On April 22, 2026, Google Cloud Next opened in Las Vegas with CEO Thomas Kurian telling the room that "you have moved beyond the pilot, the experimental phase is behind us." Over the next two days Google rolled out a unified agent platform, eighth-generation TPUs, a no-code agent builder inside Workspace, and a $750 million partner fund aimed at the world's largest consulting firms. For business buyers, the real story is not any single product. It is that Google is now betting its entire cloud strategy on the claim that agentic AI has crossed from experiment into infrastructure.
What Google Announced, in One Paragraph
Google consolidated Vertex AI and Agentspace into the Gemini Enterprise Agent Platform, a single environment for building, governing, and optimizing agents. It launched TPU 8t and TPU 8i, eighth-generation chips optimized for training and inference respectively. It introduced Google Workspace Studio, a no-code tool that lets employees describe automations in plain English across Gmail, Docs, Sheets, Drive, Meet, and Chat. It committed $750 million to consulting and systems-integration partners building on the stack, with Accenture, Deloitte, KPMG, PwC, and NTT DATA as headline partners. And it moved the Agent2Agent (A2A) protocol to version 1.2 under Linux Foundation governance, reporting 150 organizations running it in production.
Here is what each one actually means.
The Platform Consolidation Is the Headline
For two years, Vertex AI was the place you went to train and tune models on Google Cloud. Agentspace was the place you built agents for knowledge workers. They were adjacent products with overlapping features. At Cloud Next 2026, Google folded both into the Gemini Enterprise Agent Platform, and the reframing matters.
The platform includes Agent Studio for development, Agent Registry and Agent Identity for governance, Agent Gateway for controlled deployment, Agent Simulation and Agent Evaluation for testing, and Agent Observability for production monitoring. It provides access to Gemini 3.1 Pro, Gemini 3.1 Flash, and Lyria 3, plus Anthropic's Claude Opus, Sonnet, and Haiku and more than 200 other models through the Model Garden.
Our take: Google is making the bet that enterprises do not want to buy agent pieces from six vendors and stitch them together. Kurian went out of his way to contrast Google's approach with competitors "handing you the pieces, not the platform," and that framing lands because it is mostly true. Anyone who has tried to run a production agent today knows that observability, identity, and evaluation are still the weakest links. A single governed platform is genuinely useful if the governance actually works.
The risk is the same every unified platform carries. You trade best-of-breed optionality for integration simplicity, and the lock-in is real. We covered the competing approach in how Anthropic is closing the AI production gap with Managed Agents, Google's closest philosophical competitor. Both assume the market wants fewer vendors managing the full agent lifecycle. Whether that assumption holds is the multi-billion dollar question of 2026.
The $750 Million Partner Fund Is a Channel Strategy, Not a Giveaway
The $750 million partner fund is reportedly the largest single partner investment from any hyperscaler. It mixes credits, co-investment capital, training subsidies, and go-to-market funding. The partner commitments alongside it are substantial: Accenture has built 450-plus agents on Google's stack, Deloitte called this its "largest investment yet," KPMG committed $100 million, PwC $400 million, and NTT DATA dedicated 5,000 engineers to Google Cloud agent work.
Read that as a channel strategy rather than a generosity story. The large consulting firms are where enterprise AI decisions are actually made in Fortune 500 boardrooms. Getting Accenture, Deloitte, KPMG, and PwC to bake Google's agent platform into their delivery models is how Google competes with the default Microsoft-on-the-desktop and OpenAI-through-partnerships distribution those firms already have.
What this means for buyers: if you are engaging a global SI on an agentic AI program in the next twelve months, your consulting partner almost certainly has Google co-investment credits available. Ask. The economics of a pilot change meaningfully when the SI is being subsidized to train its staff on the platform you are about to standardize on. That does not mean Google is the right choice. It does mean the RFP math is different than it would have been six months ago.
Workspace Studio Puts Agents Into the Apps People Already Use
The most underrated announcement at Cloud Next 2026 is arguably Workspace Studio. It is a no-code agent builder that lives inside Gmail, Docs, Sheets, Drive, Meet, and Chat. Employees describe what they want to automate in plain language, and Gemini 3 assembles the agent. The tool connects to Asana, Jira, Mailchimp, and Salesforce out of the box, and calls external APIs through webhooks or custom Apps Script logic.
This is a direct challenge to the ecosystem of horizontal workflow automation platforms and first-generation no-code agent builders. When every Google Workspace seat ships with a prompt-driven automation tool, the addressable market for standalone automation SaaS narrows to use cases that genuinely require more than what Gemini can assemble inside the suite. That is a smaller market than many incumbents have priced into their valuations.
For businesses already on Google Workspace, the practical implication is that rank-and-file employees now have a sanctioned path to build automations without IT approval cycles. That is both a productivity unlock and a governance problem, and it is exactly the kind of tension we laid out in our governance framework for growing companies. Plan for it.
Eighth-Generation TPUs: The Infrastructure Story
The hardware announcements are easy to skip as a business reader, but they shape the cost curve of everything above. Google launched two variants. TPU 8t is optimized for training and scales to 9,600 chips and two petabytes of shared, high-bandwidth memory in a single superpod, with roughly three times the processing power of the prior Ironwood generation and up to two times the performance per watt. TPU 8i is optimized for inference, linking 1,152 chips in a pod with three times the on-chip SRAM of earlier generations, aimed specifically at running millions of agents concurrently at low latency.
The through-line is that Google is designing silicon for a world where inference costs, not training costs, dominate the AI budget. That matches what we have been seeing in client engagements. Agentic workloads run thousands of model calls per completed task, and inference efficiency determines whether a given workflow is economically viable at scale. The TPU 8i design choices, especially the on-chip SRAM increase, signal that Google expects agent inference to become the default workload shape.
This is the same economic pressure driving every hyperscaler's custom silicon story. We covered the vendor implications in our Anthropic and AWS compute deal analysis. The short version is that who owns the inference cost curve will own a lot of the downstream application margin.
A2A Protocol: The Interoperability Quiet Win
One announcement that did not grab headlines but probably should have: the Agent2Agent protocol reached version 1.2, is now governed by the Linux Foundation's Agentic AI Foundation, and is reportedly in production at 150 organizations. Microsoft, AWS, Salesforce, SAP, and ServiceNow are named as running it in production, with native support built into Google's Agent Development Kit, LangGraph, CrewAI, LlamaIndex, Semantic Kernel, and AutoGen.
A2A complements Anthropic's Model Context Protocol (MCP). MCP governs how an agent talks to tools and data sources. A2A governs how agents talk to each other across platforms and organizations. Together they are shaping up as the TCP and HTTP of the agentic era: boring, essential, and the thing that prevents a vendor from trapping your agent inside its own walled garden.
Our take: insist on A2A and MCP support in any agent platform RFP you run this year. The cost of not insisting is that a hand-off from your CRM agent to your ERP agent turns into a custom integration project every time a vendor changes its internal APIs. This is the lesson we covered in the foundations of multi-agent systems, and it only gets more important as cross-vendor agent workflows become the norm.
What This Means for Your AI Strategy
Three practical shifts follow from Cloud Next 2026.
Agent platform choice is now a strategic decision. It is no longer a tooling call you can reverse in a quarter. The governance, identity, and observability modules create real switching cost the moment you put agents into production. Pick deliberately, and pick with a multi-year horizon.
Benchmark against at least two platforms on real workflows. Google's Gemini Enterprise, Anthropic's Claude Managed Agents, and the OpenAI enterprise stack are the three realistic options for most mid-market and enterprise buyers today. Run the same workflow on two of them before committing. Demos lie. Production runs do not.
Treat open protocols as non-negotiable. Any platform that cannot speak A2A and MCP is asking you to accept future migration pain in exchange for present simplicity. That trade is only worth making if the platform is genuinely best-in-class on your specific workload, and even then, plan the egress story before you sign.
We laid out the broader strategic posture in the AI Playbook for 2026. Cloud Next 2026 did not change the playbook. It confirmed that the platform consolidation half of the playbook is now urgent rather than optional.
Key Takeaways
- Google Cloud Next 2026 opened April 22 with the unified Gemini Enterprise Agent Platform, launched as the successor to Vertex AI and Agentspace.
- Eighth-generation TPUs (8t for training, 8i for inference) target the agentic workload shape, with TPU 8i explicitly designed to run millions of concurrent agents.
- A $750 million partner fund, with Accenture, Deloitte, KPMG ($100M), PwC ($400M), and NTT DATA as headline partners, makes Google the distribution story to beat at the SI layer.
- Workspace Studio puts no-code agent building inside Gmail, Docs, Sheets, and Drive, pressuring horizontal automation SaaS incumbents.
- Agent2Agent protocol v1.2 and MCP support mean cross-vendor agent interoperability is becoming a baseline expectation.
- Agent platform selection is now a multi-year strategic commitment, not a project-level choice. Benchmark across platforms on real workloads and insist on open protocol support.
The businesses that move early on agent platform strategy will have a meaningful advantage over those still running isolated pilots. If you want to be one of them, let's start with a conversation.