Everything Google announced at Cloud Next '26 so far

Every year at Cloud Next, Google says AI is about to change your company. This year, Sundar Pichai showed up with receipts: 89% of business teams already run AI agents, the average org runs 12 of them, and Gemini Enterprise paid seats grew 40% quarter-over-quarter in Q1. Then Cloud CEO Thomas Kurian took the stage with the thesis: "Others are handing you the pieces. We're handing you the platform."

The unstated competitor is OpenAI, which shipped its own Workspace Agents the same day. The substantive difference: OpenAI shipped a product (five Codex-powered templates you deploy in ChatGPT or Slack). Google shipped an entire operating system, stitching together a platform, models, silicon, browser automation, productivity suite, security layer, and corporate funding vehicle with the same semantic layer (a map of what your data means, not just where it sits). Whether that's a real advantage or a slide-deck flex is the rest of this story.

First up, the TL;DR

Google used its Cloud Next '26 keynote to position itself as the only vendor selling an entire agentic enterprise stack in one box; Pichai opened with "the Agentic Enterprise is real," Kurian closed with a "platform, not pieces" pitch at competitors. The same week, OpenAI shipped Workspace Agents and Anthropic fake-door tested removing Claude Code from Pro, making this the biggest agent week of 2026.

Here's what happened:

• Gemini Enterprise Agent Platform is Google's new unified mission control for building, governing, and orchestrating agents at enterprise scale; paid seats up 40% quarter-over-quarter in Q1 (TechCrunch, ZDNet).
• Workspace Intelligence adds a semantic layer (a map of meaning across your content, not just the files themselves) to Gmail, Docs, Drive, Chat, and Calendar; natural-language querying, AI Overviews in Gmail, and no-code "Skills" in Workspace Studio (Testing Catalog).
• Chrome auto browse lets Gemini handle multi-step web tasks in the browser (booking travel, filling CRM forms, summarizing dashboards) with human approval at every write action (TechCrunch).
• Eighth-generation TPUs are here: TPU 8t for training (9,600-chip superpod, 3x the processing power of Ironwood, 2x better performance-per-watt) and TPU 8i for inference (80% better price-performance, built for running millions of concurrent agents simultaneously).
• Three AI security agents with Wiz (Threat Hunting, Detection Engineering, Third-Party Context) plus AI-BOM (Bill of Materials) tracking for unauthorized AI tools in the enterprise (Google Cloud blog).
• $750M corporate AI adoption fund announced alongside a single-digit multibillion-dollar Google Cloud deal with Mira Murati's Thinking Machines Lab to train frontier models on Nvidia GB300 infrastructure.

How to try it:

1. Workspace Intelligence is on by default for Gemini customers; find "Ask Gemini" in Chat or open Workspace Studio to build a no-code Skill your team can share like a Doc.
2. Chrome auto browse activates via a Chrome Enterprise policy for US Workspace admins at launch.
3. Gemini Enterprise Agent Platform lives under Google Cloud; agent-building is code-first for technical teams, no-code for business users, with built-in governance, simulation, and role-based controls.

Why this matters: Kurian's "pieces vs platform" dig wasn't subtle. OpenAI sold Workspace Agents as an evolution of GPTs; Google sold a governance stack. Both are betting on the same end-state (agents doing the routine 45 minutes of your workday), but the pitch is different. OpenAI wants to be the app your team deploys agents inside. Google wants to be the operating system your agents run on. Buyers will pick based on where they already live: Gmail-native orgs slide into Workspace Intelligence automatically, Slack-native orgs get the OpenAI default in the sidebar.

Our take: Google's proof points are real. Danfoss automated 80% of transactional decisions in its email order processing (response time dropped from 42 hours to near real-time), Macquarie Bank reclaimed 100,000 team hours, and GE Appliances deployed 800 agents across manufacturing. The awkward question is whether a full stack beats specialized tools in actual practice. Every vendor with a platform eventually loses ground to a better point solution that plugs in. Google's TPUs are impressive, but the week's other big moves (OpenAI's Workspace Agents, Anthropic's Claude Code Pro meltdown, SpaceX's $60B Cursor acquisition option) suggest the real battle is elsewhere. It's about which silicon, model, and agent fabric eats the compute, and whether Google's answer to the vendor-neutral Model Context Protocol keeps customers portable or locks them in.

How Google got to "the full stack"

Cloud Next '26 is the culmination of a 12-month push that started at I/O 2025. The pieces arrived one at a time, often quietly:

• May 2025: Gemini 2.5 Pro hits general availability with the first native 1M-token context window in a frontier model.
• October 2025: Google Cloud launches Agentspace (an internal agent directory) as a private preview.
• January 2026: Gemini 3 Pro ships with visible reasoning tokens.
• March 2026: Google open-sources Gemma 4 under Apache 2.0.
• April 2026: Workspace Intelligence, Chrome auto browse, TPU 8, and Gemini Enterprise Agent Platform all launch at Cloud Next.

The pattern: Google spent the year assembling a stack that competitors are still selling as individual products. OpenAI has ChatGPT + Codex + API but no Workspace equivalent. Anthropic has Claude + Claude Code + Routines but neither a browser agent nor a productivity suite. Microsoft has Copilot + Foundry Agent Service but no frontier model it owns end-to-end. Only Google ships every layer in-house, and Kurian's keynote was the first time it said that out loud.

Gemini Enterprise Agent Platform: what it actually does

The agent platform sits on top of Google Cloud and lets enterprise customers do five things in one console:

• Build agents in either code (Gemini SDK for technical teams) or no-code (Workspace Studio Skills for business users).
• Govern who can build, publish, and run agents; which tools each agent can access; and which write actions (sending email, updating CRM, posting in Slack) require human approval.
• Orchestrate agents that call other agents; Google supports the vendor-neutral A2A protocol (agent-to-agent messaging) alongside its own stack.
• Simulate an agent against a synthetic workload before production, so failure modes show up in a sandbox, not in customer emails.
• Audit every run through the same compliance dashboard, including agent identity with cryptographic IDs so misbehavior can be attributed to a specific agent, not a shared user account.

Kurian framed it as "the control plane for the Agentic Enterprise." The underlying pitch is IT-friendly: agents get the same treatment as any other enterprise resource (identity, policy, observability, rollback). That's not sexy, but it's what unblocks procurement at a 50,000-person company.

The 40% QoQ growth in paid Gemini Enterprise seats is the line that matters. If accurate, Google Cloud's enterprise AI business is growing faster than OpenAI's ChatGPT Business. Translation: Google has spent a year being dismissed as a distant third in frontier AI, and the enterprise scoreboard is suddenly flipping.

Workspace Intelligence: plumbing, not a product

Workspace Intelligence is the most misunderstood announcement from Cloud Next. Part of that is Google's own fault; the press release leans on terms like "semantic unifying layer" and "agentic work" that sound like a new app to open. It isn't. You'll never see a "Workspace Intelligence" button anywhere.

Here's what it actually is: a shared context layer beneath every Gemini feature in Gmail, Docs, Slides, Sheets, Drive, and Chat. Before Workspace Intelligence, every Gemini instance was siloed; the Gemini in Gmail only saw Gmail, the Gemini in Docs only saw your current doc. Ask for a Q3 project summary and you had to paste the relevant emails and docs in yourself. After Workspace Intelligence, all those Gemini instances share one brain. You ask once in any surface and Gemini reads across everything (your inbox threads, the shared Doc, Chat history, Calendar invites), then answers with citations. You didn't have to tell it where to look.

Regular users don't "use" Workspace Intelligence directly. They use the features it powers:

• Ask Gemini in Chat: a dedicated Gemini conversation in Google Chat that Google calls "a unified command line for all of your work." Type "give me my daily briefing," or "schedule 30 minutes with Sarah and Tom about Q3 planning next week," or "draft a QBR deck from our CRM data," and Gemini does the cross-app reasoning to deliver a finished result. Integrates with Asana, Jira, and Salesforce, too.
• AI Inbox in Gmail: priority inbox 2.0; automatically surfaces upcoming bill payments, appointments, and messages from people you interact with most.
• AI Overviews in Gmail search: ask natural-language questions across your inbox, get a synthesized answer with citations instead of 12 raw results to scroll through.
• Workspace Studio Skills: no-code agents your team maintains like a shared Doc. Google's example: an invoice-review Skill that compares new invoices against recent ones to catch billing errors. Build once, share with the team, update once, whole org gets the new version automatically.
• Smarter Gemini inside each app: Docs can generate infographics grounded in your company's data and match your writing style. Slides can produce full editable decks in one shot using your company templates. Sheets has a new canvas view for interactive dashboards and HubSpot or Salesforce imports. Drive has a new "Projects" organizational unit above folders that auto-organizes team files and emails per project.

Yulie Kwon Kim, Google Workspace's VP of Product, said the company is "done with AI as a passive assistant." The idea: Workspace Intelligence proactively surfaces what matters based on your current meeting, message thread, or calendar context, rather than waiting for you to ask.

The moat question: semantic context compounds. Every email, Doc, and Chat message a customer writes teaches Workspace Intelligence what that company means when it says "Q3 launch," "the Danfoss deal," or "our standard pricing." After 18 months, switching to a competitor means re-teaching everything. Google has 3 billion Workspace users across 13 million paid customers; OpenAI and Anthropic don't have a productivity suite, so they can't build the same layer even if they wanted to. That's the real moat: not the agents, but the 18 months of company vocabulary that gets absorbed once a customer turns this on.

The app gap: smarter brain, same body

Workspace Intelligence is the plumbing. The apps the plumbing feeds are a different story. Google also shipped real app-level upgrades at Cloud Next, and the honest scorecard is mixed.

What actually got better at the app level:

• Sheets canvas (Testing Catalog): a new view mode for dashboards, heatmaps, and kanban boards layered over any dataset, plus native imports from HubSpot and Salesforce. First meaningful ground Sheets has closed on Excel in years beyond real-time collaboration.
• Drive Projects: a structural addition, not an AI bolt-on. Before: Drive was folders. Now: Drive Projects are shared spaces that auto-organize files, emails, and chat context per project.
• Slides one-shot decks: generate a full editable deck in one pass that actually adheres to your company's templates and visual rules.
• Meet "Take Notes for me" in-person: Meet's note-taker now works in conference rooms, not just remote calls. Zoom doesn't have this yet.
• Rapid Enterprise Migration: a new admin-console tool that moves emails, files, and chats from Microsoft 365 to Workspace up to 5x faster. Google explicitly targeting the single biggest adoption blocker.
• Workspace MCP Server (preview): lets third-party developers plug Workspace into their AI apps via the vendor-neutral Model Context Protocol. Google opening up, not locking in.

What's still broken, or at least unchanged:

• Google Chat is still the neglected stepchild. Ask Gemini in Chat is a smart AI feature sitting on top of a chat product that has never caught up to Slack on UX, integrations, or workflow. Putting a smart Gemini in Chat doesn't fix the fact that most teams don't want to live in Google Chat.
• Meet is still behind Zoom and Teams on core meeting UX (breakout rooms, polls, whiteboarding). Gemini note-taking helps; it doesn't close the gap.
• Docs, Sheets, and Slides as document models are still 2015-era. The fundamental app design (pages of prose, grid cells, slide rectangles) hasn't been rethought. Notion's block model, Airtable's relational database model, Coda's doc-as-app model; Google hasn't answered any of those. AI is making the old tools smarter, not different.

Against competitors, the scorecard looks like this. Versus Microsoft 365 + Copilot, Google's AI plumbing is arguably smarter now, but Excel is still fundamentally better than Sheets, Teams beats Chat, and Outlook's enterprise entrenchment is hard to dislodge (which is exactly why Rapid Enterprise Migration exists). Versus Slack + ChatGPT/Claude, OpenAI's Workspace Agents play has the edge on user love, because Slack is loved and Google Chat is not. Versus Notion, Coda, and Airtable, it's a different category; Workspace Intelligence doesn't compete with teams that live in blocks and databases. And versus the emerging category of AI-native documents (Notion AI, Claude Workspaces, Cursor for docs), Google has no answer at all. Workspace Intelligence is AI retrofitted onto pre-AI apps; the next frontier is apps designed around AI from day one, and nobody has fully owned it yet.

The honest framing: Google won the plumbing war and lost the app war. Workspace Intelligence is a real moat because semantic context compounds with customer data over time. But the apps that moat is making smarter are the same apps most knowledge workers grudgingly use because their company picked Workspace, not because they love Docs. If you're already on Google, the upgrade is substantial; your existing tools just got meaningfully smarter and that effect compounds. If you're not on Google, the apps themselves are not a compelling reason to switch. The AI plumbing alone doesn't close the UX gap versus Notion, Excel, or Slack. Google is betting on "our 3 billion existing users get much better results" rather than "we'll win you over from your current stack."

That's a defensible bet; 3 billion users is a lot of users. It's also a consolidation play, not a category-redefining one. The category-redefining stuff (AI-native documents, agent-native workflows, new document paradigms) is an open frontier none of the incumbents has fully owned yet. That's where the next wave of winners probably emerges.

Chrome auto browse: the browser as the agent runtime

Chrome auto browse is the feature most likely to get copied by competitors fastest.

How it works: a US Workspace admin flips on a Chrome Enterprise policy, and any Chrome user in the org can ask Gemini to perform multi-step web tasks inside the browser. Book travel across three tabs. Fill a CRM form using data pulled from a PDF. Compare vendor pricing across four supplier sites. Summarize a candidate's portfolio before an interview. The model sees the live DOM (the browser's internal representation of the page structure) and takes actions with the user's permission at every write step.

This is categorically different from Anthropic's Claude desktop control (which takes over your entire operating system) and from OpenAI's Codex computer use (which works in a dedicated desktop app). Google's bet: the browser is already where most work happens, and scoping the agent to web workflows keeps the blast radius contained.

There's a darker reading, which TechCrunch flagged. Google also launched "Shadow IT risk detection" in Chrome Enterprise Premium, letting IT teams detect and block unsanctioned AI tools (including competitors' agents) in the enterprise browser. The generous reading is that this is a legitimate security feature. The less generous reading is that Google is leveraging corporate IT to shut down any rival agent that tries to take root in the enterprise organically; the same playbook that killed many Enterprise 2.0 apps in the 2010s.

Eighth-gen TPUs: the silicon under the pitch

Google's TPU 8 announcement is the most technical thing to come out of Cloud Next, and the most consequential for Google's long-term economics.

Two chips launched:

• TPU 8t (training): scales to 9,600 chips in a single superpod, delivers 3x the processing power of the prior Ironwood generation, and 2x better performance-per-watt. Built for training frontier models like Gemini 4 and the next generation of Google's research models.
• TPU 8i (inference): 80% better price-performance than the prior generation; optimized for running millions of concurrent agents. This is the chip Gemini Enterprise runs on, and Thinking Machines Lab signed a single-digit multibillion-dollar deal to use TPU 8i alongside Nvidia GB300s.

Bloomberg noted that Google's TPU business is the quiet threat to Nvidia's AI compute monopoly. Roughly 10% of hyperscaler spend on AI infrastructure now goes to custom silicon, and Google is the only full-stack vendor that ships its own frontier models, its own agent platform, and its own chips. Apple and AWS have custom silicon but no frontier model; OpenAI and Anthropic have frontier models but no silicon.

The second-order effect worth flagging: if Google can run Gemini agents on its own TPUs at 80% better price-performance, it can either undercut OpenAI on margin or reinvest the savings into free-tier product. The $750M fund announced at Cloud Next is a down payment on that strategy.

Security: the Wiz stack is the dark horse

Google Cloud and Wiz (the cloud security company Google acquired in 2025) announced three AI-native security agents plus AI-BOM tracking for unauthorized AI tools:

• Threat Hunting Agent proactively scans for novel attack patterns across customer workloads.
• Detection Engineering Agent auto-generates and tunes detection rules based on evolving threat intel.
• Third-Party Context Agent enriches security incidents with vendor-specific context (for example: "this alert involves Snowflake; here's the current Snowflake threat surface").

The AI-BOM piece is worth reading closely. It's Google's answer to the problem every CISO is wrestling with: employees are running dozens of unsanctioned AI tools (ChatGPT personal accounts, Claude for coding, Midjourney for graphics), and IT has no way to inventory what's actually in use. AI-BOM gives security teams a real-time ledger of every AI service running in the corporate environment, who's using it, and what data it's seeing.

This is the category that will likely matter most over the next 18 months. The one-line version: the companies that ship the best AI governance tools probably win the enterprise before the companies that ship the best models.

The competitive frame: Google vs OpenAI vs Anthropic

The three big labs are running three different plays this week:

• Google bet on the full stack. One throat to choke, one keynote to sell it, one bill for everything.
• OpenAI bet on the product layer. Workspace Agents is a single well-designed product that sits inside existing productivity surfaces (ChatGPT, Slack) and doesn't require an IT procurement cycle to adopt.
• Anthropic bet on developer loyalty, or tried to. The Claude Code Pro fake-door test Anthropic ran and reverted inside 24 hours showed the risk of underestimating power users. Simon Willison and Gergely Orosz spotted changes bleeding into public docs; Sam Altman quote-tweeted the saga with "ok boomer."

The tell is the customer each vendor is optimizing for. Google is optimizing for the CIO. OpenAI is optimizing for the team lead who wants to ship an agent next Tuesday. Anthropic is optimizing for the developer already paying $20/month, which means every pricing decision lands as a broken promise.

What this means for you

Three concrete moves based on where you sit:

If you're already on Google Workspace for Business or Enterprise: Turn on Workspace Intelligence and start using Ask Gemini in Chat this week. The semantic layer improves with use, and the 40% QoQ growth stat suggests competitors will close the product gap but not the data gap, because the data gap compounds. Give Workspace Intelligence a 90-day runway before deciding whether to commit.

If you're a Slack + ChatGPT shop: OpenAI's Workspace Agents are the faster on-ramp. You'll get 80% of what Google ships with 10% of the change-management overhead. The tradeoff is that you're accepting OpenAI's plugin ecosystem as the lock-in, rather than Google's data gravity. Both are real; pick the one that matches where your org already lives.

If you're an IT or security leader: The AI-BOM feature is what actually changes your job. Start an inventory exercise in the next two weeks on which AI tools are in use across your org; Google's announcement will force the conversation whether your CISO is ready or not. The other vendors will follow within six months.

Where this goes next

Two open questions to watch:

Will the "full stack" pitch hold up in production? Google's Cloud Next keynote was a flex. The proof is whether customers in six months report "we picked Google because it was one bill" or "we picked OpenAI because Workspace Agents shipped Tuesday, and Gemini Enterprise Agent Platform took three months to configure." Enterprise procurement has a long memory; the ratings on specific Cloud Next launches in Q4 matter more than the keynote itself.

Does semantic context actually lock customers in, or does it leak? The assumption baked into Workspace Intelligence is that once an org teaches Google's system what its data means, switching costs go up. The counter-example is Anthropic's Claude Code Pro saga this same week: power users revolt fast when they feel captured. Google's answer has to be that the platform is so good customers don't notice the lock-in; Kurian clearly believes that. The alternative future (the one OpenAI is betting on) is that agents become a commodity layer sitting on top of vendor-neutral standards like MCP and A2A, and the "full stack" moat dissolves into a premium price without a premium product.

The one thing that's no longer in question is whether agents will do the 45 minutes of your workweek you currently hate. That fight is over. The fight now is who banks the margin when they do.