Google Cloud Next 2025 Deep Dive — Gemini Enterprise, TPU Ironwood, Wiz Deal: The Ground Shift Under Agent-First Management

Hello, this is Ryuta Hamamoto from TIMEWELL.

Watching the Google Cloud Next 2025 keynote live from Las Vegas (22 April 2026 week) gave me strong déjà vu. One year earlier the same stage said enterprise AI was going live. Japanese CEOs largely half-believed it. This year Thomas Kurian opened with: the experimental phase is already behind us; the problem is how to deploy AI across production. Screens showed ~75% of Google Cloud customers running some AI product in real operations. Not pilots. Revenue and daily work.

Sundar Pichai then showed Google itself: roughly 75% of new internal code is already AI-generated, with human engineers reviewing and approving. From ~30% in fall 2024 and ~50% in fall 2025, the jump in half a year is enough to make boards sit up. Below is the keynote through a Japanese-enterprise lens. Dense numbers and names, all on one line: agent-first management.

What Kurian's "production AI" declaration means

Kurian's first minutes rewrote the era, not a product sheet. "The experimental phase is behind us" and "the real challenge begins: how do you move AI into production across your entire enterprise?" landed across English-language media. He stacked that claim against 1,302 generative AI use cases Google Cloud collected from customers in production. Roughly double 2024's official list.

"Production" here is not a SaaS feature bolt-on. It is a unified stack of models, data, agents, security, and apps. Google's search/YouTube/Chrome/Android vertical integration brought to enterprise AI. That draws a hard line against best-of-breed mashups: in the agent era, tightly coupled data–model–security stacks win on speed and safety; bolted pipes create security holes and ops friction.

Behind the 75% figure sits CapEx: Google from USD 52.5B (2024) to USD 91.4B (2025) to a planned USD 175–185B (2026). About 3.5× in two years, AI-driven. Infinite pilots cannot recover that. Japanese buyers are not spectators: as Google floods the market, AI cost curves fall and the center of gravity shifts from "still trying" to "already running." Late firms lose on decision speed, not model choice. I have been saying a version of that for a year; the keynote made it harder to ignore.

Sundar Pichai on Google's self-transformation

Pichai's sharpest image: Google becoming an organization managed by agents. The 75% code figure is not "autocomplete with extra steps." Planner, Orchestrator, and Coder agents share feature work: Planner shapes design, Orchestrator decomposes and assigns, Coder implements and tests. Humans approve Planner designs and review Coder diffs. Internal metric: about 6× engineer throughput for equivalent features.

Marketing too: ideation through copy, creative, LP, delivery, measurement, and iteration on Gemini Enterprise marketing agents cut production lead time to ~30% of before. Three days to about one, three times the experiments per period, and learning curves that compound.

"Customer zero" means Google runs Gemini Enterprise on itself first (90k+ engineers, marketing, legal, HR, finance) and feeds failures outward. Vendors shift from "build and sell" to "run and harden." Product and operations fuse. Whether you treat your own agents with the same intensity decides the five-year gap: production use, including failure, converted into institutional knowledge. That is what separates "we installed Gemini Enterprise / ZEROCK" from agent-first management.

Engineering value redefines. The remaining 25% is not "humans still type code." Review, architecture, ethics, UX, and incident root cause dominate: judge, connect, own. Next-gen IDEs like Antigravity, agent environments like Claude Code Skills, plugins like Superpowers redesign human touchpoints. See also Antigravity and Workspace CLI. By 2026 the field is agents write, humans decide.

Gemini Enterprise five layers

Gemini Enterprise is an enterprise AI operating system. Kurian and Karthik Narain stacked five layers bottom-up: AI Hypercomputer, Agentic Data Cloud, Agent Defense, Agent Platform, Agent Apps & Taskbooks. Vertical integration of infrastructure with upward capture of UX. AWS/Azure history did not plant vendors this deep in the app layer.

AI Hypercomputer. TPU, CPU, network, storage designed for AI workloads; absolute cost and performance (next section).

Agentic Data Cloud. BigQuery, Cloud Storage, warehouses as a single fact base for agents. Knowledge Catalog maps schemas and semantics as a dynamic graph so agents know which column is which KPI. Smart Storage feeds file content and metadata via Object Context API; Cross-Cloud Lakehouse queries AWS/Azure data without copy via Apache Iceberg; Lightning Engine for Spark claims 4.5× open-source and 2× major competitor price-performance. Data access is the agent bottleneck. This layer decides production smoothness. TIMEWELL's ZEROCK uses GraphRAG for the same idea: preserve shop-floor meaning of "what this phrase on this doc means" for Japanese work contexts.

Agent Defense. Where the Wiz deal shows most: AI Application Protection Platform, inline security hooks on generated code, AI-BOM inventory of models/IDE extensions/external MCP servers, Agent Identity and Gateway for per-agent zero trust. Advance other layers without this and you put luxury furniture in a house with no walls.

Agent Platform (Gemini Enterprise Agent Platform). Evolved Vertex AI: Agent Studio (low-code) plus ADK (code-first). Runtime multi-day state, Memory Bank for durable context, long-running agents through nights and holidays. Governance: Agent Identity, Registry, Gateway; Simulation, Evaluation, Observability. Closer to an HR system for digital employees than a "tool."

Agent Apps & Taskbooks. Workspace-extended Gemini Enterprise App (Ask Gemini, Deep Think). Taskbooks template month-end close, recruiting, customer onboarding, export control with agents, data links, and permissions pre-bundled. Explicit Microsoft 365 interop (Canvas → Office export) is a land-without-replacing-Office play.

One idea through the stack: invite a new class of "employee" safely and productively by designing data context, defense, roles, tools, and behavior end to end. "Which vendor's which layer in which order" ages poorly against a stack that trades layer choice for zero inter-layer friction.

Strategic placement of the new models

Model news looked modest; strategy was not. Gemini 3.1 Pro, Gemini 3.1 Flash Image (Nano Banana 2), Veo 3.1 Lite, Lyria 3 Pro, and first-class Anthropic Claude Opus 4.7 support.

3.1 Pro is optimized for complex workflow orchestration: multi-step tool-using chains, not single Q&A. Agent value is long stable reasoning chains, not isolated IQ. Flash Image is low-latency multimodal for interactive UX. Veo/Lyria embed SynthID watermarking and IP protection so marketing and training video become legally operable.

Claude Opus 4.7 as a peer model is the "not Gemini-only" signal. Narain: customers pick the best model for the task. Multi-model routing (Claude for contracts, Gemini Flash for chat, Nano Banana for images, Lyria for audio) becomes an application quality problem. Watermarks and audit logs become table stakes for listed Japanese firms.

Agent platform depth

Not a Vertex rename. Build / extend / govern / optimize cover full agent lifecycle: Studio, Runtime, Identity, Memory Bank, Simulation, MCP.

Studio is visual low-code; ADK is TypeScript/Python code-first on the same agent spec. Marketers draft, engineers harden, ship. Runtime multi-day state makes month-end and quarterly report agents real. Memory Bank holds "last week's A-co deal" and failure lessons so agents participate in organizational learning. Identity + Registry + Gateway extend zero trust from people to agents.

Simulation pre-prod, Evaluation continuous quality, Observability traces "why this conclusion." Without that, ops and audit will not trust agents. I have sat in enough Japanese audit rooms to know that last sentence is not optional.

Native MCP is decisive: Google Cloud and Workspace services behind a single MCP endpoint for agents. MCP (Anthropic, Nov 2024; donated to Agentic AI Foundation under Linux Foundation Dec 2025) is industry standard with OpenAI, Microsoft, AWS, Google, Cloudflare. By March 2026, 10k+ active MCP servers and 97M monthly SDK downloads. Forrester: 30% of enterprise app vendors ship own MCP servers in 2026. Combined with Agent2Agent (A2A) (tools via MCP, agent-to-agent via A2A) you get two layers. A2A went from 50 partners (2025) to 150 production orgs (2026), with Microsoft, AWS, Salesforce, SAP, ServiceNow aboard. Cross-vendor handoffs (Agentforce → Vertex → ServiceNow) are live.

Japanese agent projects usually fail not on models or tools but on governance and observability. Identity/Registry/Gateway/Observability finally give IT and audit a shared base. Agent-as-team practice already shows in development (Claude Code Agent Teams); the platform side is catching up.

TPU Ironwood, AI CPUs, network impact

Infrastructure is dry for boards and decisive for multi-year cost. Headline: 8th-gen TPU 8t and 8i, Arm Axion N4A, NVIDIA Vera Rubin NVL72, Virgo Network, Managed Lustre storage.

TPU 8t (training): new ICI, up to 9,600 TPUs per superpod, 2 PB shared HBM; 3× vs 7th-gen Ironwood, up to 2× efficiency per watt. Ironwood (Nov 2025) was already 4,614 TFLOPs/chip, 42.5 ExaFLOPS/pod. 8t is another generational step.

TPU 8i (inference): Boardfly topology, 1,152 TPUs/superpod, 3× on-chip SRAM, Collectives Acceleration Engine, ~80% better inference price-performance. Agents running 24/7 make token cost a P&L line. 80% better is "run agents at 1/5–1/3 the old cost" territory. "Millions of agents simultaneously, cost-efficiently" tracks architecture, not slogan.

Axion N4A handles non-AI general work cheaply so front ends and preprocessing leave TPUs for train/infer. Vera Rubin and H200 keep GPU customers. Virgo doubles inter-superpod bandwidth; Managed Lustre hits ~10 TB/s. Bottlenecks often sit in I/O, not FLOPs.

Citadel Securities moved quant research to Google Cloud, >USD 400B daily trading volume, joint tooling for job state, profiling, efficiency, mixing on-demand and preemptible capacity for 2–4× speed and ~30% cost cut. Japanese banks exploring quant, market risk, and non-financial risk agents can assume similar infrastructure is available.

"All clouds are roughly the same" is outdated when dedicated silicon opens ~80% cost gaps on the same workload.

Cases that sketch "agent management"

Five customers stood out: Walmart, Virgin Voyages, KPMG, NASA, Citadel Securities, plus long-time Signify (Philips Hue), on GKE at 200M transactions/day including 18M remote lighting commands. IoT×agent already in consumer life.

Walmart: Pixel Fold for store staff; Gemini Enterprise agents for customer Q&A, inventory, supply-chain triage. HQ/field decision walls flatten.

Virgin Voyages: Rovey concierge agent on Gemini Enterprise + Google Distributed Cloud; creative timelines cut up to 60%. Offline-tolerant agents matter for Japanese plants, construction, ships, aircraft.

KPMG: KPMG Law on Google AI; Agentspace internally; A2A design partner. Consultancies and auditors that run agents, not only recommend them, hold the next axis. Signal for Japanese SI and Big Four peers.

NASA: Gemini Enterprise agents on major Artemis II elements. Mission-critical production is a maturity marker (quality, governance, observability past a bar).

Citadel again: not just infra swap. Research process rebuilt cloud-and-agent-native, with Google ML performance specialists embedded. Buy the vendor into your org. That depth is where Japanese agent programs will split winners from the rest.

Shopping and Food ordering agents under Gemini Enterprise for Customer Experience, with Omnichannel Gateway keeping context across chat, voice, app, web, phone, and store. CS as a management problem of data integration and accountability redesign.

Security for the agent era

Francis deSouza's security segment was the spine-stiffener: Wiz deal, Agentic Defense, M-Trends 2026 numbers.

Google agreed to buy Wiz for USD 32B (Mar 2025), closed Mar 2026. Google's largest deal. Wiz crossed USD 1B ARR in 2025 on cloud/container/AI vulnerability management.

M-Trends 2026: Mean Time to Exploit of −7 days. Average exploitation before patch. AI accelerates discovery and exploit codegen. React2Shell saw crypto miners ~48 hours after disclosure. Google Threat Intelligence reports commercialized LLM exploit services on dark markets.

Agentic Defense ships Threat Hunting, Detection Engineering, and Third-Party Context agents into AIPP with Wiz assets. Zero-trust defense assuming autonomous internal agents. Google Cloud Fraud Defense (reCAPTCHA evolution) GA; AI-BOM inventories frameworks, models, IDE extensions, MCP servers. SBOM for AI.

Demo numbers: threat handoff to analysis complete in 22 seconds; SOC response from 30 minutes to 60 seconds. Outsourced SOC quality now hinges on whether the vendor uses agents.

Boards, three questions I would force into the next risk committee: (1) attacker speed already exceeds human response; (2) inventory agents and data they touch; (3) per-agent identity and privilege. That checklist also selects Gemini Enterprise Identity features and domestic-region platforms like ZEROCK (AWS Japan, knowledge control, prompt control, audit logs) for Japanese data-protection fit.

How Japanese companies should re-read this

Straight conclusion: not doing AI agents is no longer a defensible management decision. "Still experimental," "no talent," "security fears" all lost their 2026 spring warrants. Experiments ended; agents are designed for talent shortage; security is faster with agents.

1. Experimental phase is over. 75% of Google Cloud customers in production, 150 A2A orgs, 75% of Google's new code AI-made. Saying "we're still in PoC" is announcing lag while competitors accumulate customer data and process gains.

2. Talent shortage flips. Not adopting agents deepens shortage. Japanese coverage frames 2026 as the year agents leave trial for ROI execution. Peers redraw agent vs human work and concentrate humans on judgment and customers. Like Google's Planner/Orchestrator/Coder split, Japanese finance, legal, sales ops, and HR should identify agent-ready work and redeploy people to decisions and relationships.

3. Security anxiety flips. Not using agents is the bigger risk when MTTExploit is −7 days. Agentic Defense, Wiz integration, and GraphRAG knowledge control (ZEROCK) can meet Japanese audit needs while matching attacker AI. Agents are defensive tools; "agents are dangerous" is the old frame.

Japanese win path: specialized agents where commercial custom, regulation, and Japanese nuance dominate. Do not race general models. Inject proprietary process knowledge. Export control with METI classification expertise as TRAFEED; finance regulation, IFRS, labor, HR evaluation, compliance training, medical docs, quality docs. General agents lose to specialized stacks on accuracy and governance.

Sequence: not big-bang. TIMEWELL's WARP (WARP / WARP NEXT / WARP BASIC) is monthly AI consulting for leadership and field. Start one process (invoices, inquiries, approvals, contract review), then horizontalize governance/observability patterns, then standardize identity/registry/gateway, then marketplace and external agents. Question one: "Can we put the first agent in production next week?" not "must we change everything tomorrow."

Three-year view and the next move

Google Cloud Next 2025 was a declaration that management premises rewrote. Kurian's pilot-era end is number-backed. CapEx doubling, Wiz at USD 32B, 75% AI-written code. The current does not reverse. Gartner's "40% of enterprise apps with task-specific agents by 2026" will be past tense by year-end. McKinsey's USD 2.6–4.4T annual agent value is hundreds of trillions of yen of reallocation. Opportunity and crisis together for Japan.

Three questions I put to Japanese firms for the rest of 2026: Which smallest unit of work can an agent own next week? Who designs data, model, security, and accountability for it? When it still runs in six months, what should the people who used to own that work spend time on? Only boards that answer the third honestly harvest agent-first management.

In a flood of foundation news, difference is decision speed and context depth. TIMEWELL's ZEROCK and WARP own those two together. GraphRAG hands company knowledge to agents; WARP co-runs monthly executive judgment and ops improvement. In a "no reason left not to start" year, join the "already starting" side with a small, fast, high-learning first move. Reach out with a concrete case. We want to build the next three years' landscape with operators watching the same 2026.

References

1. Google Cloud Blog, Next '26 day 1 recap — https://cloud.google.com/blog/topics/google-cloud-next/next26-day-1-recap
2. Google Cloud Blog, Introducing Gemini Enterprise Agent Platform — https://cloud.google.com/blog/products/ai-machine-learning/introducing-gemini-enterprise-agent-platform
3. Capacity Global, Google Cloud declares the pilot era over — https://capacityglobal.com/news/google-cloud-declares-pilot-era-over/
4. Google Cloud Blog, 10 industry leaders building the agentic enterprise — https://blog.google/innovation-and-ai/infrastructure-and-cloud/google-cloud/cloud-next-2026-customer-round-up/
5. Stratechery, Interview with Thomas Kurian — https://stratechery.com/2026/an-interview-with-google-cloud-ceo-thomas-kurian-about-the-agentic-moment/
6. Google Blog, Ironwood TPU — https://blog.google/innovation-and-ai/infrastructure-and-cloud/google-cloud/ironwood-tpu-age-of-inference/
7. TechCrunch, Google completes $32B Wiz acquisition — https://techcrunch.com/2026/03/11/google-completes-32b-acquisition-of-wiz/
8. Google Cloud Blog, Welcoming Wiz — https://cloud.google.com/blog/products/identity-security/google-completes-acquisition-of-wiz
9. Google Cloud Blog, M-Trends 2026 — https://cloud.google.com/blog/topics/threat-intelligence/m-trends-2026
10. Gartner, 40% of enterprise apps with task-specific AI agents by 2026 — https://www.gartner.com/en/newsroom/press-releases/2025-08-26-gartner-predicts-40-percent-of-enterprise-apps-will-feature-task-specific-ai-agents-by-2026-up-from-less-than-5-percent-in-2025
11. SiliconANGLE, Maximizing Gemini — https://siliconangle.com/2026/04/22/maximizing-gemini-google-cloud-makes-bid-build-operating-system-enterprise-ai/
12. Constellation Research, Full-stack AI edge — https://www.constellationr.com/insights/news/google-cloud-presses-full-stack-ai-edge-new-tpus-agentic-data-cloud-gemini-enterprise
13. The Register, AI agent sprawl — https://www.theregister.com/2026/04/22/google_enterprise
14. PR Newswire, A2A Protocol surpasses 150 organizations — https://www.prnewswire.com/news-releases/a2a-protocol-surpasses-150-organizations-lands-in-major-cloud-platforms-and-sees-enterprise-production-use-in-first-year-302737641.html
15. Fortune, Alphabet $185B AI spending — https://fortune.com/2026/02/04/alphabet-google-ai-spending-supply-constraints/
16. Anthropic, Donating MCP — https://www.anthropic.com/news/donating-the-model-context-protocol-and-establishing-of-the-agentic-ai-foundation
17. Google Cloud, Citadel Securities quant research — https://cloud.google.com/transform/citadel-securities-reimagine-quantitative-reseach-cloud-scale-speed
18. Nikkei, 2026 AI agents contribute to Japanese profits — https://www.nikkei.com/article/DGXZQOUB16BQ90W5A211C2000000/
19. EnterpriseZine, 2026 year of agent execution — https://enterprisezine.jp/article/detail/23596
20. The Next Web, Google Cloud Next 2026 — https://thenextweb.com/news/google-cloud-next-ai-agents-agentic-era