Meituan says its new LongCat-2.0 model was trained start to finish on Chinese-made chips — and it wants you to notice.
The food-delivery-turned-everything company unveiled LongCat-2.0, a 1.6-trillion-parameter model it claims is the first of its scale trained end-to-end on domestic silicon. The announcement is less about the model's capabilities and more about its supply chain: Meituan is signaling that US export controls, which have blocked Nvidia's most advanced chips from reaching Chinese buyers, have not stopped large-scale AI training. The company did not specify which domestic chips were used.
The timing matters. Washington has steadily tightened restrictions on advanced semiconductor exports to China, betting that cutting off access to leading-edge hardware would slow Chinese AI development. If Meituan's claim holds up, it suggests Chinese firms are further along in working around that bottleneck than US policymakers may have assumed — whether through stockpiled inventory, Huawei's Ascend line, or other domestic alternatives.
That caveat — "if it holds up" — is doing real work here. Meituan has every incentive to present this as a clean domestic-chips story, and the announcement offers no independent verification. Chinese AI labs have made bold hardware claims before that later came with asterisks.