I do think the 8th-generation TPU lineup shows what full-stack control means: a training chip at three times Ironwood's processing power, an inference chip at 80% better performance per dollar. Axion now hosts them, replacing x86. We're also among the first with NVIDIA Vera Rubin — Broadcom remains our TPU partner. And we're opening TPUs to select customers for their own data centers: Thinking Machines Lab, Hudson River Trading, Boston Dynamics. Revenue there is a small percent in 2026, majority in 2027, lumpy with ship timing. Very early innings.