A Food-Delivery Company Just Trained a Frontier Model Without a Single Nvidia Chip

The company that just showed Nvidia's chip export controls might not be the moat everyone priced them as isn't a frontier AI lab. It delivers food.

Meituan open-sourced LongCat-2.0 on June 30 — a 1.6-trillion-parameter mixture-of-experts model with a native 1-million-token context window, released under a permissive MIT license. The detail that matters more than the parameter count: it was pre-trained and served entirely on domestic Chinese ASICs, no Nvidia hardware anywhere in the pipeline, at a scale no one outside a dedicated frontier lab had previously managed on non-Nvidia silicon.

It had already been winning before anyone knew its name. Under the codename "Owl Alpha," the model was quietly topping OpenRouter's usage charts on coding tasks for weeks — near-frontier performance, unattributed, until Meituan claimed it. That sequence is its own signal: buyers picked the model on merit before they knew who built it or on what hardware, which is not how a compute-scarcity story is supposed to play out.

Two theses that a large share of AI infrastructure capital has been priced against just took a direct hit. The first is that model quality is bottlenecked by Nvidia access, which is the entire premise behind export controls functioning as an American moat. The second is that "AI infrastructure investing" is functionally synonymous with "compute scarcity investing." If a scaled, near-frontier model can be trained on domestic ASICs by a company whose core business is delivering lunch, chip access stops being the constraint that separates who can build frontier AI from who can't.

The caveat is real and worth sitting with: as of this week, LongCat-2.0's weights are listed as "coming soon," with only the inference framework and infrastructure code actually released. Meituan has announced the achievement before fully substantiating it in public — a pattern China-watchers have learned to treat with some skepticism until the download link works. Call the moat-erosion thesis provisional, not proven.

But provisional is still directionally uncomfortable for a specific class of venture bet: capital deployed on the assumption that chip scarcity is a durable, multi-year wall between American and Chinese AI capability. If Meituan's claim holds up even partially, frontier-scale model training is becoming a commodity capability that any sufficiently capitalized tech company can bolt onto its balance sheet — the way any retailer can now stand up a data warehouse, not the way only a handful of companies on earth can build a fab.

Where does that leave the value? Exactly where the application-layer thesis has been arguing it belongs all along — in data, distribution, and the depth of integration into a real workflow, not in who has the biggest GPU cluster. A model lab's compute advantage is worth less every quarter that a food-delivery company proves it can be replicated from a standing start. The defensible position was never the chip. It was always what you do with the customer relationship once the chip stops being scarce.