A foundry tuning exercise that missed its performance targets has pushed a Korean edge AI chip designer’s second processor back about six months.
Mobilint had expected Regulus, its next neural processing unit, to reach volume production by the end of 2026. The schedule now points to no earlier than June, with the design under revision.
The chip was handed to Asicland, a TSMC value chain aggregator, to be optimised for the foundry process. Industry sources say the result fell well short of what Mobilint had specified, with responsibility shared by both companies.
The cost shows up in guidance. Mobilint has halved its revenue target for the year, to KRW 15B from KRW 30B, and it extended its Asicland contract by a year, to the end of 2027, lifting the value about 7 percent to KRW 16.5B.
Customers are feeling the slip. A planned shipment of 1,000 Regulus chips for drone maker Daon I&C has not landed, and Daon has postponed its own mass production schedule. Mobilint says a second account, Shinsegae I&C, received all the samples it asked for and met its requirements for running language models and speech recognition at the edge.
Regulus is aimed at robots, drones, appliances and vehicle recorders, and runs on a TSMC node, unlike the first-generation Aries, which Samsung built at 14 nanometers. Mobilint is staying with Asicland rather than restarting the design, judging that alternatives such as GUC and Alchip serve big tech customers first.
