AI infrastructure diversifies
- Hyperscalers are building alternatives to Nvidia, with Google unveiling new training and inference chips and a $750M partner fund. - Google previewed its Ironwood TPU and an eighth-generation split into training and inference chips to be made at TSMC's 2nm node. - The move signals hyperscalers are creating parallel silicon and partnership paths, reducing single-vendor dependency and changing enterprise sourcing calculus. (cnbc.com; thenextweb.com)
Google used its Cloud Next conference on April 22 to show a new shape for AI hardware: one chip for training models, another for running them after deployment. (cnbc.com) The company said its eighth-generation Tensor Processing Units, called TPU 8t and TPU 8i, will both become available later in 2026. TPU 8t is aimed at pre-training and other heavy model-building work, while TPU 8i is built for inference and reinforcement learning. (cloud.google.com) Inference is the part customers feel: a user sends a prompt, and the model responds. Google said the split reflects different bottlenecks in pre-training, post-training, and real-time serving, which now need different hardware designs. (techcrunch.com; cloud.google.com) Google also tied the chip launch to a services push. On the same day, it announced a $750 million fund for its 120,000-member partner ecosystem to support agentic artificial intelligence development, adoption, and training. (googlecloudpresscorner.com; cloud.google.com) That pairing matters because cloud companies are no longer selling only rented computing power. Google is packaging chips, models, software, and consulting channels together as enterprises decide where to build and run AI systems. (blog.google; googlecloudpresscorner.com) Google is not alone in building custom silicon. CNBC reported Amazon is pursuing a similar approach, Microsoft announced a second-generation AI chip in January, and Meta said last week that it is working with Broadcom on multiple AI processors. (cnbc.com) Google’s own TPU effort predates the current boom. The company began using internally designed AI processors in 2015 and started renting them to cloud customers in 2018, years before today’s scramble for Nvidia supply. (cnbc.com) The new chips are still being pitched as complements, not replacements, for Nvidia systems. Google said its cloud will also offer Nvidia’s Vera Rubin later this year, and TechCrunch reported the two companies are working together on networking software for Nvidia-based systems in Google’s cloud. (techcrunch.com) Google said TPU 8t delivers 2.8 times the performance of the seventh-generation Ironwood TPU at the same price, while the inference-focused chip offers 80% better performance. It also said TPU 8t can scale to 9,600 chips in a single superpod and that the broader system is designed to link more than 1 million TPUs in one cluster. (cnbc.com; cloud.google.com; techcrunch.com) The immediate result is not an end to Nvidia’s lead. It is a cloud market where the biggest providers are building more of their own chips, more of their own software, and more of their own sales channels at the same time. (techcrunch.com; cnbc.com)
