entire industry.
The method is called distillation. In simple terms, a large, expensive "teacher" model is asked thousands of questions, and a much smaller "student" model is trained to copy its answers and reasoning. The student never sees how the teacher was built; it simply learns to imitate the output. The result is a model that costs a fraction as much to run, yet performs remarkably close to the original on the tasks it was trained for.
Why this blew up
Distillation went from academic curiosity to boardroom topic in early 2025, when Chinese lab DeepSeek released a reasoning model built at a fraction of the cost of its American rivals. The shock briefly wiped hundreds of billions off Nvidia's market value in a single day, as investors questioned whether cutting-edge AI really required cutting-edge budgets. OpenAI later said it had evidence DeepSeek may have distilled from its models' outputs — something its terms of service forbid.
Since then, university researchers have demonstrated that capable reasoning models can be distilled for a few hundred pounds' worth of computing power. Not millions — hundreds.
That creates an awkward problem for the big labs. Their public interfaces are precisely how customers use their products, but they're also how competitors can copy years of research. If anyone can query your flagship model at scale and compress its abilities into their own software, the advantage of having built it starts to evaporate. That's why the debate has moved beyond engineering circles into policy discussions in Washington, touching on export controls and intellectual property.
What it means for smaller firms
For UK business owners, this is largely good news, for three reasons.
Prices keep falling. Distilled models are far cheaper to run, and that competition is dragging down the cost of AI across the board. Tasks that required a premium subscription a year ago increasingly work fine on budget-tier tools.
AI is coming to ordinary hardware. Small models can run on laptops, phones and in-store devices rather than distant data centres. That matters if you handle sensitive customer data or work somewhere with patchy connectivity — the AI can stay on your premises.
More choice, less lock-in. If capability can be copied, no single provider holds a permanent monopoly on smart software. That gives smaller buyers more negotiating power and more alternatives if a supplier raises prices.
The caution: cheap distilled models are good at what they were trained to copy, and less reliable outside it. If you're building AI into your business, test it on your actual tasks rather than trusting benchmark claims.
What to watch next
Expect the big labs to fight back with tighter rate limits, watermarked outputs and stricter terms of use — and expect legal disputes over who is allowed to learn from whose model. Also watch for regulation: governments are waking up to the fact that AI capability can now cross borders through nothing more than an API subscription.
The strategic lesson is simple. In AI, the moat is no longer just having the biggest model — it's controlling distribution, data and customer relationships. For small businesses, that shift means more affordable tools arriving faster than anyone predicted.