The AI frontier just stopped being a moat

The interesting question in AI just stopped being who has the biggest model. It became what gets built now that the frontier has been copied, shared, and rebuilt cheaper than the labs expected.

Most readers will read Moonshot AI's Kimi K3 release as a US/China horse race: a Beijing start-up catching up, the Nasdaq selling off chip names, the usual geopolitics. The pattern underneath is quieter. The Week's coverage pegs K3 as the world's first open-weight model in the three-trillion-parameter class, with Moonshot claiming parity with Anthropic's Opus 4.8 and OpenAI's 5.5 on coding and agent benchmarks at roughly a third of the compute cost. Open-weight means the learned parameters are downloadable and modifiable. That is not a national victory. It is frontier infrastructure becoming a public good.

The mechanism is portable. A capability that lived inside a handful of US labs is now something a developer, a mid-sized company, or a regulator can run, inspect, and fork. The competitive question shifts from who can train the biggest model to who builds the most defensible thing on top of it. The labs that treated the frontier as a moat are now competing on a field the moat no longer encloses.

The Anthropic and OpenAI accusations about distillation are real, as is Graham Webster's observation that the gap closed under years of US export controls — "there is real innovation going on," he said. Both can be true. The next round of AI value will be settled in the products, policies, and institutions that run on top of a frontier anyone can now download.