Google Cloud Next Push

Google used Cloud Next on April 22 to pitch itself as a full enterprise artificial intelligence stack, not just a model vendor. (blog.google) The company’s new lineup spans data, model building, agents and chips, with Sundar Pichai highlighting a Gemini Enterprise Agent Platform and new eighth-generation Tensor Processing Units, or TPUs. Google said TPU 8i is tuned for inference — the step where a trained model answers requests — while TPU 8t is tuned for training. (blog.google) Google also pushed a cross-cloud data pitch. Its Cloud blog said the new lakehouse roadmap centers on managed Apache Iceberg tables, Spark processing and “always-on context” so agents can work across operational and analytical data, while a separate infrastructure post grouped Smart Storage and Knowledge Catalog under a “unified data layer.” (cloud.google.com, cloud.google.com) A lakehouse is the layer companies use to keep raw files and query-ready tables in one system, instead of splitting them between a data lake and a warehouse. Apache Iceberg is the open table format many cloud vendors now support so customers can move data across engines without rewriting everything. (cloud.google.com, cloud.google.com) That matters in 2026 because enterprise buyers are trying to run agents against data spread across clouds, old databases and object storage, while also controlling where models run and what they can access. Google’s cross-cloud post framed the problem as one of compute placement, secure connectivity, governance and “digital sovereignty” for data that cannot leave a region or environment. (cloud.google.com) The chip push is part of the same argument. Google said TPU 8i can connect 1,152 TPUs in one pod and adds three times more on-chip static random-access memory, or SRAM, than the prior design, which it said is aimed at serving “millions of agents” with lower latency and lower cost. (blog.google, blog.google) Google is also building on last year’s inference hardware launch. At Cloud Next 2025, it introduced the seventh-generation Ironwood TPU as its first TPU designed specifically for inference, and this week it said the new eighth-generation family delivers three times Ironwood’s processing power and up to twice the performance per watt. (blog.google, blog.google) On the data side, Google has been laying groundwork for months. Two weeks before Next, the company said BigQuery would preview read-and-write interoperability with Iceberg-compatible engines including Trino and Spark through a Google-managed Iceberg REST Catalog. (cloud.google.com) Google’s message to information-technology teams is that buying a model is no longer the whole job. The harder work now sits in choosing where data lives, which accelerator runs each workload, and how an agent gets governed access across clouds without breaking cost or compliance rules. (cloud.google.com, blog.google)