Meituan trains a 1.6T-parameter model on 50,000 domestic chips 美团使用50,000颗国产芯片训练1.6万亿参数模型
Meituan reportedly trained a 1.6 trillion-parameter AI model on 50,000 domestic chips, framing the run as a workaround to U.S. export controls. The report matters because it points to Chinese internet firms building large clusters around local silicon instead of waiting for unrestricted access to Nvidia-class parts. It does not prove performance parity, but it shows that scale-out architecture is becoming part of the export-control contest. 报道称,美团使用50,000颗国产芯片训练了一个1.6万亿参数AI模型,并将其视为绕开美国出口管制的路径。其意义在于,中国互联网公司正在围绕本土硅片搭建大型集群,而不是等待不受限制地获得Nvidia级别器件。这并不证明性能已经对等,但说明横向扩展架构正成为出口管制博弈的一部分。