DeepSeek vs OpenAI: how far behind is China's frontier?

The head-to-head, mid-2026

Is DeepSeek just distilling Western models?

Distillation allegations exist, but Epoch AI characterizes DeepSeek innovations like Multi-head Latent Attention (MLA) and fine-grained Mixture-of-Experts as genuine advances — not mere copying. And there's a twist: a world where the frontier can be cheaply distilled is itself a world where the proprietary moat doesn't hold.

Why the narrow gap matters

Aschenbrenner predicted open source would fade and proprietary algorithms would form a durable US moat. The opposite happened — capability diffused, pricing collapsed, and the gap compressed to months. That's the evidence behind the scorecard's one Wrong verdict, and it directly weakens the US–China moat story.

Frequently asked questions

As of mid-2026, open-weight models like DeepSeek V4 trail the proprietary frontier by only ~3–6 months, at a fraction of the cost, with genuine architectural innovation — a much narrower gap than expected.

Distillation claims exist, but Epoch AI rates innovations like Multi-head Latent Attention and fine-grained Mixture-of-Experts as genuine advances. A frontier that can be cheaply distilled also undermines the durable-moat thesis.

Yes, but not durably. The lead is months, not years, and closing — which is why Aschenbrenner's 'open source fades, proprietary moat holds' prediction is graded Wrong.