Google held Cloud Next ‘26 in Las Vegas from April 22 to 24, and the keynote landed with a clearer thesis than any of the company’s prior AI events: enterprises are not buying chatbots, they are buying agents, and the cloud platform that wins is the one that makes building and running them cheap, observable, and safe. Across 32,000 in-person attendees and 260 announcements, three threads carried the story — the Gemini Enterprise Agent Platform, eighth-generation TPUs, and the Agentic Data Cloud.
This is Google’s most coherent answer yet to the question that has been hanging over enterprise AI since MCP crossed 97 million monthly installs: once agents are the unit of work, what does the cloud look like underneath?
The Gemini Enterprise Agent Platform
The headline announcement is the Gemini Enterprise Agent Platform — an end-to-end workspace for building, governing, and scaling agents. The pitch is that everything required to ship an agent into production lives in one console: model choice (Gemini 3 Pro, Gemini 3 Flash, third-party models via API), tools and connectors, identity and access policies, deployment, monitoring, evals, and cost attribution.
The piece that distinguishes this from earlier agent products is governance. Most enterprises currently have agents running in three or four places — a Copilot Studio bot here, an internal Python LangGraph project there, a vendor-provided agent in a SaaS tool — and no single surface to enforce data-access policies across them. The Gemini Enterprise Agent Platform tries to be that surface, with role-based access controls, audit logs, and policy enforcement at the agent-to-tool boundary.
For platform teams, the value proposition is straightforward: instead of negotiating policy with every agent vendor, define it once in Gemini Enterprise and apply it everywhere. Whether the abstraction holds when half your agents live in Microsoft’s ecosystem is the open question.
8th-Generation TPUs: Training and Inference Specialization
The hardware story is that Google split its TPU roadmap into two SKUs:
- TPU 8t (training) — optimized for frontier-scale model training. Scales to 9,600 TPUs and 2 PB of shared high-bandwidth memory in a single superpod via new Inter-Chip Interconnect (ICI) topology. Delivers 3x the processing power of Ironwood (the prior generation) and up to 2x more performance per watt.
- TPU 8i (inference) — optimized for low-latency serving. Uses a new “Boardfly” topology to directly connect 1,152 TPUs in a single pod, with 3x more on-chip SRAM to host larger KV caches entirely on-silicon, plus a dedicated Collectives Acceleration Engine. Delivers 80% better performance per dollar for inference versus the prior generation.
The strategic move is the split. Training and inference have different bottlenecks — training is bandwidth-bound and benefits from massive memory pooling, while inference is latency-bound and benefits from cache locality. NVIDIA solves both with the same Blackwell-class part and lets the buyer eat the inefficiency. Google is betting that purpose-built silicon at each end of the workload pipeline wins on total cost of ownership.
The inference number — 80% better performance per dollar — is the one that matters for the agent economy. If millions of concurrent agents need to be served per enterprise, the cost of inference per agent action is the variable that determines whether the workload pencils out. TPU 8i is engineered around exactly that scaling regime.
The Agentic Data Cloud
The third pillar is data. Agents are useless without governed access to the right data, and most enterprises are still arguing about which catalog is authoritative. Google’s answer is the Agentic Data Cloud — a cross-cloud Lakehouse plus a Knowledge Catalog that exposes data to agents with the same access controls humans would have.
The cross-cloud framing matters. Google is not pretending enterprises will move all their data to BigQuery. The Lakehouse layer is designed to query data where it sits — including in S3 and Azure Storage — and the Knowledge Catalog enriches that data with semantic metadata that agents can navigate. This is the same insight that drove the MCP standardization push at the protocol layer, applied at the data layer.
What Else Shipped
Beyond the three pillars, the announcements that will matter in six months:
- Gemini 3 Pro and Gemini 3 Deep Think generally available, with Deep Think gated initially to AI Ultra subscribers and safety testers
- Gemma 4 expanded availability — already shipped April 2 under Apache 2.0, now fully integrated into Vertex AI
- Deep Research Max — Gemini’s research mode extended to autonomously orchestrate multi-day investigations
- Learn Mode in Colab — a tutor-style overlay for data scientists picking up unfamiliar tools
- Gemini for Mac — a native macOS app with screen context awareness and global shortcut launch
The Mac client is the consumer-side echo of what OpenAI did with ChatGPT 5.5’s unified super-app. The implicit recognition: assistants that live in browser tabs lose to assistants that live one keystroke away.
What It Adds Up To
The honest read on Cloud Next ‘26 is that Google has finally built the platform story it needed. A year ago, the company had a credible model (Gemini), credible silicon (TPU), and credible data infrastructure (BigQuery) — but they did not snap together into a single agentic story. They do now.
The competitive question is whether Microsoft’s Copilot ecosystem and OpenAI’s distribution can be displaced by a more vertically integrated alternative. Most enterprises will not migrate wholesale — the more likely outcome is that Google captures net-new agent workloads while Microsoft retains the installed base inside Office and Teams. The size of the new market is what determines who wins on revenue.
For automation buyers, the practical takeaway is to evaluate the Gemini Enterprise Agent Platform on three axes that the slick demo will not stress: how it handles agents you did not build, how its policy model survives a federated identity audit, and what the actual TPU 8i bill looks like at your concurrency.