Google will ship two AI chips a year
Amin Vahdat said the old two-year cycle is too slow for the AI race.
Google is speeding up its own AI chip roadmap. The company will move from a two-year generation cycle to releasing two custom chips a year, and may go faster later, AI infrastructure chief Amin Vahdat said on Wednesday at Semicon Taiwan in Taipei.
The point is to keep silicon in step with large language models and other generative systems. A two-year hardware cycle leaves a gap between new model designs and the chips that run them. More frequent chips let Google iterate on power, memory bandwidth, and interconnects alongside software updates.
Google also said it will expand research and development space in Taiwan by about 60%. Custom chips sit at the centre of its cloud and internal AI services strategy.
Google did not name exact ship dates for the next chips. It said the two-a-year cadence starts now and could rise further.
Hyperscalers are racing to design silicon rather than wait only on Nvidia and AMD. Meta has talked about putting its own Iris AI chip into production this month. Faster Google TPUs mean sharper competition on cloud AI price and speed, which Indian startups and IT services firms buy into every day.
It also tightens the squeeze on foundry capacity in Taiwan. More aggressive schedules need deeper ties with manufacturers and packaging partners.
Watch for the first product names under the new cadence, and whether Google Cloud packages them for customers with clear pricing. Rival cloud chip plans will set the bar for how fast "two a year" has to mean in practice.