Key Takeaways
- AI distillation allows smaller "student" models to learn from larger "teacher" models, enabling China to build competitive AI systems without advanced chips restricted by US export controls
- DeepSeek emerged as the headline case study — a Chinese lab created a capable, affordable model through distillation that challenged Western AI dominance
- The technique transforms a traditional academic compression method into a national security flashpoint, reshaping US chip export policy and congressional debate
- Distillation works like reverse-engineering a recipe by taste rather than ingredients — China needs outputs, not the original hardware or training data
- This arms race is forcing policymakers to fundamentally rethink what "controlling AI" means in practice
The AI distillation race China refers to the strategic competition where Chinese AI labs use knowledge distillation — training smaller "student" models on outputs from larger "teacher" models — to build cheap, capable systems without needing the massive chip stockpiles US export controls were designed to deny them. It turned an academic compression technique into a genuine national security debate.
Knowledge distillation, explained without a computer science degree
Knowledge distillation is a training method where a smaller "student" model learns to mimic a larger "teacher" model. Instead of training from scratch on raw internet text (expensive, slow, chip-hungry), the student just learns to copy the teacher's outputs — its predictions, its reasoning patterns, sometimes its internal probability distributions.
Think of it like an apprentice electrician shadowing a master for six months instead of doing a four-year degree. You don't get everything. But you get most of the useful stuff, fast, for a fraction of the cost.
Reportedly, distilled models can achieve approximately 70-90% of the performance of larger models while using 40-60% fewer parameters. That ratio is the entire reason this became an arms race instead of a footnote in an AI research paper. Nine times out of ten, a business doesn't need the biggest model available — it needs the cheapest model that's good enough. Distillation is how you get there.
How the race unfolded, 2022 to now
This didn't happen overnight. It built quietly for two years before anyone outside AI labs noticed.
- 2022-2023: Major AI labs reportedly began accelerating research into knowledge distillation as a way to compress large language models into smaller, more deployable versions — mostly for cost and latency reasons, not geopolitics.
- Mid-2023: According to reports, Chinese AI companies and research institutions intensified distillation efforts, recognising it as a potential workaround to US export restrictions on advanced AI hardware and models.
- 2023-2024: US policymakers reportedly grew increasingly concerned, viewing distillation as a way to reverse-engineer or improve upon restricted models without needing the restricted chips themselves.
- Early 2024: Multiple Chinese companies reportedly achieved competitive results using models distilled from larger Western systems — proof the approach wasn't just theoretical.
- Mid-2024: The US reportedly considered expanding export controls to cover not just model weights, but the techniques and training data that enable distillation itself.
- Late 2024: Both US and Chinese entities kept publishing research on distillation efficiency and model compression, treating it as a core competitive battleground rather than a side project.
Every step of that timeline is the same story on repeat: China finds an efficiency workaround, the US tries to regulate the workaround, and the workaround evolves anyway. Wash, rinse, distill.
The DeepSeek distillation controversy
DeepSeek is the name that turned this from an inside-baseball AI policy topic into front-page news. The Chinese lab released models that reportedly performed competitively against Western frontier systems, at a fraction of the reported training cost — and did it, critics allege, partly by distilling from outputs of established US models.
The controversy isn't really about whether distillation is technically impressive (it clearly is). It's about the ethics and legality of training on another company's model outputs without permission, and about what it means when export controls on hardware get outflanked by clever training methods that need less hardware to begin with.
American AI labs poured billions into raw compute. Then a smaller, leaner competitor turns up allegedly having learned by watching over their shoulder. It's the AI equivalent of studying for finals by copying someone else's crib notes — except the notes were extremely good, and the "someone else" spent a fortune writing them.
The numbers behind the hype
Strip away the geopolitics for a second and the numbers alone explain why every lab on earth is obsessed with distillation:
- Distilled models can reportedly hit 70-90% of a larger model's performance.
- They do it using 40-60% fewer parameters — meaning less compute, less energy, less chip demand.
- Fewer parameters means cheaper inference, which means cheaper products, which means faster deployment at scale.
That efficiency math is exactly why distillation matters more to China's AI strategy than raw model size does. If you can't buy (or aren't allowed to buy) the newest Nvidia chips in bulk, the smart move isn't to sulk about it — it's to build something that needs fewer chips to run well. Distillation is the shortcut. Not a cheat code exactly, but close enough that Washington started treating it that way.
Can distillation actually bypass US chip export controls?
Partially, yes — and that's the uncomfortable bit for policymakers. Export controls were built to restrict access to the most powerful chips, on the theory that without them, you can't train frontier-scale AI models. Distillation punches a hole in that theory.
If a smaller student model can learn 70-90% of a teacher model's capability using far fewer parameters, then a country facing chip restrictions doesn't need to match the original training compute at all. It just needs enough compute to run the distillation process on outputs it can access — through APIs, published research, or open-weight releases.
That's reportedly why the US considered expanding controls in mid-2024 to cover not just chips and model weights, but the techniques and training data that make distillation possible in the first place. It's a much harder thing to police than a chip shipment. You can put a chip on a restricted-export list. You can't exactly put a training method on one — the maths doesn't care which country it runs in.
How the US is responding
Washington's response has been part regulation, part reinvestment. Policymakers have debated tightening controls around API access to frontier models (to make distillation harder), while US labs have simultaneously ramped up their own distillation research — because if the technique works this well, ignoring it isn't a strategy, it's a self-own.
There's a real irony here worth sitting with: the same labs lobbying for tighter export controls are also racing to build better distilled models themselves, because smaller and cheaper is good business regardless of which country you're in. Export control policy is trying to slow down a technique that every serious AI lab, American included, actually wants to get better at.
The open-source twist nobody planned for
Here's the part most coverage glosses over. Distillation isn't just a China-vs-US story — it's also an open-source story, and open-source doesn't respect export control maps.
When a lab publishes model weights or research openly (which plenty of US and Chinese labs both do, for reputation and recruitment reasons), it hands the entire world a teacher model for free. You don't need to be a nation-state actor to distil from an open model — a well-funded university lab or even a serious hobbyist with enough GPU credits can do it. Export controls assume the choke point is hardware. Open-source releases quietly remove that choke point for anyone patient enough to work the distillation process themselves.
That means the "race" isn't really two countries anymore. It's two countries, plus every open-source community sitting in the stands, occasionally wandering onto the pitch.
Where distillation goes from here
The next phase of this fight probably isn't about bigger models at all — it's about who distils fastest and cheapest. Reportedly, both US and Chinese entities continued publishing research on distillation efficiency and model compression through late 2024, treating it as a core focus area rather than a side quest.
Expect three things to keep happening: export control debates expanding beyond hardware into "does this technique count as a controlled technology", frontier labs quietly restricting API access to make distillation harder for competitors, and smaller/regional AI labs (not just Chinese ones) leaning harder into distillation because it's simply the most cost-effective way to compete without a hyperscaler's budget.
The moral, if there is one: the race to build the biggest model may already be less important than the race to build the smartest small one.
My honest take: the export control strategy has a hole in it, and it's shaped like an equation
Here's my one strong opinion on this, and I'll back it with the numbers already on the table: chip export controls alone cannot contain AI capability diffusion, because distillation lets you achieve 70-90% of a frontier model's performance with 40-60% fewer parameters — and parameters, not raw silicon, are what export controls were actually designed to gatekeep.
Restricting Nvidia's top chips slows down who can train the next GPT-scale teacher model from scratch. Fair enough, that part works. But it does almost nothing to stop a well-resourced lab from distilling a smaller, still-highly-capable model from any teacher model whose outputs are reachable — via API, via leaked weights, via published research. You can't export-control a technique that runs on maths anyone can read in an arXiv paper.
The actionable consequence for policymakers: if you're serious about slowing distillation-based capability transfer, the fight has to move from hardware controls to access controls — rate-limiting and monitoring who can pull enormous volumes of outputs from frontier model APIs, because that's the actual raw material distillation needs. Chips get you the teacher. API access gets you the student. Right now, the US is guarding the front door of the building while the fire escape stays wide open.
Where this doesn't apply: for genuinely frontier capability — the next major reasoning leap, not incremental performance — distillation caps out at whatever the teacher model already knows. A student can't outlearn its teacher. So if the concern is "will China copy their way to the next GPT-5-level breakthrough," the honest answer is no, not through distillation alone. That's a different, harder problem, and conflating the two just muddies good policy with bad panic.
What is AI distillation and how does China use it?
AI distillation trains a smaller "student" model to mimic a larger "teacher" model's outputs, producing a compact model that keeps most of the capability at a fraction of the training cost. Chinese AI companies reportedly use it to build competitive models without needing the largest chip stockpiles US export controls restrict — cheaper compute, similar results, fewer questions asked.
Is China copying US AI models through distillation?
Partly, and that's exactly the controversy. Distillation involves learning from a teacher model's outputs, which critics argue crosses a line when that teacher model belongs to a competitor who didn't consent to it. DeepSeek's rise reportedly reignited this debate, though the technique itself is standard practice across the whole AI industry, not a uniquely Chinese invention.
How does model distillation work in AI development?
A large teacher model generates outputs — predictions, reasoning steps, probability scores — and a smaller student model is trained to replicate those outputs as closely as possible. The student never sees the original training data directly. It just learns to imitate the teacher's behaviour, which reportedly gets it to 70-90% of the teacher's performance using 40-60% fewer parameters.
DeepSeek vs OpenAI: who is winning the AI distillation race?
Depends how you score it. DeepSeek reportedly achieved competitive performance at dramatically lower reported cost, which is a win on efficiency. OpenAI and other Western labs still reportedly lead on raw frontier capability, since distillation can't exceed what the original teacher model knows. It's less a knockout and more a points decision, round by round.
How much cheaper is AI distillation compared to training from scratch?
Training a frontier model from scratch requires enormous compute for the full training run. Distillation skips most of that by learning from an existing model's outputs instead, reportedly enabling 70-90% of the performance with 40-60% fewer parameters — which translates directly into less compute, less energy, and lower deployment cost per query.
What is knowledge distillation in machine learning for beginners?
It's teaching a small AI model by having it copy a bigger, smarter one — like a student learning by watching a expert solve problems rather than working through every textbook page themselves. The small model ends up faster and cheaper to run, while keeping most of the original's smarts. Not quite cloning. More like a really efficient study buddy.
Can distillation bypass US chip export controls in China?
To a meaningful extent, yes. Export controls restrict access to top-tier chips needed for training massive models from scratch, but distillation needs far less compute since it learns from an existing model's outputs. Reportedly, this is exactly why the US considered expanding controls in 2024 to cover distillation-enabling techniques and training data, not just hardware.
Is the AI distillation race with China actually a real threat?
It's real, but it's specific — a threat to the effectiveness of hardware-based export controls, not necessarily to overall US AI leadership at the frontier. Distillation lets China build cheap, competitive, "good enough" models fast. It does not let anyone leapfrog past a teacher model's actual ceiling, since a student can't out-know its teacher.
Will distillation eventually make export controls pointless?
Not pointless, but incomplete on their own. Export controls still slow down who can train the next-generation frontier model from raw scratch, which matters. But reportedly, policymakers are already looking at API access limits and technique-level restrictions, because chip controls alone don't stop distillation from an accessible teacher model.
So that's the state of play: a compression trick born in research labs now sits at the centre of US-China tech policy, congressional hearings, and a genuinely fascinating game of cat-and-mouse where the mouse keeps finding smaller doors. Export controls guard the chips. Distillation just asks nicely to borrow the recipe instead. Whether Washington closes that gap or keeps chasing it is the actual story to watch in 2025 — not who has the biggest model, but who's smart enough to need the smallest one.
